Evidence · fMRIPrep
fMRIPrep’s confound estimation
The confounds fMRIPrep computes for every BOLD run: its desc-confounds_timeseries.tsv and the JSON beside it, with CompCor, DVARS, framewise displacement, the motion parameters and their expansion, the tissue signals, the spike and cosine regressors. LARMORX’s port is a replica of fMRIPrep 25.2.7’s confounds workflow, nipype 1.12.0’s confounds module and niworkflows 1.15.4’s steps, with the arithmetic of numpy, SciPy, pandas, nitime and nipy under them. In reproduce it gives the originals’ outputs bit for bit in each of the 88 cases of its five records, and run as fMRIPrep ran it, it writes fMRIPrep’s own confounds files of 36 runs column by column. Its default, fix, corrects the originals’ defects or refuses them with a clear error, each with the original’s result on request.
- Ports
- replica
larmorxApache-2.0: cratelarmorx-confounds;lx.confounds,lx.niworkflows,larmorx nipype runand larmorprepx's confounds workflow, translated from fMRIPrep 25.2.7, nipype 1.12.0 and niworkflows 1.15.4 (fMRIPrep 25.2.6's own releases on request); from the next release
- replica
- Status
- replica experimental
- Records
- fMRIPrep's and niworkflows' steps and the workflow · nipype's confounds module and command line, with its full comparison · numpy's and SciPy's linear algebra and FFTs, with its full comparison
- Reference
nipypecli run· API pages confounds, nipype-run, niworkflows-helpers
What is compared
After head-motion correction, fMRIPrep estimates, for every BOLD run, the time series that a later analysis may regress out: the confounds. Its init_bold_confs_wf (fMRIPrep 25.2.7) computes them with nipype’s and niworkflows’ interfaces and writes them as one table with a JSON that describes each column:
- The motion parameters from the head-motion matrices (FSL’s convention), their derivatives and squares, the root mean square deviation, and framewise displacement (nipype’s
FramewiseDisplacement); - DVARS, raw, standardised and voxel-wise (nipype’s
ComputeDVARS); - CompCor: the anatomical components of the white matter, the CSF and both (aCompCor, on masks resampled from the anatomical tissue maps with a Gaussian kernel), the temporal components of the most variable voxels (tCompCor) and the components of the brain’s edge (crownCompCor), with the cosine regressors of the high-pass filter;
- the tissue signals: the global, white-matter and CSF means, and their expansion;
- the outlier regressors: one column per volume whose framewise displacement or standardised DVARS passes its threshold.
The replica runs beside the originals, run in-process: nipype 1.12.0, niworkflows 1.15.4 and fMRIPrep 25.2.7’s interfaces with numpy, SciPy, pandas and nitime. fMRIPrep’s own files are those of its runs, with fMRIPrep 25.2.6 and its image’s releases (numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3).
larmorx nipype run nipype.algorithms.confounds ComputeDVARS bold.nii.gz brainmask.nii.gz --save_allimport larmorx as lx
# nipype's DVARS: standardised, raw and voxel-wise
dvars = lx.confounds.compute_dvars("bold.nii.gz", "brainmask.nii.gz", n_threads=8)
# nipype's framewise displacement from FSL's motion parameters
fd = lx.confounds.framewise_displacement("motion_params.txt", "FSL", out_dir="work")The API pages describe every function and option, and larmorprepx’s whole confounds workflow of a BOLD run (bold_confounds). How to read what follows: how to read the evidence.
Its port
fMRIPrep, nipype, niworkflows, numpy, SciPy, pandas, nitime and nipy are permissive (Apache-2.0, BSD-3-Clause), so the replica is in larmorx itself. Their Python is the specification, with the arithmetic of the library releases they call.
- The workflow is a plain Python function (
larmorx.pipelines.larmorprepx.confounds.bold_confounds), with no nipype and no graph engine. Each node writes its files into the folder nipype would give it, so a run can be compared with fMRIPrep’s node folders file by file. - The numeric routines run in Rust (
larmorx-confounds): numpy’s and SciPy’s linear algebra (OpenBLAS’ kernels, LAPACK’s SVD and least-squares paths), SciPy’s FFTs, numpy’s reductions and percentiles, and pandas’ reading and writing of tables, each in the order of the original’s code, son_threadsnever changes a bit. - The programs underneath run through LARMORX’s own replicas: ANTs’
antsApplyTransformsfor the masks’ resampling (lx.ants). - nipype’s own command line for these interfaces,
nipypecli run, islarmorx nipype run.
A clean room of these steps is planned after FreeSurfer (roadmap).
The verdicts at a glance
The replica is a translation of the originals’ source, so the question is whether it gives their bytes.
Select a segment to list its cases below.
The records run the originals in-process as the reference, in three Python environments:
- the newest library releases (numpy 2.5.4 with OpenBLAS 0.3.34, SciPy 1.18.1, pandas 3.0.6), LARMORX’s reference;
- the releases the other comparisons share (numpy 2.3.5, SciPy 1.15.3);
- fMRIPrep 25.2.6’s image (numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3).
They compare the replica with them in reproduce (upstream_defects="reproduce"): every array, file, table cell, header, warning, message and exit status, bit for bit. The 88 cases of the five records pass, with 4,181 checks. Each case is one of the comparison’s check scripts (or one of its sections) run as a process of its own, one check per line it prints.
The replica’s records against the originals: fMRIPrep 25.2.7, nipype 1.12.0, niworkflows 1.15.4, numpy, SciPy and pandas (every case passes)
| Record | Suite | Tier | Cases | Pass | Checks | From commit |
|---|---|---|---|---|---|---|
| confounds-fmriprep | confounds-fmriprep | standard | 16 | 16 | 990 | 9a106cf16763 |
| confounds-nipype | confounds-nipype | standard | 25 | 25 | 588 | 34272ae120de |
| confounds-nipype-full | confounds-nipype | full | 3 | 3 | 90 | b7420b224553 |
| confounds-linalg | confounds-linalg | standard | 32 | 32 | 1,673 | 9a106cf16763 |
| confounds-linalg-full | confounds-linalg | full | 12 | 12 | 840 | e54c3fb52532 |
Each record runs the investigation’s check scripts (validation/investigations/confounds) as separate processes in the environment each case names, from a clean commit: e1 numpy 2.5.4 / OpenBLAS 0.3.34 with OPENBLAS_L2_SIZE=512, SciPy 1.18.1, pandas 3.0.6 (and e1-zen, this machine’s GEMM_Q 512); e2 numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3; e3 fMRIPrep 25.2.6’s image (numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3). One check per summary line a script prints, its final line and its exit status; the negative controls pass when the releases are told apart. The full tiers: fMRIPrep’s 108 real CompCor nodes in every environment, the linear algebra’s every size class.
fMRIPrep’s own confounds files
The replica’s workflow ran each of the 36 confounds workflows of fMRIPrep 25.2.6’s runs end to end, from the workflow’s own inputs, as fMRIPrep ran it (reproduce, fMRIPrep 25.2.6’s releases). Each desc-confounds_timeseries.tsv is compared with fMRIPrep’s file cell by cell as text, the columns matched by name: their order is that of niworkflows’ set, which changes from one process to the next (PIPE-43), so it is never reproduced. Each JSON is compared byte for byte.
The only cells that differ are in the motion chain, where SciPy’s rotation calls the C library’s atan2 and the replica computes a correctly rounded one (decision D4: the C library’s code is not translated). Run again with fMRIPrep’s own motion table, every cell is fMRIPrep’s. The table also gives what the newest releases change (the originals’ own step, no verdict) and what fix changes.
fMRIPrep’s own confounds files, column by column: every one of the 36 confounds workflows of fMRIPrep 25.2.6’s runs
| Workflow (fMRIPrep's run / BOLD run) | BOLD series | Columns | Cells | ≠ fMRIPrep's file | ≠ with its motion table | JSON byte for byte | The step: cells changed | fix: cells changed | fix: columns negated |
|---|---|---|---|---|---|---|---|---|---|
| the 36 files25 files identical column by column, 11 once glibc’s atan2 (D4) is taken from fMRIPrep’s own motion table; the step adds 2 motion-outlier columns and removes 4 | – | – | 1,862,699 | 93 | – | 36 of 36 | 263,181 | 218,474 | 60 |
ds000005-sub-01-noreconall/bold_task_mixedgamblestask_run_01_wf8 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. | 64 × 64 × 34 × 240 | 222 | 53,280 | 8 | 0 | yes | 6,435 | 6,367 | 0 |
ds000005-sub-01-noreconall/bold_task_mixedgamblestask_run_02_wf | 64 × 64 × 34 × 240 | 211 | 50,640 | 0 | – | yes | 8,654 | 9,166 | 12 |
ds000005-sub-01-noreconall/bold_task_mixedgamblestask_run_03_wf16 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. | 64 × 64 × 34 × 240 | 218 | 52,320 | 16 | 0 | yes | 15,466 | 15,068 | 36 |
ds000210-sub-02-noreconall/bold_task_cuedSGT_run_01_echo_1_wf | 64 × 64 × 33 × 260 | 157 | 40,820 | 0 | – | yes | 8,028 | 6,936 | 0 |
ds000210-sub-02-noreconall/bold_task_rest_run_01_echo_1_wf | 72 × 72 × 46 × 204 | 211 | 43,044 | 0 | – | yes | 5,669 | 5,604 | 0 |
ds000228-sub-pixar008-noreconall/bold_task_pixar_wf | 64 × 64 × 32 × 168 | 167 | 28,056 | 0 | – | yes | 6,035 | 4,865 | 0 |
ds000228-sub-pixar008-noreconall-syn/bold_task_pixar_wf | 64 × 64 × 32 × 168 | 166 | 27,888 | 0 | – | yes | 5,028 | 4,894 | 0 |
ds000256-sub-CTS203-noreconall/bold_task_restbaseline_wf | 64 × 64 × 32 × 164 | 110 | 18,040 | 0 | – | yes | 6,553 | 4,214 | 0 |
ds000256-sub-CTS203-noreconall/bold_task_restfixationfixed_wfThe step adds 2 column(s) (motion_outlier69, motion_outlier70) and removes 0 (): FD crosses its threshold at other volumes with fMRIPrep 25.2.7’s motion parameters. | 64 × 64 × 32 × 164 | 152 | 24,928 | 0 | – | yes | 6,676 | 3,956 | 0 |
ds000256-sub-CTS203-noreconall/bold_task_restfixationnofeedback_wf5 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. The step adds 0 column(s) () and removes 1 (motion_outlier91): FD crosses its threshold at other volumes with fMRIPrep 25.2.7’s motion parameters. | 64 × 64 × 32 × 164 | 177 | 29,028 | 5 | 0 | yes | 4,415 | 3,984 | 0 |
ds000258-sub-21262-noreconall/bold_task_rest_echo_1_wf8 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. The step adds 0 column(s) () and removes 1 (motion_outlier28): FD crosses its threshold at other volumes with fMRIPrep 25.2.7’s motion parameters. | 64 × 64 × 30 × 239 | 178 | 42,542 | 8 | 0 | yes | 6,230 | 6,772 | 0 |
ds002422-sub-02-noreconall/bold_task_arithm_run_01_wf8 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. The step adds 0 column(s) () and removes 1 (motion_outlier27): FD crosses its threshold at other volumes with fMRIPrep 25.2.7’s motion parameters. | 64 × 64 × 24 × 210 | 253 | 53,130 | 8 | 0 | yes | 6,508 | 5,301 | 0 |
ds002422-sub-02-noreconall/bold_task_arithm_run_02_wf | 64 × 64 × 24 × 210 | 226 | 47,460 | 0 | – | yes | 7,549 | 5,347 | 0 |
ds002422-sub-02-noreconall/bold_task_rest_wfThe step adds 0 column(s) () and removes 1 (motion_outlier22): FD crosses its threshold at other volumes with fMRIPrep 25.2.7’s motion parameters. | 64 × 64 × 36 × 200 | 197 | 39,400 | 0 | – | yes | 11,022 | 6,132 | 1 |
ds002603-sub-01-noreconall/bold_task_wm_acq_Mb4Mm3Tr550_wf8 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. | 80 × 80 × 36 × 588 | 332 | 195,216 | 8 | 0 | yes | 16,401 | 16,592 | 0 |
ds002603-sub-01-noreconall/bold_task_wm_acq_SeqMm3Tr2000_wf | 80 × 80 × 36 × 160 | 162 | 25,920 | 0 | – | yes | 3,987 | 4,571 | 1 |
ds002603-sub-01-noreconall-epi/bold_task_wm_acq_Mb4Mm3Tr550_wf8 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. | 80 × 80 × 36 × 588 | 327 | 192,276 | 8 | 0 | yes | 25,219 | 16,556 | 0 |
ds002603-sub-01-noreconall-epi/bold_task_wm_acq_SeqMm3Tr2000_wf | 80 × 80 × 36 × 160 | 161 | 25,760 | 0 | – | yes | 3,988 | 4,393 | 0 |
ds003592-sub-280-noreconall/bold_ses_1_task_rest_echo_1_wf | 64 × 64 × 43 × 200 | 132 | 26,400 | 0 | – | yes | 6,987 | 7,315 | 10 |
ds004884-sub-M2079-noreconall/bold_ses_217_task_naming40_acq_epfid2_dir_AP_run_2_wf | 64 × 64 × 33 × 60 | 102 | 6,120 | 0 | – | yes | 1,847 | 1,671 | 0 |
ds004884-sub-M2079-noreconall/bold_ses_217_task_rest_acq_epfidp2p2_dir_AP_run_9_wf | 64 × 64 × 34 × 196 | 216 | 42,336 | 0 | – | yes | 5,073 | 5,479 | 0 |
ds004884-sub-M2079-noreconall-syn/bold_ses_217_task_naming40_acq_epfid2_dir_AP_run_2_wf | 64 × 64 × 33 × 60 | 101 | 6,060 | 0 | – | yes | 2,147 | 1,632 | 0 |
ds004884-sub-M2079-noreconall-syn/bold_ses_217_task_rest_acq_epfidp2p2_dir_AP_run_9_wf | 64 × 64 × 34 × 196 | 214 | 41,944 | 0 | – | yes | 6,049 | 5,456 | 0 |
ds004958-sub-fsm64ps-noreconall/bold_task_rest_wf8 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. | 104 × 104 × 72 × 600 | 456 | 273,600 | 8 | 0 | yes | 19,188 | 16,472 | 0 |
ds004958-sub-fsm64ps-noreconall-pepolar/bold_task_rest_wf8 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. | 104 × 104 × 72 × 600 | 452 | 271,200 | 8 | 0 | yes | 15,602 | 16,860 | 0 |
ds005003-sub-RS001-noreconall/bold_ses_1_task_motor_wf8 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. | 64 × 64 × 36 × 80 | 116 | 9,280 | 8 | 0 | yes | 2,056 | 2,093 | 0 |
ds005003-sub-RS001-noreconall/bold_ses_1_task_rest_wf | 64 × 64 × 36 × 80 | 113 | 9,040 | 0 | – | yes | 1,985 | 2,180 | 0 |
ds005003-sub-RS001-noreconall-syn/bold_ses_1_task_motor_wf8 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. | 64 × 64 × 36 × 80 | 117 | 9,360 | 8 | 0 | yes | 2,135 | 2,095 | 0 |
ds005003-sub-RS001-noreconall-syn/bold_ses_1_task_rest_wf | 64 × 64 × 36 × 80 | 111 | 8,880 | 0 | – | yes | 1,906 | 2,174 | 0 |
ds005040-sub-002-noreconall/bold_task_clip_run_1_wf | 64 × 64 × 40 × 63 | 123 | 7,749 | 0 | – | yes | 1,537 | 1,828 | 0 |
ds005454-sub-16-noreconall/bold_task_rest_wf | 160 × 160 × 96 × 280 | 335 | 93,800 | 0 | – | yes | 30,515 | 8,240 | 0 |
ds005752-sub-ON01016-noreconall/bold_ses_01_task_rest_dir_forward_wf | 64 × 64 × 48 × 200 | 142 | 28,400 | 0 | – | yes | 4,951 | 5,825 | 0 |
ds005752-sub-ON01016-noreconall/bold_ses_01_task_rest_dir_reverse_wf | 64 × 64 × 48 × 10 | 63 | 630 | 0 | – | yes | 237 | 271 | 0 |
ds005752-sub-ON01016-noreconall-direct/bold_ses_01_task_rest_dir_forward_wf | 64 × 64 × 48 × 200 | 140 | 28,000 | 0 | – | yes | 4,947 | 5,996 | 0 |
ds005752-sub-ON01016-noreconall-direct/bold_ses_01_task_rest_dir_reverse_wf | 64 × 64 × 48 × 10 | 63 | 630 | 0 | – | yes | 237 | 294 | 0 |
ds006736-sub-004-noreconall/bold_task_freeRecall_wf | 74 × 64 × 58 × 69 | 138 | 9,522 | 0 | – | yes | 1,919 | 1,875 | 0 |
The cache entry confounds/workflows (confounds_run.py): each workflow run end to end by the replica’s bold_confounds from its own inputs (its nodes’ pickled inputs), in reproduce with fMRIPrep 25.2.6’s releases (fMRIPrep 25.2.6, niworkflows 1.15.2, numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3), its desc-confounds_timeseries.tsv compared with the run’s own file cell by cell as text, the columns matched by name (their order is niworkflows’ set order, which changes from process to process: PIPE-43, never reproduced). Where a cell differs, the run is made again with fMRIPrep’s own motion table in place of the replica’s (glibc’s atan2, D4). The step: the replica in reproduce with the newest releases (fMRIPrep 25.2.7, niworkflows 1.15.4, numpy 2.5.4, SciPy 1.18.1, pandas 3.0.6) against fMRIPrep 25.2.6’s: the originals’ own step (each node of it the newest originals’ bit for bit: the records). fix, the default: the replica’s corrections (newest releases) against reproduce; a column negated: GLIBC-07’s cosine (expected).
Every column of the 36 files, by name (numbered columns by family)
| Column | Files | Columns | Cells | ≠ fMRIPrep's file | The step: changed | fix: changed | fix: negated |
|---|---|---|---|---|---|---|---|
rot_xfMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 9e-14 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,269 | 7,153 | 0 |
rot_x_derivative1fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.9e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,272 | 7,317 | 0 |
rot_x_derivative1_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 7.5e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,277 | 7,302 | 0 |
rot_x_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 7.2e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,269 | 7,265 | 0 |
rot_yfMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.8e-13 of the column’s largest value). | 36 | 36 | 7,431 | 12 | 7,367 | 6,564 | 0 |
rot_y_derivative1fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 4.6e-13 of the column’s largest value). | 36 | 36 | 7,431 | 24 | 7,329 | 7,052 | 0 |
rot_y_derivative1_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 6.5e-13 of the column’s largest value). | 36 | 36 | 7,431 | 22 | 7,312 | 7,063 | 0 |
rot_y_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 3.3e-13 of the column’s largest value). | 36 | 36 | 7,431 | 11 | 7,308 | 6,856 | 0 |
rot_zfMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.8e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,410 | 6,979 | 0 |
rot_z_derivative1fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 5.9e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,354 | 7,213 | 0 |
rot_z_derivative1_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 9.3e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,354 | 7,209 | 0 |
rot_z_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 4.2e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,365 | 7,115 | 0 |
trans_xfMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 4.1e-15 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,409 | 5,950 | 0 |
trans_x_derivative1fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.3e-14 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,372 | 7,100 | 0 |
trans_x_derivative1_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 3.4e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,374 | 6,886 | 0 |
trans_x_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.6e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,409 | 6,895 | 0 |
trans_yfMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.8e-15 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,431 | 5,669 | 0 |
trans_y_derivative1fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 5.3e-15 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,382 | 6,876 | 0 |
trans_y_derivative1_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.7e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,378 | 7,129 | 0 |
trans_y_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 6.5e-14 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,431 | 7,036 | 0 |
trans_zfMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.4e-15 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,431 | 4,526 | 0 |
trans_z_derivative1fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 4e-15 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,388 | 6,931 | 0 |
trans_z_derivative1_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 3e-14 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,381 | 6,909 | 0 |
trans_z_power2fMRIPrep’s FSLMotionParams (the head-motion matrices) and niworkflows’ expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: fMRIPrep 25.2.7’s FSLMotionParams (the centre of gravity of the reference minus its minimum, the matrix not transposed, the angles negated: confounds-motion.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 2e-14 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 7,431 | 5,895 | 0 |
framewise_displacementfMRIPrep’s FramewiseDisplacement, FSLRMSDeviation and niworkflows’ spikes. reproduce: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The step: follows the motion parameters of fMRIPrep 25.2.7 (FD, the RMS deviation’s inputs, the spikes). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 8.5e-14 of the column’s largest value). | 36 | 36 | 7,431 | 24 | 7,384 | 7,350 | 0 |
motion_outlierXXfMRIPrep’s FramewiseDisplacement, FSLRMSDeviation and niworkflows’ spikes. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: follows the motion parameters of fMRIPrep 25.2.7 (FD, the RMS deviation’s inputs, the spikes). The step adds 2 such columns and removes 4 over the 36 files (FD crosses its threshold at other volumes). | 31 | 540 | 95,679 | 0 | 188 | 0 | 0 |
rmsdfMRIPrep’s FramewiseDisplacement, FSLRMSDeviation and niworkflows’ spikes. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 9.5e-16 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 6,191 | 0 |
dvarsnipype’s ComputeDVARS. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.8e-16 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 45 | 0 |
std_dvarsnipype’s ComputeDVARS. reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: SciPy 1.18’s FFT (ducc0 for pocketfft) and numpy 2.5’s sums (confounds-dvars.md). fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.9e-16 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 31 | 292 | 0 |
a_comp_cor_XXnipype’s ACompCor and TCompCor (niworkflows’ robust wrappers). reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: numpy 2.5.4’s SVD under LAPACK 3.12.1 (DGESDD’s reflectors: the components’ last digits, and the sign of a vector whose sign the SVD leaves free: confounds-blas.md). fix, the default: GLIBC-07: the default’s cosine basis takes the correctly rounded cos (LARMORX’s rule of correctly rounded maths; reproduce glibc’s cos), whose last bits decide the sign of an SVD vector (arbitrary in itself) and the 10th decimal of a few values; a column and its negative remove the same signal, so a model gives the same cleaned data (confounds-compcor.md) (37 columns negated, 32 by one unit in the 10th decimal). | 36 | 1,478 | 513,328 | 0 | 14,825 | 8,800 | 37 |
c_comp_cor_XXnipype’s ACompCor and TCompCor (niworkflows’ robust wrappers). reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: numpy 2.5.4’s SVD under LAPACK 3.12.1 (DGESDD’s reflectors: the components’ last digits, and the sign of a vector whose sign the SVD leaves free: confounds-blas.md). fix, the default: GLIBC-07: the default’s cosine basis takes the correctly rounded cos (LARMORX’s rule of correctly rounded maths; reproduce glibc’s cos), whose last bits decide the sign of an SVD vector (arbitrary in itself) and the 10th decimal of a few values; a column and its negative remove the same signal, so a model gives the same cleaned data (confounds-compcor.md) (12 columns negated, 1 by one unit in the 10th decimal). | 36 | 600 | 172,533 | 0 | 5,508 | 2,881 | 12 |
edge_comp_XXnipype’s ACompCor and TCompCor (niworkflows’ robust wrappers). reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: numpy 2.5.4’s SVD under LAPACK 3.12.1 (DGESDD’s reflectors: the components’ last digits, and the sign of a vector whose sign the SVD leaves free: confounds-blas.md). fix, the default: GLIBC-07: the default’s cosine basis takes the correctly rounded cos (LARMORX’s rule of correctly rounded maths; reproduce glibc’s cos), whose last bits decide the sign of an SVD vector (arbitrary in itself) and the 10th decimal of a few values; a column and its negative remove the same signal, so a model gives the same cleaned data (confounds-compcor.md) (11 columns negated, 4 by one unit in the 10th decimal). | 36 | 834 | 178,044 | 0 | 13,903 | 2,164 | 11 |
t_comp_cor_XXnipype’s ACompCor and TCompCor (niworkflows’ robust wrappers). reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: numpy 2.5.4’s SVD under LAPACK 3.12.1 (DGESDD’s reflectors: the components’ last digits, and the sign of a vector whose sign the SVD leaves free: confounds-blas.md). | 36 | 149 | 39,093 | 0 | 1,109 | 0 | 0 |
w_comp_cor_XXnipype’s ACompCor and TCompCor (niworkflows’ robust wrappers). reproduce: fMRIPrep’s cells, as text, in every file that has the column. The step: numpy 2.5.4’s SVD under LAPACK 3.12.1 (DGESDD’s reflectors: the components’ last digits, and the sign of a vector whose sign the SVD leaves free: confounds-blas.md). fix, the default: GLIBC-07: the default’s cosine basis takes the correctly rounded cos (LARMORX’s rule of correctly rounded maths; reproduce glibc’s cos), whose last bits decide the sign of an SVD vector (arbitrary in itself) and the 10th decimal of a few values; a column and its negative remove the same signal, so a model gives the same cleaned data (confounds-compcor.md) (0 columns negated, 16 by one unit in the 10th decimal). | 36 | 1,453 | 505,275 | 0 | 43,530 | 17 | 0 |
cosineXXtCompCor’s pre-filter file (the cosine basis, the dummy scans). reproduce: fMRIPrep’s cells, as text, in every file that has the column. | 33 | 174 | 40,209 | 0 | 0 | 0 | 0 |
non_steady_state_outlierXXtCompCor’s pre-filter file (the cosine basis, the dummy scans). reproduce: fMRIPrep’s cells, as text, in every file that has the column. | 16 | 23 | 6,436 | 0 | 0 | 0 | 0 |
csfniworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 3.4e-16 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 1,009 | 0 |
csf_derivative1niworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 3.5e-14 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 2,775 | 0 |
csf_derivative1_power2niworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 4.2e-14 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 2,968 | 0 |
csf_power2niworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 9.3e-16 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 2,213 | 0 |
csf_wmniworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 4.8e-16 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 1,009 | 0 |
global_signalniworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 3.3e-16 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 610 | 0 |
global_signal_derivative1niworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.1e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 2,713 | 0 |
global_signal_derivative1_power2niworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1.1e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 3,190 | 0 |
global_signal_power2niworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 7.6e-16 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 1,407 | 0 |
tcompcorniworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 3.7e-16 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 1,085 | 0 |
white_matterniworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 4.8e-16 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 920 | 0 |
white_matter_derivative1niworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 1e-13 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 3,001 | 0 |
white_matter_derivative1_power2niworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 9.6e-14 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 3,366 | 0 |
white_matter_power2niworkflows’ SignalExtraction and its expansion. reproduce: fMRIPrep’s cells, as text, in every file that has the column. fix, the default: PD-01, rule A: fix reads the intermediate tables’ numbers correctly rounded where pandas’ parser is not (its last bit; confounds-pandas.md), so a value computed from them can move by its last bits (at most 7.9e-16 of the column’s largest value). | 36 | 36 | 7,431 | 0 | 0 | 1,578 | 0 |
The same runs as the table above, summed by column. The verdict is the port’s (reproduce with fMRIPrep 25.2.6’s releases against fMRIPrep’s own files); the step is the originals’ own change from fMRIPrep 25.2.6’s releases to the newest (a yardstick, no verdict); the default’s changes are the corrections by default named with each run (PD-01 and GLIBC-07 in the tables; PIPE-41 adds the five signals’ entries to each JSON). Negated: the confounds files keep the retained components only; over the CompCor nodes’ own components files (every component, 108 files; the default against reproduce, both with the newest releases: the findings count the same with fMRIPrep 25.2.6’s releases, whose LAPACK leaves other signs free) the default changes 18 files, negates 60 columns in 5 of them and moves 53 columns by one unit in the 10th decimal (GLIBC-07: docs/findings/confounds-compcor.md).
Every node file of the 36 workflows (reproduce, fMRIPrep 25.2.6’s releases) against fMRIPrep’s node folders
| Check | Files | Equal | Differ |
|---|---|---|---|
| node files = fMRIPrep’s (others) | 970 | 970 | 0 |
| aCompCor mask values (Gaussian, glibc exp/log): header and thresholds equal | 218 | 218 | 0 |
| motion_params differs only in rot_* (D4) | 12 | 12 | 0 |
| tcc_metadata_fmt = fMRIPrep’s | 36 | 36 | 0 |
| acc_metadata_fmt = fMRIPrep’s | 36 | 36 | 0 |
| crowncc_metadata_fmt = fMRIPrep’s | 36 | 36 | 0 |
| merge_confound_metadata2 = fMRIPrep’s | 36 | 36 | 0 |
| node files = fMRIPrep’s (motion chain) | 74 | 74 | 0 |
| node files = fMRIPrep’s (motion chain, by column name) | 50 | 50 | 0 |
| The replica files fMRIPrep’s nipype deleted | acc_msk_bin 108, acompcor 36, crowncompcor 36 | – | – |
| aCompCor mask voxels with other values (glibc exp/log) | acc_msk_bin 2, acc_msk_brain 4,714, acc_msk_tfm 5,298 | – | – |
| motion chain cells changed (D4 atan2) | concat 36, fdisp 24, model_expansion 96, motion_params 15, spike_regressors 93 | – | – |
check_workflow.py’s compare_nodes and compare_dicts on each run: every file the replica writes against the file of the same name in fMRIPrep’s node folder (decompressed; the motion chain’s tables cell by cell by column name). The aCompCor masks after -n Gaussian hold other values where vnl’s erf goes through glibc’s exp and log in fMRIPrep’s ANTs and correctly rounded ones in the replica’s ANTs port (the Gaussian example below: ANTs 2.6.5 with CORE-MATH gives the replica’s bytes); no voxel crosses the masks’ 0.5 and 0.99, so aCompCor reads the same masks. nipype deletes the files of outputs nothing reads; the replica keeps them.
Four runs of fMRIPrep
Each example shows one confounds workflow of fMRIPrep’s runs: the CompCor masks over the BOLD reference (fMRIPrep’s files, the replica in reproduce and in fix), the confounds’ time series, and every column of its table with its verdict. The mask shown is chosen above the viewer.
ds000005: an adult, 3 T, single-band
Worked example
fMRIPrep's confounds of a real BOLD run: an adult, 3 T, single-band
What this shows: In reproduce, with fMRIPrep 25.2.6’s releases, the replica writes every CompCor mask of fMRIPrep’s run, voxel for voxel and header for header, and its confounds table holds every column of fMRIPrep’s file, cell for cell, but 8 cells of the motion chain, where glibc’s atan2 rounds an angle otherwise than the replica’s correctly rounded one (with fMRIPrep’s own motion table, none differs); its JSON byte for byte. In fix, its default, with the newest releases, the masks are the same bytes, at 1 and 8 threads alike; the table below says, column by column, what the newest releases and the corrections change.
- Replica against fMRIPrep's outputBit-identical
fMRIPrep 25.2.6’s confounds workflow (bold_task_mixedgamblestask_run_01_wf of OpenNeuro ds000005 (Tom et al., 2007), snapshot 00001 sub-01) run end to end by the replica from the workflow’s own inputs: in reproduce with fMRIPrep 25.2.6’s releases, compared with the run’s own files (its CompCor masks drawn over the BOLD reference, and its confounds table column by column), with the newest releases (the originals’ own step) and in fix, its default (the corrections), at 8 threads and, in fix, at 1.
- Data
- OpenNeuro ds000005 (Tom et al., 2007), snapshot 00001, sub-01 (PDDL-1.0); shown inside fMRIPrep’s BOLD mask, dilated by three voxels
- Workflow
- fMRIPrep 25.2.6,
bold_task_mixedgamblestask_run_01_wf/bold_confounds_wf(andds_confounds): 64 × 64 × 34 voxels of 3.12 × 3.12 × 4 mm, 240 volumes (float32), TR 2.0 s, 1 dummy scan(s) - Confounds file
sub-01_task-mixedgamblestask_run-01_desc-confounds_timeseries.tsv: 222 columns, 53,280 cells; the replica’s file has fMRIPrep’s name (yes) and another column order (PIPE-43); its JSON byte for byte- Verdict of the file
- divergence: 8 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of
reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. - Programs
- The replica 0.1.0rc1 (
bold_confounds, experimental) at commit5fbdd0f07f6f, in numpy 2.5.4 (OpenBLAS 0.3.34, LAPACK 3.12.1, OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6, nipype 1.12.0, niworkflows 1.15.4, nitime 0.12.1, nibabel 5.4.2 (oracles/python-sdcflows-2.17), fMRIPrep 25.2.7 from its tag; fMRIPrep’s files from its run (fMRIPrep 25.2.6 in its image (niworkflows 1.15.2, nipype 1.10, numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3): the runs’ own files) - Notices
- e3: GLIBC-07
reproduce3, PD-01reproduce8, PIPE-41reproduce1, PIPE-43 fix 1; sp: GLIBC-07reproduce3, PD-01reproduce8, PIPE-41reproduce1, PIPE-43 fix 1; new: GLIBC-07reproduce3, PD-01reproduce8, PIPE-41reproduce1, PIPE-43 fix 1; fix: GLIBC-07 fix 3, PD-01 fix 8, PIPE-41 fix 1, PIPE-43 fix 1; fix-1: GLIBC-07 fix 3, PD-01 fix 8, PIPE-41 fix 1, PIPE-43 fix 1 - Grid
- 64 × 64 × 34 voxels of 3.12 × 3.12 × 4 mm; shown: 57 × 59 × 34 around the brain
Data shown
- Mixed-gambles task (OpenNeuro ds000005, version 00001), Open Data Commons Public Domain Dedication and Licence 1.0.




Bit-identical139,264 of 139,264 voxels equal
Bit-identical139,264 of 139,264 voxels equal
Axial slice 15 of 34; the masks drawn over the BOLD reference. The interactive viewer needs JavaScript; the metrics below are complete without it.
- fMRIPrep's output (fMRIPrep 25.2.6's run)
- replica, reproduce
- replica, fix (the default)
Each panel one column of the confounds table (n/a cells, the first volume of FD and DVARS, left out). The file holds 222 columns; 0 motion-outlier columns (one per volume over the FD or standardised-DVARS threshold). The lines of fMRIPrep’s file and of the replica in reproduce coincide; the table below counts every cell.
This file’s every column (numbered columns by family): the replica against fMRIPrep’s file, the originals’ step, the default
| Column | Columns | Cells | ≠ fMRIPrep's | largest |Δ| | The step: changed | largest |Δ| | fix: changed | largest |Δ| |
|---|---|---|---|---|---|---|---|---|
rot_xfix, the default: PD-01’s correctly rounded reading (≤ 6.5e-14 of the column’s largest value). | 1 | 240 | 0 | – | 160 | 8.7e−7 | 232 | 9.93e−17 |
rot_x_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 1.9e-13 of the column’s largest value). | 1 | 240 | 0 | – | 174 | 3.66e−7 | 237 | 1.89e−16 |
rot_x_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.5e-13 of the column’s largest value). | 1 | 240 | 0 | – | 179 | 4.95e−10 | 237 | 1.43e−19 |
rot_x_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.2e-13 of the column’s largest value). | 1 | 240 | 0 | – | 160 | 2.14e−9 | 235 | 2.66e−19 |
rot_yreproduce: rot_y: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 6.2e-14 of the column’s largest value). | 1 | 240 | 1 | 3e−16 | 203 | 9.28e−7 | 210 | 9.99e−17 |
rot_y_derivative1reproduce: rot_y_derivative1: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 3.7e-13 of the column’s largest value). | 1 | 240 | 2 | 3e−16 | 211 | 7.98e−7 | 215 | 1.85e−16 |
rot_y_derivative1_power2reproduce: rot_y_derivative1_power2: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 2.2e-13 of the column’s largest value). | 1 | 240 | 2 | 1.29e−19 | 207 | 2.38e−10 | 216 | 5.47e−20 |
rot_y_power2reproduce: rot_y_power2: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 1.2e-13 of the column’s largest value). | 1 | 240 | 1 | 3.29e−19 | 187 | 2.14e−9 | 214 | 3.03e−19 |
rot_zfix, the default: PD-01’s correctly rounded reading (≤ 1.3e-13 of the column’s largest value). | 1 | 240 | 0 | – | 238 | 1.91e−6 | 180 | 9.89e−17 |
rot_z_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 5.9e-13 of the column’s largest value). | 1 | 240 | 0 | – | 231 | 1.14e−6 | 195 | 1.67e−16 |
rot_z_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 6.5e-13 of the column’s largest value). | 1 | 240 | 0 | – | 229 | 4.95e−10 | 197 | 5.3e−20 |
rot_z_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.2e-13 of the column’s largest value). | 1 | 240 | 0 | – | 215 | 2.14e−9 | 191 | 1.25e−19 |
trans_xfix, the default: PD-01’s correctly rounded reading (≤ 1.7e-15 of the column’s largest value). | 1 | 240 | 0 | – | 228 | 0.0002 | 167 | 9.89e−17 |
trans_x_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 1.3e-14 of the column’s largest value). | 1 | 240 | 0 | – | 234 | 0.00034 | 206 | 2.41e−16 |
trans_x_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.8e-13 of the column’s largest value). | 1 | 240 | 0 | – | 234 | 4.73e−6 | 212 | 9.91e−17 |
trans_x_power2fix, the default: PD-01’s correctly rounded reading (≤ 3.2e-14 of the column’s largest value). | 1 | 240 | 0 | – | 228 | 1.46e−5 | 205 | 1.04e−16 |
trans_yfix, the default: PD-01’s correctly rounded reading (≤ 8.8e-16 of the column’s largest value). | 1 | 240 | 0 | – | 240 | 0.000259 | 218 | 9.71e−17 |
trans_y_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 2.4e-15 of the column’s largest value). | 1 | 240 | 0 | – | 239 | 0.000481 | 224 | 1.87e−16 |
trans_y_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.7e-14 of the column’s largest value). | 1 | 240 | 0 | – | 239 | 4.66e−5 | 236 | 1.1e−16 |
trans_y_power2fix, the default: PD-01’s correctly rounded reading (≤ 8.9e-15 of the column’s largest value). | 1 | 240 | 0 | – | 240 | 5.07e−5 | 239 | 1.09e−16 |
trans_zfix, the default: PD-01’s correctly rounded reading (≤ 4.2e-16 of the column’s largest value). | 1 | 240 | 0 | – | 240 | 0.000212 | 191 | 9.71e−17 |
trans_z_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 1.7e-15 of the column’s largest value). | 1 | 240 | 0 | – | 239 | 0.000274 | 228 | 1.8e−16 |
trans_z_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 9.5e-15 of the column’s largest value). | 1 | 240 | 0 | – | 239 | 2.31e−5 | 223 | 1.08e−16 |
trans_z_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.3e-15 of the column’s largest value). | 1 | 240 | 0 | – | 240 | 8.38e−5 | 237 | 1.25e−16 |
framewise_displacementreproduce: framewise_displacement: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 6e-14 of the column’s largest value). | 1 | 240 | 2 | 1.5e−14 | 239 | 0.000585 | 239 | 1.09e−14 |
rmsdfix, the default: PD-01’s correctly rounded reading (≤ 8e-16 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 218 | 9.71e−17 |
dvars | 1 | 240 | 0 | – | 0 | – | 0 | – |
std_dvarsfix, the default: PD-01’s correctly rounded reading (≤ 1.9e-16 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 4 | 2.22e−16 |
a_comp_cor_XXfix, the default: 0 columns negated, 2 by one unit in the 10th decimal (GLIBC-07, expected). | 56 | 13,440 | 0 | – | 4 | 1e−10 | 2 | 1e−10 |
c_comp_cor_XX | 29 | 6,960 | 0 | – | 717 | 0.449 | 0 | – |
edge_comp_XX | 24 | 5,760 | 0 | – | 0 | – | 0 | – |
t_comp_cor_XX | 8 | 1,920 | 0 | – | 239 | 0.412 | 0 | – |
w_comp_cor_XXfix, the default: 0 columns negated, 2 by one unit in the 10th decimal (GLIBC-07, expected). | 56 | 13,440 | 0 | – | 2 | 1e−10 | 3 | 1e−10 |
cosineXX | 6 | 1,440 | 0 | – | 0 | – | 0 | – |
non_steady_state_outlierXX | 1 | 240 | 0 | – | 0 | – | 0 | – |
csf | 1 | 240 | 0 | – | 0 | – | 0 | – |
csf_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 2.5e-17 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 74 | 4.44e−16 |
csf_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.1e-17 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 86 | 3.55e−15 |
csf_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.4e-16 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 59 | 5.82e−11 |
csf_wm | 1 | 240 | 0 | – | 0 | – | 0 | – |
global_signal | 1 | 240 | 0 | – | 0 | – | 0 | – |
global_signal_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 9.3e-17 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 93 | 2.22e−16 |
global_signal_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 7.8e-17 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 138 | 4.44e−16 |
global_signal_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.1e-16 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 12 | 5.82e−11 |
tcompcorfix, the default: PD-01’s correctly rounded reading (≤ 1.2e-16 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 49 | 5.68e−14 |
white_matter | 1 | 240 | 0 | – | 0 | – | 0 | – |
white_matter_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 9.8e-17 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 87 | 2.22e−16 |
white_matter_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.7e-16 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 146 | 8.88e−16 |
white_matter_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.9e-16 of the column’s largest value). | 1 | 240 | 0 | – | 0 | – | 12 | 5.82e−11 |
Cells compared as text (the 10 decimals pandas and numpy write), the columns by name. ≠ fMRIPrep’s: the replica in reproduce with fMRIPrep 25.2.6’s releases against the run’s own file; the step: reproduce with the newest releases against fMRIPrep 25.2.6’s (the originals’ own change, no verdict); the default: the replica’s corrections against reproduce, both with the newest releases.
Every comparison, computed in float64 over every voxel of the outputs
| Comparison (A − B), verdict and reason | Output | max |Δ| | Values that differ | RMS of Δ | max |Δ| ÷ range | 1 − NCC |
|---|---|---|---|---|---|---|
| The replica against fMRIPrep's output | ||||||
| replica, reproduce − fMRIPrep's outputBit-identicalEvery mask fMRIPrep’s run wrote, voxel for voxel and header for header (the masks’ values above 0.99 kept, as fMRIPrep’s Binarize keeps them). | every output | identical: 0 of 139,264 values differ | ||||
replica, fix − fMRIPrep's outputBit-identicalIn fix, its default (the newest releases, the corrections), the replica gives the same masks: no correction reaches them on this run. | every output | identical: 0 of 139,264 values differ | ||||
| The replica against itself: thread counts, fix and reproduce | ||||||
replica: fix − reproduceBit-identicalNo correction by default changes a mask here: fix and reproduce give the same bytes. | every output | identical: 0 of 139,264 values differ | ||||
| replica, fix: 1 thread − 8 threadsBit-identicalThe replica’s result does not depend on its thread count: the same bytes at 1 and 8 threads (every file of the run, the confounds table included). | every output | identical: 0 of 139,264 values differ | ||||
Δ is A − B over every voxel of the masks’ files (float64 masks: the probabilities above 0.99 kept, as fMRIPrep’s Binarize writes them; uint8 for tCompCor’s and the crown). The confounds table is compared cell by cell as text in the table below. The records’ threshold: in reproduce, every array, file, table cell, header, warning, message and exit status bit-identical to the originals’ (or fMRIPrep’s own files).
Provenance of this example
- Data: OpenNeuro ds000005 (Tom et al., 2007), snapshot 00001, sub-01, PDDL-1.0
- Inputs: fMRIPrep’s confounds workflow ds000005-sub-01-noreconall/bold_task_mixedgamblestask_run_01_wf (sha256 47ee409dbe6815c5…)
- Runs: figures cache entry
confounds/workflows, computed 2026-11-13T08:32:56+00:00 with larmorx5fbdd0f07f6f; numpy 2.5.4 (OpenBLAS 0.3.34, LAPACK 3.12.1, OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6, nipype 1.12.0, niworkflows 1.15.4, nitime 0.12.1, nibabel 5.4.2 (oracles/python-sdcflows-2.17), fMRIPrep 25.2.7 from its tag; fMRIPrep 25.2.6 in its image (niworkflows 1.15.2, nipype 1.10, numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3): the runs’ own files; every command, its environment, threads and exit status are in itsRUN_INFO.json - Workflow:
ds000005-sub-01-noreconall/work/fmriprep_25_2_wf/sub_01_wf/bold_task_mixedgamblestask_run_01_wf/bold_confounds_wf
Every number above is in confounds.json, exported by python -m larmorx_validation evidence confounds.
ds000256: a child’s run with strong head motion
Worked example
fMRIPrep's confounds of a real BOLD run: a 7.5-year-old's rest run with strong head motion, 3 T
What this shows: In reproduce, with fMRIPrep 25.2.6’s releases, the replica writes every CompCor mask of fMRIPrep’s run, voxel for voxel and header for header, and its confounds table holds every column of fMRIPrep’s file, cell for cell, but 5 cells of the motion chain, where glibc’s atan2 rounds an angle otherwise than the replica’s correctly rounded one (with fMRIPrep’s own motion table, none differs); its JSON byte for byte. In fix, its default, with the newest releases, the masks are the same bytes, at 1 and 8 threads alike; the table below says, column by column, what the newest releases and the corrections change.
- Replica against fMRIPrep's outputBit-identical
fMRIPrep 25.2.6’s confounds workflow (bold_task_restfixationnofeedback_wf of OpenNeuro ds000256 (Greene et al., 2018) sub-CTS203) run end to end by the replica from the workflow’s own inputs: in reproduce with fMRIPrep 25.2.6’s releases, compared with the run’s own files (its CompCor masks drawn over the BOLD reference, and its confounds table column by column), with the newest releases (the originals’ own step) and in fix, its default (the corrections), at 8 threads and, in fix, at 1.
- Data
- OpenNeuro ds000256 (Greene et al., 2018), sub-CTS203 (PDDL-1.0); shown inside fMRIPrep’s BOLD mask, dilated by three voxels
- Workflow
- fMRIPrep 25.2.6,
bold_task_restfixationnofeedback_wf/bold_confounds_wf(andds_confounds): 64 × 64 × 32 voxels of 4 × 4 × 4 mm, 164 volumes (float32), TR 2.5 s, 0 dummy scan(s) - Confounds file
sub-CTS203_task-restfixationnofeedback_desc-confounds_timeseries.tsv: 177 columns, 29,028 cells; the replica’s file has fMRIPrep’s name (yes) and another column order (PIPE-43); its JSON byte for byte- Verdict of the file
- divergence: 5 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of
reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. - Programs
- The replica 0.1.0rc1 (
bold_confounds, experimental) at commit5fbdd0f07f6f, in numpy 2.5.4 (OpenBLAS 0.3.34, LAPACK 3.12.1, OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6, nipype 1.12.0, niworkflows 1.15.4, nitime 0.12.1, nibabel 5.4.2 (oracles/python-sdcflows-2.17), fMRIPrep 25.2.7 from its tag; fMRIPrep’s files from its run (fMRIPrep 25.2.6 in its image (niworkflows 1.15.2, nipype 1.10, numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3): the runs’ own files) - Notices
- e3: GLIBC-07
reproduce3, PD-01reproduce9, PIPE-41reproduce1, PIPE-43 fix 1; sp: GLIBC-07reproduce3, PD-01reproduce9, PIPE-41reproduce1, PIPE-43 fix 1; new: GLIBC-07reproduce3, PD-01reproduce9, PIPE-41reproduce1, PIPE-43 fix 1; fix: GLIBC-07 fix 3, PD-01 fix 9, PIPE-41 fix 1, PIPE-43 fix 1; fix-1: GLIBC-07 fix 3, PD-01 fix 9, PIPE-41 fix 1, PIPE-43 fix 1 - Grid
- 64 × 64 × 32 voxels of 4 × 4 × 4 mm; shown: 46 × 54 × 32 around the brain
Data shown
- Behavioral interventions for reducing head motion in children (OpenNeuro ds000256, version 00002), Open Data Commons Public Domain Dedication and Licence 1.0.




Bit-identical131,072 of 131,072 voxels equal
Bit-identical131,072 of 131,072 voxels equal
Axial slice 19 of 32; the masks drawn over the BOLD reference. The interactive viewer needs JavaScript; the metrics below are complete without it.
- fMRIPrep's output (fMRIPrep 25.2.6's run)
- replica, reproduce
- replica, fix (the default)
Each panel one column of the confounds table (n/a cells, the first volume of FD and DVARS, left out). The file holds 177 columns; 92 motion-outlier columns (one per volume over the FD or standardised-DVARS threshold). The lines of fMRIPrep’s file and of the replica in reproduce coincide; the table below counts every cell.
This file’s every column (numbered columns by family): the replica against fMRIPrep’s file, the originals’ step, the default
| Column | Columns | Cells | ≠ fMRIPrep's | largest |Δ| | The step: changed | largest |Δ| | fix: changed | largest |Δ| |
|---|---|---|---|---|---|---|---|---|
rot_xfix, the default: PD-01’s correctly rounded reading (≤ 2.3e-16 of the column’s largest value). | 1 | 164 | 0 | – | 164 | 0.00342 | 155 | 9.71e−17 |
rot_x_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 4.3e-16 of the column’s largest value). | 1 | 164 | 0 | – | 163 | 0.00353 | 160 | 1.96e−16 |
rot_x_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 6.6e-16 of the column’s largest value). | 1 | 164 | 0 | – | 163 | 0.00241 | 162 | 1.35e−16 |
rot_x_power2fix, the default: PD-01’s correctly rounded reading (≤ 7e-16 of the column’s largest value). | 1 | 164 | 0 | – | 164 | 0.000387 | 163 | 1.25e−16 |
rot_yreproduce: rot_y: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 1.9e-15 of the column’s largest value). | 1 | 164 | 1 | 1.98e−16 | 164 | 0.0339 | 163 | 9.76e−17 |
rot_y_derivative1reproduce: rot_y_derivative1: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 4.8e-15 of the column’s largest value). | 1 | 164 | 2 | 1.01e−16 | 163 | 0.0339 | 162 | 2.41e−16 |
rot_y_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 3.8e-14 of the column’s largest value). | 1 | 164 | 0 | – | 163 | 0.00327 | 162 | 9.32e−17 |
rot_y_power2fix, the default: PD-01’s correctly rounded reading (≤ 3.9e-14 of the column’s largest value). | 1 | 164 | 0 | – | 164 | 0.000388 | 164 | 1.04e−16 |
rot_zfix, the default: PD-01’s correctly rounded reading (≤ 4.9e-16 of the column’s largest value). | 1 | 164 | 0 | – | 164 | 0.00505 | 158 | 9.89e−17 |
rot_z_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 9.4e-16 of the column’s largest value). | 1 | 164 | 0 | – | 163 | 0.00632 | 162 | 1.99e−16 |
rot_z_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.2e-15 of the column’s largest value). | 1 | 164 | 0 | – | 163 | 0.00042 | 161 | 1.01e−16 |
rot_z_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.6e-15 of the column’s largest value). | 1 | 164 | 0 | – | 164 | 0.000388 | 162 | 1.06e−16 |
trans_xfix, the default: PD-01’s correctly rounded reading (≤ 1.1e-16 of the column’s largest value). | 1 | 164 | 0 | – | 164 | 5.67 | 73 | 4.44e−16 |
trans_x_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 1e-16 of the column’s largest value). | 1 | 164 | 0 | – | 163 | 5.8 | 125 | 4.44e−16 |
trans_x_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 2e-16 of the column’s largest value). | 1 | 164 | 0 | – | 163 | 54 | 130 | 3.55e−15 |
trans_x_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.8e-16 of the column’s largest value). | 1 | 164 | 0 | – | 164 | 50.1 | 95 | 2.66e−15 |
trans_yfix, the default: PD-01’s correctly rounded reading (≤ 7.1e-17 of the column’s largest value). | 1 | 164 | 0 | – | 164 | 23.4 | 67 | 1.78e−15 |
trans_y_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 1.4e-16 of the column’s largest value). | 1 | 164 | 0 | – | 163 | 23.2 | 126 | 3.55e−15 |
trans_y_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 7e-17 of the column’s largest value). | 1 | 164 | 0 | – | 163 | 100 | 131 | 4.26e−14 |
trans_y_power2fix, the default: PD-01’s correctly rounded reading (≤ 9.2e-17 of the column’s largest value). | 1 | 164 | 0 | – | 164 | 177 | 114 | 5.68e−14 |
trans_zfix, the default: PD-01’s correctly rounded reading (≤ 8.7e-17 of the column’s largest value). | 1 | 164 | 0 | – | 164 | 7.52 | 58 | 1.78e−15 |
trans_z_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 1.3e-16 of the column’s largest value). | 1 | 164 | 0 | – | 163 | 7.89 | 121 | 2.22e−15 |
trans_z_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 5.8e-17 of the column’s largest value). | 1 | 164 | 0 | – | 163 | 265 | 122 | 1.78e−14 |
trans_z_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.4e-16 of the column’s largest value). | 1 | 164 | 0 | – | 164 | 172 | 83 | 5.68e−14 |
framewise_displacementreproduce: framewise_displacement: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 2.5e-16 of the column’s largest value). | 1 | 164 | 2 | 1.07e−14 | 163 | 11.8 | 155 | 1.42e−14 |
motion_outlierXX | 92 | 15,088 | 0 | – | 0 | – | 0 | – |
rmsdfix, the default: PD-01’s correctly rounded reading (≤ 1.2e-16 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 76 | 3.55e−15 |
dvarsfix, the default: PD-01’s correctly rounded reading (≤ 3e-17 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 5 | 1.42e−14 |
std_dvarsfix, the default: PD-01’s correctly rounded reading (≤ 1.6e-17 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 3 | 1.11e−16 |
a_comp_cor_XX | 3 | 492 | 0 | – | 0 | – | 0 | – |
c_comp_cor_XX | 7 | 1,148 | 0 | – | 0 | – | 0 | – |
edge_comp_XX | 24 | 3,936 | 0 | – | 328 | 1.07 | 0 | – |
t_comp_cor_XX | 1 | 164 | 0 | – | 0 | – | 0 | – |
w_comp_cor_XX | 3 | 492 | 0 | – | 0 | – | 0 | – |
cosineXX | 5 | 820 | 0 | – | 0 | – | 0 | – |
csffix, the default: PD-01’s correctly rounded reading (≤ 1.1e-16 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 46 | 5.68e−14 |
csf_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 1.1e-15 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 94 | 1.14e−13 |
csf_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 1e-15 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 92 | 1.09e−11 |
csf_power2fix, the default: PD-01’s correctly rounded reading (≤ 4.2e-16 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 86 | 1.16e−10 |
csf_wmfix, the default: PD-01’s correctly rounded reading (≤ 1e-16 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 6 | 5.68e−14 |
global_signalfix, the default: PD-01’s correctly rounded reading (≤ 1.1e-16 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 13 | 5.68e−14 |
global_signal_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 8.7e-16 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 44 | 1.14e−13 |
global_signal_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 5.3e-16 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 45 | 9.09e−12 |
global_signal_power2fix, the default: PD-01’s correctly rounded reading (≤ 3.1e-16 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 28 | 8.73e−11 |
tcompcorfix, the default: PD-01’s correctly rounded reading (≤ 9.4e-17 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 11 | 5.68e−14 |
white_matterfix, the default: PD-01’s correctly rounded reading (≤ 1e-16 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 4 | 5.68e−14 |
white_matter_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 8.6e-16 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 34 | 6.04e−14 |
white_matter_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.5e-15 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 49 | 7.28e−12 |
white_matter_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.7e-16 of the column’s largest value). | 1 | 164 | 0 | – | 0 | – | 14 | 8.73e−11 |
Cells compared as text (the 10 decimals pandas and numpy write), the columns by name. ≠ fMRIPrep’s: the replica in reproduce with fMRIPrep 25.2.6’s releases against the run’s own file; the step: reproduce with the newest releases against fMRIPrep 25.2.6’s (the originals’ own change, no verdict); the default: the replica’s corrections against reproduce, both with the newest releases.
Every comparison, computed in float64 over every voxel of the outputs
| Comparison (A − B), verdict and reason | Output | max |Δ| | Values that differ | RMS of Δ | max |Δ| ÷ range | 1 − NCC |
|---|---|---|---|---|---|---|
| The replica against fMRIPrep's output | ||||||
| replica, reproduce − fMRIPrep's outputBit-identicalEvery mask fMRIPrep’s run wrote, voxel for voxel and header for header (the masks’ values above 0.99 kept, as fMRIPrep’s Binarize keeps them). | every output | identical: 0 of 131,072 values differ | ||||
replica, fix − fMRIPrep's outputBit-identicalIn fix, its default (the newest releases, the corrections), the replica gives the same masks: no correction reaches them on this run. | every output | identical: 0 of 131,072 values differ | ||||
| The replica against itself: thread counts, fix and reproduce | ||||||
replica: fix − reproduceBit-identicalNo correction by default changes a mask here: fix and reproduce give the same bytes. | every output | identical: 0 of 131,072 values differ | ||||
| replica, fix: 1 thread − 8 threadsBit-identicalThe replica’s result does not depend on its thread count: the same bytes at 1 and 8 threads (every file of the run, the confounds table included). | every output | identical: 0 of 131,072 values differ | ||||
Δ is A − B over every voxel of the masks’ files (float64 masks: the probabilities above 0.99 kept, as fMRIPrep’s Binarize writes them; uint8 for tCompCor’s and the crown). The confounds table is compared cell by cell as text in the table below. The records’ threshold: in reproduce, every array, file, table cell, header, warning, message and exit status bit-identical to the originals’ (or fMRIPrep’s own files).
Provenance of this example
- Data: OpenNeuro ds000256 (Greene et al., 2018), sub-CTS203, PDDL-1.0
- Inputs: fMRIPrep’s confounds workflow ds000256-sub-CTS203-noreconall/bold_task_restfixationnofeedback_wf (sha256 7015c05cc61f2c96…)
- Runs: figures cache entry
confounds/workflows, computed 2026-11-13T08:32:56+00:00 with larmorx5fbdd0f07f6f; numpy 2.5.4 (OpenBLAS 0.3.34, LAPACK 3.12.1, OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6, nipype 1.12.0, niworkflows 1.15.4, nitime 0.12.1, nibabel 5.4.2 (oracles/python-sdcflows-2.17), fMRIPrep 25.2.7 from its tag; fMRIPrep 25.2.6 in its image (niworkflows 1.15.2, nipype 1.10, numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3): the runs’ own files; every command, its environment, threads and exit status are in itsRUN_INFO.json - Workflow:
ds000256-sub-CTS203-noreconall/work/fmriprep_25_2_wf/sub_CTS203_wf/bold_task_restfixationnofeedback_wf/bold_confounds_wf
Every number above is in confounds.json, exported by python -m larmorx_validation evidence confounds.
ds004958: multiband 8, 600 volumes
Worked example
fMRIPrep's confounds of a real BOLD run: multiband 8, 600 volumes
What this shows: In reproduce, with fMRIPrep 25.2.6’s releases, the replica writes every CompCor mask of fMRIPrep’s run, voxel for voxel and header for header, and its confounds table holds every column of fMRIPrep’s file, cell for cell, but 8 cells of the motion chain, where glibc’s atan2 rounds an angle otherwise than the replica’s correctly rounded one (with fMRIPrep’s own motion table, none differs); its JSON byte for byte. In fix, its default, with the newest releases, the masks are the same bytes, at 1 and 8 threads alike; the table below says, column by column, what the newest releases and the corrections change.
- Replica against fMRIPrep's outputBit-identical
fMRIPrep 25.2.6’s confounds workflow (bold_task_rest_wf of OpenNeuro ds004958 sub-fsm64ps) run end to end by the replica from the workflow’s own inputs: in reproduce with fMRIPrep 25.2.6’s releases, compared with the run’s own files (its CompCor masks drawn over the BOLD reference, and its confounds table column by column), with the newest releases (the originals’ own step) and in fix, its default (the corrections), at 8 threads and, in fix, at 1.
- Data
- OpenNeuro ds004958, sub-fsm64ps (CC0-1.0); shown inside fMRIPrep’s BOLD mask, dilated by three voxels
- Workflow
- fMRIPrep 25.2.6,
bold_task_rest_wf/bold_confounds_wf(andds_confounds): 104 × 104 × 72 voxels of 2 × 2 × 2 mm, 600 volumes (float32), TR 0.8 s, 3 dummy scan(s) - Confounds file
sub-fsm64ps_task-rest_desc-confounds_timeseries.tsv: 456 columns, 273,600 cells; the replica’s file has fMRIPrep’s name (yes) and another column order (PIPE-43); its JSON byte for byte- Verdict of the file
- divergence: 8 cells of the motion chain (framewise_displacement, rot_y, rot_y_derivative1, rot_y_derivative1_power2, rot_y_power2): glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of
reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. The JSON byte for byte. - Programs
- The replica 0.1.0rc1 (
bold_confounds, experimental) at commit5fbdd0f07f6f, in numpy 2.5.4 (OpenBLAS 0.3.34, LAPACK 3.12.1, OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6, nipype 1.12.0, niworkflows 1.15.4, nitime 0.12.1, nibabel 5.4.2 (oracles/python-sdcflows-2.17), fMRIPrep 25.2.7 from its tag; fMRIPrep’s files from its run (fMRIPrep 25.2.6 in its image (niworkflows 1.15.2, nipype 1.10, numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3): the runs’ own files) - Notices
- e3: GLIBC-07
reproduce3, PD-01reproduce9, PIPE-41reproduce1, PIPE-43 fix 1; sp: GLIBC-07reproduce3, PD-01reproduce9, PIPE-41reproduce1, PIPE-43 fix 1; new: GLIBC-07reproduce3, PD-01reproduce9, PIPE-41reproduce1, PIPE-43 fix 1; fix: GLIBC-07 fix 3, PD-01 fix 9, PIPE-41 fix 1, PIPE-43 fix 1; fix-1: GLIBC-07 fix 3, PD-01 fix 9, PIPE-41 fix 1, PIPE-43 fix 1 - Grid
- 104 × 104 × 72 voxels of 2 × 2 × 2 mm; shown: 75 × 96 × 72 around the brain
Data shown
- The FreeSurfer Maintenance Dataset (OpenNeuro ds004958, version 1.0.0, doi:10.18112/openneuro.ds004958.v1.0.0), CC0 1.0 public-domain dedication.




Bit-identical778,752 of 778,752 voxels equal
Bit-identical778,752 of 778,752 voxels equal
Axial slice 34 of 72; the masks drawn over the BOLD reference. The interactive viewer needs JavaScript; the metrics below are complete without it.
- fMRIPrep's output (fMRIPrep 25.2.6's run)
- replica, reproduce
- replica, fix (the default)
Each panel one column of the confounds table (n/a cells, the first volume of FD and DVARS, left out). The file holds 456 columns; 5 motion-outlier columns (one per volume over the FD or standardised-DVARS threshold). The lines of fMRIPrep’s file and of the replica in reproduce coincide; the table below counts every cell.
This file’s every column (numbered columns by family): the replica against fMRIPrep’s file, the originals’ step, the default
| Column | Columns | Cells | ≠ fMRIPrep's | largest |Δ| | The step: changed | largest |Δ| | fix: changed | largest |Δ| |
|---|---|---|---|---|---|---|---|---|
rot_xfix, the default: PD-01’s correctly rounded reading (≤ 9.7e-15 of the column’s largest value). | 1 | 600 | 0 | – | 600 | 2.91e−5 | 593 | 9.97e−17 |
rot_x_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 8.7e-14 of the column’s largest value). | 1 | 600 | 0 | – | 599 | 8.79e−6 | 598 | 2.24e−16 |
rot_x_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 5.9e-14 of the column’s largest value). | 1 | 600 | 0 | – | 599 | 2.58e−8 | 598 | 3.85e−19 |
rot_x_power2fix, the default: PD-01’s correctly rounded reading (≤ 7.2e-13 of the column’s largest value). | 1 | 600 | 0 | – | 600 | 2.88e−7 | 595 | 7.62e−17 |
rot_yreproduce: rot_y: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 1.9e-14 of the column’s largest value). | 1 | 600 | 1 | 2e−16 | 600 | 4.09e−5 | 590 | 1e−16 |
rot_y_derivative1reproduce: rot_y_derivative1: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 1.8e-13 of the column’s largest value). | 1 | 600 | 2 | 2e−16 | 599 | 1.77e−5 | 599 | 2.57e−16 |
rot_y_derivative1_power2reproduce: rot_y_derivative1_power2: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 1.3e-13 of the column’s largest value). | 1 | 600 | 2 | 5.42e−20 | 599 | 3.22e−8 | 598 | 2.78e−19 |
rot_y_power2reproduce: rot_y_power2: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 3.5e-14 of the column’s largest value). | 1 | 600 | 1 | 6.68e−19 | 600 | 2.88e−7 | 596 | 9.79e−19 |
rot_zfix, the default: PD-01’s correctly rounded reading (≤ 1.6e-14 of the column’s largest value). | 1 | 600 | 0 | – | 600 | 2.55e−5 | 591 | 9.95e−17 |
rot_z_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 1.1e-13 of the column’s largest value). | 1 | 600 | 0 | – | 599 | 7.8e−6 | 598 | 1.94e−16 |
rot_z_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 4.7e-14 of the column’s largest value). | 1 | 600 | 0 | – | 599 | 2.81e−8 | 598 | 1.55e−19 |
rot_z_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.5e-14 of the column’s largest value). | 1 | 600 | 0 | – | 600 | 2.88e−7 | 593 | 9.55e−19 |
trans_xfix, the default: PD-01’s correctly rounded reading (≤ 2.4e-16 of the column’s largest value). | 1 | 600 | 0 | – | 600 | 0.0064 | 462 | 9.71e−17 |
trans_x_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 3.6e-15 of the column’s largest value). | 1 | 600 | 0 | – | 599 | 0.00185 | 590 | 2.19e−16 |
trans_x_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.8e-14 of the column’s largest value). | 1 | 600 | 0 | – | 599 | 0.000171 | 527 | 1.04e−16 |
trans_x_power2fix, the default: PD-01’s correctly rounded reading (≤ 7.5e-16 of the column’s largest value). | 1 | 600 | 0 | – | 600 | 0.00466 | 590 | 1.25e−16 |
trans_yfix, the default: PD-01’s correctly rounded reading (≤ 1.7e-16 of the column’s largest value). | 1 | 600 | 0 | – | 600 | 0.0124 | 392 | 9.71e−17 |
trans_y_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 5.3e-16 of the column’s largest value). | 1 | 600 | 0 | – | 599 | 0.00281 | 521 | 1.94e−16 |
trans_y_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.1e-15 of the column’s largest value). | 1 | 600 | 0 | – | 599 | 0.000721 | 580 | 1.53e−16 |
trans_y_power2fix, the default: PD-01’s correctly rounded reading (≤ 4.3e-16 of the column’s largest value). | 1 | 600 | 0 | – | 600 | 0.0119 | 570 | 1.39e−16 |
trans_zfix, the default: PD-01’s correctly rounded reading (≤ 1.9e-16 of the column’s largest value). | 1 | 600 | 0 | – | 600 | 0.0115 | 175 | 2.22e−16 |
trans_z_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 7.1e-16 of the column’s largest value). | 1 | 600 | 0 | – | 599 | 0.00223 | 575 | 2.71e−16 |
trans_z_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 8.3e-16 of the column’s largest value). | 1 | 600 | 0 | – | 599 | 0.000625 | 546 | 1.21e−16 |
trans_z_power2fix, the default: PD-01’s correctly rounded reading (≤ 3.4e-16 of the column’s largest value). | 1 | 600 | 0 | – | 600 | 0.0191 | 313 | 4.44e−16 |
framewise_displacementreproduce: framewise_displacement: glibc’s atan2 (SciPy’s Rotation.as_euler calls the C library’s) rounds these angles otherwise than the replica’s correctly rounded atan2, in both modes (decision D4: glibc’s code is LGPL, never translated; a gap of reproduce, GLIBC-07’s row in the register): with fMRIPrep’s own motion table in place of the replica’s, every column of the file is fMRIPrep’s. fix, the default: PD-01’s correctly rounded reading (≤ 1.9e-14 of the column’s largest value). | 1 | 600 | 2 | 1.01e−14 | 599 | 0.0062 | 596 | 1.54e−14 |
motion_outlierXX | 5 | 3,000 | 0 | – | 0 | – | 0 | – |
rmsdfix, the default: PD-01’s correctly rounded reading (≤ 2e-16 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 494 | 9.71e−17 |
dvarsfix, the default: PD-01’s correctly rounded reading (≤ 1.7e-16 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 1 | 1.42e−14 |
std_dvarsfix, the default: PD-01’s correctly rounded reading (≤ 1.5e-16 of the column’s largest value). | 1 | 600 | 0 | – | 2 | 1e−6 | 9 | 2.22e−16 |
a_comp_cor_XXfix, the default: 0 columns negated, 6 by one unit in the 10th decimal (GLIBC-07, expected). | 166 | 99,600 | 0 | – | 2,396 | 0.327 | 6 | 1e−10 |
c_comp_cor_XX | 37 | 22,200 | 0 | – | 1 | 1e−10 | 0 | – |
edge_comp_XX | 24 | 14,400 | 0 | – | 597 | 0.239 | 0 | – |
t_comp_cor_XX | 7 | 4,200 | 0 | – | 0 | – | 0 | – |
w_comp_cor_XXfix, the default: 0 columns negated, 5 by one unit in the 10th decimal (GLIBC-07, expected). | 166 | 99,600 | 0 | – | 1,205 | 0.285 | 5 | 1e−10 |
cosineXX | 6 | 3,600 | 0 | – | 0 | – | 0 | – |
non_steady_state_outlierXX | 3 | 1,800 | 0 | – | 0 | – | 0 | – |
csffix, the default: PD-01’s correctly rounded reading (≤ 2.3e-16 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 102 | 9.09e−13 |
csf_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 6.6e-15 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 234 | 9.11e−13 |
csf_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 7.3e-15 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 252 | 1.4e−10 |
csf_power2fix, the default: PD-01’s correctly rounded reading (≤ 5.9e-16 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 211 | 9.31e−9 |
csf_wmfix, the default: PD-01’s correctly rounded reading (≤ 3.6e-16 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 74 | 9.09e−13 |
global_signalfix, the default: PD-01’s correctly rounded reading (≤ 3e-16 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 71 | 9.09e−13 |
global_signal_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 4.4e-14 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 241 | 1.36e−12 |
global_signal_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.1e-14 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 255 | 1.13e−11 |
global_signal_power2fix, the default: PD-01’s correctly rounded reading (≤ 6.2e-16 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 71 | 5.59e−9 |
tcompcorfix, the default: PD-01’s correctly rounded reading (≤ 2.7e-16 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 95 | 9.09e−13 |
white_matterfix, the default: PD-01’s correctly rounded reading (≤ 3.6e-16 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 99 | 9.09e−13 |
white_matter_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 7.8e-14 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 273 | 1.36e−12 |
white_matter_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 5.9e-14 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 286 | 1.81e−11 |
white_matter_power2fix, the default: PD-01’s correctly rounded reading (≤ 7.2e-16 of the column’s largest value). | 1 | 600 | 0 | – | 0 | – | 109 | 4.66e−9 |
Cells compared as text (the 10 decimals pandas and numpy write), the columns by name. ≠ fMRIPrep’s: the replica in reproduce with fMRIPrep 25.2.6’s releases against the run’s own file; the step: reproduce with the newest releases against fMRIPrep 25.2.6’s (the originals’ own change, no verdict); the default: the replica’s corrections against reproduce, both with the newest releases.
Every comparison, computed in float64 over every voxel of the outputs
| Comparison (A − B), verdict and reason | Output | max |Δ| | Values that differ | RMS of Δ | max |Δ| ÷ range | 1 − NCC |
|---|---|---|---|---|---|---|
| The replica against fMRIPrep's output | ||||||
| replica, reproduce − fMRIPrep's outputBit-identicalEvery mask fMRIPrep’s run wrote, voxel for voxel and header for header (the masks’ values above 0.99 kept, as fMRIPrep’s Binarize keeps them). | every output | identical: 0 of 778,752 values differ | ||||
replica, fix − fMRIPrep's outputBit-identicalIn fix, its default (the newest releases, the corrections), the replica gives the same masks: no correction reaches them on this run. | every output | identical: 0 of 778,752 values differ | ||||
| The replica against itself: thread counts, fix and reproduce | ||||||
replica: fix − reproduceBit-identicalNo correction by default changes a mask here: fix and reproduce give the same bytes. | every output | identical: 0 of 778,752 values differ | ||||
| replica, fix: 1 thread − 8 threadsBit-identicalThe replica’s result does not depend on its thread count: the same bytes at 1 and 8 threads (every file of the run, the confounds table included). | every output | identical: 0 of 778,752 values differ | ||||
Δ is A − B over every voxel of the masks’ files (float64 masks: the probabilities above 0.99 kept, as fMRIPrep’s Binarize writes them; uint8 for tCompCor’s and the crown). The confounds table is compared cell by cell as text in the table below. The records’ threshold: in reproduce, every array, file, table cell, header, warning, message and exit status bit-identical to the originals’ (or fMRIPrep’s own files).
Provenance of this example
- Data: OpenNeuro ds004958, sub-fsm64ps, CC0-1.0
- Inputs: fMRIPrep’s confounds workflow ds004958-sub-fsm64ps-noreconall/bold_task_rest_wf (sha256 0bbb1326f28919c9…)
- Runs: figures cache entry
confounds/workflows, computed 2026-11-13T08:32:56+00:00 with larmorx5fbdd0f07f6f; numpy 2.5.4 (OpenBLAS 0.3.34, LAPACK 3.12.1, OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6, nipype 1.12.0, niworkflows 1.15.4, nitime 0.12.1, nibabel 5.4.2 (oracles/python-sdcflows-2.17), fMRIPrep 25.2.7 from its tag; fMRIPrep 25.2.6 in its image (niworkflows 1.15.2, nipype 1.10, numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3): the runs’ own files; every command, its environment, threads and exit status are in itsRUN_INFO.json - Workflow:
ds004958-sub-fsm64ps-noreconall/work/fmriprep_25_2_wf/sub_fsm64ps_wf/bold_task_rest_wf/bold_confounds_wf
Every number above is in confounds.json, exported by python -m larmorx_validation evidence confounds.
ds003592: an older adult’s multi-echo run
Worked example
fMRIPrep's confounds of a real BOLD run: an older adult's multi-echo run (the optimally combined series)
What this shows: In reproduce, with fMRIPrep 25.2.6’s releases, the replica writes every CompCor mask of fMRIPrep’s run, voxel for voxel and header for header, and its confounds table holds every column of fMRIPrep’s file, cell for cell, and its JSON byte for byte. In fix, its default, with the newest releases, the masks are the same bytes; the table below says, column by column, what the newest releases and the corrections change.
- Replica against fMRIPrep's outputBit-identical
fMRIPrep 25.2.6’s confounds workflow (bold_ses_1_task_rest_echo_1_wf of OpenNeuro ds003592 (Spreng et al.) sub-280) run end to end by the replica from the workflow’s own inputs: in reproduce with fMRIPrep 25.2.6’s releases, compared with the run’s own files (its CompCor masks drawn over the BOLD reference, and its confounds table column by column), with the newest releases (the originals’ own step) and in fix, its default (the corrections), at 8 threads and, in fix, at 1.
- Data
- OpenNeuro ds003592 (Spreng et al.), sub-280 (CC0-1.0); shown inside fMRIPrep’s BOLD mask, dilated by three voxels
- Workflow
- fMRIPrep 25.2.6,
bold_ses_1_task_rest_echo_1_wf/bold_confounds_wf(andds_confounds): 64 × 64 × 43 voxels of 3.38 × 3.38 × 3 mm, 200 volumes (float32), TR 3.0 s, 0 dummy scan(s) - Confounds file
sub-280_ses-1_task-rest_desc-confounds_timeseries.tsv: 132 columns, 26,400 cells; the replica’s file has fMRIPrep’s name (yes) and another column order (PIPE-43); its JSON byte for byte- Verdict of the file
- bit-identical: Every column of fMRIPrep’s file, cell for cell as text (matched by name: the order of niworkflows’ set, PIPE-43, is never reproduced), and its JSON byte for byte.
- Programs
- The replica 0.1.0rc1 (
bold_confounds, experimental) at commit5fbdd0f07f6f, in numpy 2.5.4 (OpenBLAS 0.3.34, LAPACK 3.12.1, OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6, nipype 1.12.0, niworkflows 1.15.4, nitime 0.12.1, nibabel 5.4.2 (oracles/python-sdcflows-2.17), fMRIPrep 25.2.7 from its tag; fMRIPrep’s files from its run (fMRIPrep 25.2.6 in its image (niworkflows 1.15.2, nipype 1.10, numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3): the runs’ own files) - Notices
- e3: GLIBC-07
reproduce3, PD-01reproduce9, PIPE-41reproduce1, PIPE-43 fix 1; new: GLIBC-07reproduce3, PD-01reproduce9, PIPE-41reproduce1, PIPE-43 fix 1; fix: GLIBC-07 fix 3, PD-01 fix 9, PIPE-41 fix 1, PIPE-43 fix 1 - Grid
- 64 × 64 × 43 voxels of 3.38 × 3.38 × 3 mm; shown: 53 × 59 × 43 around the brain
Data shown
- Neurocognitive aging data release (OpenNeuro ds003592, version 1.0.13, doi:10.18112/openneuro.ds003592.v1.0.13), CC0 1.0 public-domain dedication.




Bit-identical176,128 of 176,128 voxels equal
Bit-identical176,128 of 176,128 voxels equal
Axial slice 21 of 43; the masks drawn over the BOLD reference. The interactive viewer needs JavaScript; the metrics below are complete without it.
- fMRIPrep's output (fMRIPrep 25.2.6's run)
- replica, reproduce
- replica, fix (the default)
Each panel one column of the confounds table (n/a cells, the first volume of FD and DVARS, left out). The file holds 132 columns; 40 motion-outlier columns (one per volume over the FD or standardised-DVARS threshold). The lines of fMRIPrep’s file and of the replica in reproduce coincide; the table below counts every cell.
This file’s every column (numbered columns by family): the replica against fMRIPrep’s file, the originals’ step, the default
| Column | Columns | Cells | ≠ fMRIPrep's | largest |Δ| | The step: changed | largest |Δ| | fix: changed | largest |Δ| |
|---|---|---|---|---|---|---|---|---|
rot_xfix, the default: PD-01’s correctly rounded reading (≤ 1.3e-14 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 6.29e−6 | 198 | 9.95e−17 |
rot_x_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 2.7e-14 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 3.91e−6 | 199 | 1.94e−16 |
rot_x_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.1e-14 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 3.54e−8 | 199 | 1.07e−18 |
rot_x_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.3e-14 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 8.93e−8 | 198 | 1.42e−18 |
rot_yfix, the default: PD-01’s correctly rounded reading (≤ 1.6e-14 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 7.67e−6 | 188 | 9.75e−17 |
rot_y_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 8.2e-14 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 4.89e−6 | 196 | 1.79e−16 |
rot_y_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 4.1e-14 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 1.48e−8 | 194 | 1.98e−19 |
rot_y_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.9e-14 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 8.93e−8 | 192 | 1.1e−18 |
rot_zfix, the default: PD-01’s correctly rounded reading (≤ 6.8e-14 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 4.77e−5 | 186 | 9.97e−17 |
rot_z_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 2.3e-13 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 2.55e−5 | 194 | 1.82e−16 |
rot_z_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 2.1e-13 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 2.46e−8 | 193 | 1.35e−19 |
rot_z_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.1e-13 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 8.93e−8 | 188 | 2.38e−19 |
trans_xfix, the default: PD-01’s correctly rounded reading (≤ 3.5e-16 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 0.00845 | 163 | 9.8e−17 |
trans_x_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 2.7e-15 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 0.0039 | 193 | 2.78e−16 |
trans_x_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 9.9e-15 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 0.000234 | 190 | 1.02e−16 |
trans_x_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.6e-15 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 0.00434 | 198 | 1.25e−16 |
trans_yfix, the default: PD-01’s correctly rounded reading (≤ 2.3e-16 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 0.0118 | 139 | 9.71e−17 |
trans_y_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 6.1e-16 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 0.00714 | 180 | 2.08e−16 |
trans_y_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 9.4e-16 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 0.00292 | 193 | 1.11e−16 |
trans_y_power2fix, the default: PD-01’s correctly rounded reading (≤ 7.2e-16 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 0.00753 | 192 | 1.32e−16 |
trans_zfix, the default: PD-01’s correctly rounded reading (≤ 5.9e-17 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 0.00631 | 144 | 9.71e−17 |
trans_z_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 1.6e-16 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 0.00313 | 170 | 2.5e−16 |
trans_z_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 6.8e-17 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 0.00473 | 188 | 1.67e−16 |
trans_z_power2fix, the default: PD-01’s correctly rounded reading (≤ 4.6e-17 of the column’s largest value). | 1 | 200 | 0 | – | 200 | 0.00545 | 190 | 1.25e−16 |
framewise_displacementfix, the default: PD-01’s correctly rounded reading (≤ 8e-15 of the column’s largest value). | 1 | 200 | 0 | – | 199 | 0.00926 | 198 | 1.59e−14 |
motion_outlierXX | 40 | 8,000 | 0 | – | 0 | – | 0 | – |
rmsdfix, the default: PD-01’s correctly rounded reading (≤ 6e-17 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 134 | 9.71e−17 |
dvarsfix, the default: PD-01’s correctly rounded reading (≤ 1.6e-16 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 2 | 1.42e−14 |
std_dvarsfix, the default: PD-01’s correctly rounded reading (≤ 4e-17 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 8 | 1.11e−16 |
a_comp_cor_XX | 3 | 600 | 0 | – | 0 | – | 0 | – |
c_comp_cor_XX | 5 | 1,000 | 0 | – | 0 | – | 0 | – |
edge_comp_XXfix, the default: 10 columns negated, 0 by one unit in the 10th decimal (GLIBC-07, expected). | 24 | 4,800 | 0 | – | 2,000 | 0.789 | 2,000 | 0.789 |
t_comp_cor_XX | 3 | 600 | 0 | – | 0 | – | 0 | – |
w_comp_cor_XX | 7 | 1,400 | 0 | – | 0 | – | 0 | – |
cosineXX | 8 | 1,600 | 0 | – | 0 | – | 0 | – |
csf | 1 | 200 | 0 | – | 0 | – | 0 | – |
csf_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 8.9e-17 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 44 | 3.55e−15 |
csf_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 7.1e-17 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 37 | 1.14e−13 |
csf_power2 | 1 | 200 | 0 | – | 0 | – | 0 | – |
csf_wm | 1 | 200 | 0 | – | 0 | – | 0 | – |
global_signal | 1 | 200 | 0 | – | 0 | – | 0 | – |
global_signal_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 1e-16 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 42 | 1.78e−15 |
global_signal_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.2e-17 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 51 | 3.55e−15 |
global_signal_power2fix, the default: PD-01’s correctly rounded reading (≤ 1.6e-16 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 14 | 5.82e−11 |
tcompcor | 1 | 200 | 0 | – | 0 | – | 0 | – |
white_matterfix, the default: PD-01’s correctly rounded reading (≤ 1.4e-16 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 53 | 5.68e−14 |
white_matter_derivative1fix, the default: PD-01’s correctly rounded reading (≤ 1.3e-14 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 101 | 1.14e−13 |
white_matter_derivative1_power2fix, the default: PD-01’s correctly rounded reading (≤ 1e-14 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 111 | 7.82e−13 |
white_matter_power2fix, the default: PD-01’s correctly rounded reading (≤ 3.7e-16 of the column’s largest value). | 1 | 200 | 0 | – | 0 | – | 55 | 5.82e−11 |
Cells compared as text (the 10 decimals pandas and numpy write), the columns by name. ≠ fMRIPrep’s: the replica in reproduce with fMRIPrep 25.2.6’s releases against the run’s own file; the step: reproduce with the newest releases against fMRIPrep 25.2.6’s (the originals’ own change, no verdict); the default: the replica’s corrections against reproduce, both with the newest releases.
Every comparison, computed in float64 over every voxel of the outputs
| Comparison (A − B), verdict and reason | Output | max |Δ| | Values that differ | RMS of Δ | max |Δ| ÷ range | 1 − NCC |
|---|---|---|---|---|---|---|
| The replica against fMRIPrep's output | ||||||
| replica, reproduce − fMRIPrep's outputBit-identicalEvery mask fMRIPrep’s run wrote, voxel for voxel and header for header (the masks’ values above 0.99 kept, as fMRIPrep’s Binarize keeps them). | every output | identical: 0 of 176,128 values differ | ||||
replica, fix − fMRIPrep's outputBit-identicalIn fix, its default (the newest releases, the corrections), the replica gives the same masks: no correction reaches them on this run. | every output | identical: 0 of 176,128 values differ | ||||
| The replica against itself: fix and reproduce | ||||||
replica: fix − reproduceBit-identicalNo correction by default changes a mask here: fix and reproduce give the same bytes. | every output | identical: 0 of 176,128 values differ | ||||
Δ is A − B over every voxel of the masks’ files (float64 masks: the probabilities above 0.99 kept, as fMRIPrep’s Binarize writes them; uint8 for tCompCor’s and the crown). The confounds table is compared cell by cell as text in the table below. The records’ threshold: in reproduce, every array, file, table cell, header, warning, message and exit status bit-identical to the originals’ (or fMRIPrep’s own files).
Provenance of this example
- Data: OpenNeuro ds003592 (Spreng et al.), sub-280, CC0-1.0
- Inputs: fMRIPrep’s confounds workflow ds003592-sub-280-noreconall/bold_ses_1_task_rest_echo_1_wf (sha256 0f012050203d22e8…)
- Runs: figures cache entry
confounds/workflows, computed 2026-11-13T08:32:56+00:00 with larmorx5fbdd0f07f6f; numpy 2.5.4 (OpenBLAS 0.3.34, LAPACK 3.12.1, OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6, nipype 1.12.0, niworkflows 1.15.4, nitime 0.12.1, nibabel 5.4.2 (oracles/python-sdcflows-2.17), fMRIPrep 25.2.7 from its tag; fMRIPrep 25.2.6 in its image (niworkflows 1.15.2, nipype 1.10, numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3): the runs’ own files; every command, its environment, threads and exit status are in itsRUN_INFO.json - Workflow:
ds003592-sub-280-noreconall/work/fmriprep_25_2_wf/sub_280_ses_1_wf/bold_ses_1_task_rest_echo_1_wf/bold_confounds_wf
Every number above is in confounds.json, exported by python -m larmorx_validation evidence confounds.
The masks’ Gaussian resampling
fMRIPrep resamples the anatomical tissue maps onto the BOLD grid with ANTs’ antsApplyTransforms -n Gaussian before it thresholds them into the aCompCor masks. The Gaussian kernel’s error function goes through the C library’s exp and log in fMRIPrep’s ANTs, and through correctly rounded ones in LARMORX’s replica of ANTs. So the resampled probabilities differ in their last bits, and the example traces the cause: ANTs 2.6.5 run with correctly rounded maths gives the replica’s bytes, and run as built gives fMRIPrep’s. No voxel crosses the masks’ thresholds (0.5 and 0.99), so the masks aCompCor reads are fMRIPrep’s.
Worked example
The aCompCor masks' Gaussian resampling, and why its last bits differ
What this shows: The replica gives the bytes of ANTs 2.6.5’s antsApplyTransforms when its exp and log are correctly rounded. Against fMRIPrep’s own file 99 voxels differ in their last bits: fMRIPrep’s ANTs computes the Gaussian kernel’s error function with glibc’s exp and log, and ANTs as built here gives fMRIPrep’s bytes. No voxel crosses the masks’ threshold, so the masks the confounds use are fMRIPrep’s.
- Replica against ANTs antsApplyTransforms 2.6.5Bit-identical
fMRIPrep resamples the T1w’s CSF, WM and CSF + WM probability maps onto the BOLD reference with antsApplyTransforms’ Gaussian interpolation (its acc_msk_tfm node), then keeps the brain (0.5) and binarizes (0.99). Its three items on OpenNeuro ds000005 (Tom et al., 2007), snapshot 00001 sub-01: fMRIPrep’s files, ANTs 2.6.5 as built and with correctly rounded maths, fMRIPrep’s own ANTs run again in its image, and the replica’s ANTs port as the workflow runs it.
- Data
- OpenNeuro ds000005 (Tom et al., 2007), snapshot 00001, sub-01 (PDDL-1.0); shown inside fMRIPrep’s BOLD mask, dilated by three voxels
- Node
- fMRIPrep 25.2.6,
bold_task_mixedgamblestask_run_01_wf/bold_confounds_wf/acc_msk_tfm(three items) - Command
antsApplyTransforms --default-value 0 --float 0 --input /out/sub-01/anat/sub-01_label-CSF_probseg.nii.gz --interpolation Gaussian --output sub-01_label-CSF_probseg_trans.nii.gz --reference-image /out/sub-01/func/sub-01_task-mixedgamblestask_run-01_desc-brain_mask.nii.gz --transform [ /out/sub-01/func/sub-01_task-mixedgamblestask_run-01_from-boldref_to-T1w_mode-image_desc-coreg_xfm.txt, 1 ](the first item; /out is the run’s output folder)- Programs
- ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0, one thread, as built and with CORE-MATH preloaded; fMRIPrep 25.2.6’s image’s antsApplyTransforms; the replica at commit
5fbdd0f07f6f - Grid
- 64 × 64 × 34 voxels of 3.12 × 3.12 × 4 mm; shown: 57 × 59 × 34 around the brain
Data shown
- Mixed-gambles task (OpenNeuro ds000005, version 00001), Open Data Commons Public Domain Dedication and Licence 1.0.





Bit-identical139,264 of 139,264 voxels equal

Axial slice 26 of 34; the crosshair where the replica and fMRIPrep's file differ most. The interactive viewer needs JavaScript; the metrics below are complete without it.
Every comparison, computed in float64 over every voxel of the outputs
| Comparison (A − B), verdict and reason | Output | max |Δ| | Values that differ | RMS of Δ | max |Δ| ÷ range | 1 − NCC |
|---|---|---|---|---|---|---|
| The original against itself: the yardstick | ||||||
| ANTs antsApplyTransforms 2.6.5, correctly rounded maths − fMRIPrep's outputreferenceThe same ANTs program with correctly rounded exp and log: where vnl’s erf, under the Gaussian kernel, reaches glibc’s last bits. | CSF probability in BOLD space | 2.22e−16 | 76 of 139,264 (0.0546 %) | 1.83e−18 | 2.22e−16 | 5.69e−35 |
| WM probability in BOLD space | 2.22e−16 | 10 of 139,264 (0.00718 %) | 8.53e−19 | 2.22e−16 | 4.95e−35 | |
| CSF + WM probability in BOLD space | 2.22e−16 | 13 of 139,264 (0.00933 %) | 9.18e−19 | 2.22e−16 | 4.81e−35 | |
| ANTs antsApplyTransforms 2.6.5 as built (glibc's maths) − fMRIPrep's outputreferenceANTs 2.6.5 as the repository builds it, with glibc’s maths, gives fMRIPrep’s bytes: fMRIPrep’s ANTs computes the same on this call. | every output | identical: 0 of 139,264 values differ | ||||
| fMRIPrep's image's ANTs, run again − fMRIPrep's outputreferencefMRIPrep’s own antsApplyTransforms, run again in its image, gives its file’s bytes: the node is deterministic. | every output | identical: 0 of 139,264 values differ | ||||
| The replica against the original | ||||||
| replica − ANTs antsApplyTransforms 2.6.5, 1 thread, correctly rounded mathsBit-identicalANTs 2.6.5’s bytes when its exp and log are correctly rounded. | every output | identical: 0 of 139,264 values differ | ||||
| The replica against fMRIPrep's output | ||||||
replica, reproduce − fMRIPrep's outputIdentical with correctly rounded maths99 voxels differ in their last bits, where vnl’s erf goes through glibc’s exp and log in fMRIPrep’s ANTs and through correctly rounded ones in LARMORX’s replica of ANTs (LARMORX’s rule of correctly rounded maths, in fix and reproduce): ANTs 2.6.5 with CORE-MATH preloaded gives the replica’s bytes, as built fMRIPrep’s. No voxel crosses ApplyMask’s 0.5 or Binarize’s 0.99: aCompCor reads the same masks. | CSF probability in BOLD space | 2.22e−16 | 76 of 139,264 (0.0546 %) | 1.83e−18 | 2.22e−16 | 5.69e−35 |
| WM probability in BOLD space | 2.22e−16 | 10 of 139,264 (0.00718 %) | 8.53e−19 | 2.22e−16 | 4.95e−35 | |
| CSF + WM probability in BOLD space | 2.22e−16 | 13 of 139,264 (0.00933 %) | 9.18e−19 | 2.22e−16 | 4.81e−35 | |
| The replica against itself: fix and reproduce | ||||||
replica: fix − reproduceBit-identicalfix and reproduce compute the resampling with correctly rounded maths. | every output | identical: 0 of 139,264 values differ | ||||
Δ is A − B over every voxel of the files (float64 probabilities). The records’ threshold: in reproduce, every array, file, table cell, header, warning, message and exit status bit-identical to the originals’ (or fMRIPrep’s own files).
The three items: the replica (reproduce) against fMRIPrep’s files
| Map | Voxels above 0 | Voxels that differ | largest |Δ| | Crossing 0.99 | Crossing 0.5 | = ANTs with CORE-MATH |
|---|---|---|---|---|---|---|
| CSF probability in BOLD space | 23,600 | 76 | 2.22e−16 | 0 | 0 | yes |
| WM probability in BOLD space | 3,427 | 10 | 2.22e−16 | 0 | 0 | yes |
| CSF + WM probability in BOLD space | 3,843 | 13 | 2.22e−16 | 0 | 0 | yes |
The masks aCompCor reads keep the values above 0.99 inside the brain (0.5): no voxel crosses either threshold, so the components are computed from the same voxels. The other shown runs’ items give the same picture (the page’s table of the Gaussian resampling).
Provenance of this example
- Data: OpenNeuro ds000005 (Tom et al., 2007), snapshot 00001, sub-01, PDDL-1.0
- Inputs: fMRIPrep’s acc_msk_tfm item 0 of ds000005-sub-01-noreconall (sha256 3319f623fc1bd816…); fMRIPrep’s acc_msk_tfm item 1 of ds000005-sub-01-noreconall (sha256 67184e02fd8d209b…); fMRIPrep’s acc_msk_tfm item 2 of ds000005-sub-01-noreconall (sha256 9dd023ced221a77b…)
- Runs: figures cache entry
confounds/gaussian, computed 2026-11-13T08:10:44+00:00 with larmorxd1d1228d38e0; ANTs v2.6.5 (fdce4d2f84) on ITK v5.4.5 (f51594ad88), x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.3.0-20) 14.3.0; fMRIPrep 25.2.6’s image (its ANTs); every command, its environment, threads and exit status are in itsRUN_INFO.json - LARMORX’s runs:
confounds/workflows
Every number above is in confounds.json, exported by python -m larmorx_validation evidence confounds.
The aCompCor masks’ Gaussian resampling on the three shown runs: the chain of causes
| Item | Voxels ≠ fMRIPrep's | ANTs 2.6.5 as built = fMRIPrep's | fMRIPrep's ANTs again = its file | replica = ANTs with correctly rounded maths | Threshold crossings |
|---|---|---|---|---|---|
| ds000005: CSF probability in BOLD space | 76 | yes | yes | yes | 0 |
| ds000005: WM probability in BOLD space | 10 | yes | yes | yes | 0 |
| ds000005: CSF + WM probability in BOLD space | 13 | yes | yes | yes | 0 |
| ds000256: CSF probability in BOLD space | 44 | yes | yes | yes | 0 |
| ds000256: WM probability in BOLD space | 8 | yes | yes | yes | 0 |
| ds000256: CSF + WM probability in BOLD space | 8 | yes | yes | yes | 0 |
| ds004958: CSF probability in BOLD space | 327 | yes | yes | yes | 0 |
| ds004958: WM probability in BOLD space | 51 | yes | yes | yes | 0 |
| ds004958: CSF + WM probability in BOLD space | 66 | yes | yes | yes | 0 |
The cache entry confounds/gaussian: each acc_msk_tfm item’s command (its command.txt and inputs) run by ANTs 2.6.5 (the repository’s GCC 14.3 build) at one thread, as built and with CORE-MATH’s correctly rounded exp and log preloaded, and by fMRIPrep 25.2.6’s own antsApplyTransforms in its image; the replica’s file from the workflow’s reproduce run. Identical with correctly rounded maths: the replica gives ANTs’ bytes when ANTs’ maths are correctly rounded, and ANTs as built gives fMRIPrep’s.
Corrections by default
Every defect the replica found in the originals was decided by the register’s rules (A: corrected by default; B: refused with a clear error; C: kept and documented; D: measured first). upstream_defects="reproduce" (on the command line --upstream-defects reproduce from the next release) gives the original’s behaviour, its errors and warnings included.
- 49 entries were registered: PIPE-41 to PIPE-83 in nipype, niworkflows and fMRIPrep, PD-01 to PD-03 in pandas, NP-09 and NP-10 in numpy, SCI-41 in SciPy.
- 29 are corrected by default (two undefined conversions clamped) and 12 refused with a clear error, 41 with the switch; among them every confounds JSON’s missing tissue-signal entries (PIPE-41) and nipype’s file names with a blank.
- GLIBC-07, the C library’s cosine: in
fixthe cosine regressors take the correctly roundedcos, which turns the sign of some CompCor components (the multi-echo example above).
The originals’ defects found while porting (PIPE-41 to PIPE-83, PD-01 to PD-03, NP-09, NP-10, SCI-41) and GLIBC-07’s cosine
| Entry | Rule | What the original does (the findings' words) | the replica in fix, its default | Switch |
|---|---|---|---|---|
| PIPE-41 | rule A | merge_confound_metadata.inputs.in3 is set to the tissue signals’ metadata ({“Method”: “Mean”} for global_signal, csf, white_matter, csf_wm, tcompcor) and then connected to the crown’s metadata, which replaces it: the confounds’ JSON never describes those five columns (all 36 files of fMRIPrep’s runs). | – | yes |
| PIPE-42 | rule A, with larmorprepx’s reports | merge_compcor_metadata = Merge(2) with the crown wired to in3: the CompCor variance reportlet never shows crownCompCor. | – | – |
| PIPE-43 | class b, never reproduced | The expansion’s column order changes between fMRIPrep processes | – | yes |
| PIPE-44 | rule A | a mask that gives no component (an empty mask with ‘all’) ends the loop over the masks (break): the later masks are skipped, and with no component at all failure_mode=‘NaN’ stops with ValueError: need at least one array to concatenate | the loop goes on with the next mask; no component and empty metadata when none is found | yes |
| PIPE-45 | rule A | only voxels whose float sum over time is NaN are zeroed: a voxel holding an infinity reaches the filter and the SVD, which fails (every component NaN, or ValueError), while a finite voxel whose float32 sum overflows both ways is zeroed | exactly the voxels that hold a non-finite value are zeroed | yes |
| PIPE-46 | rule A, after measuring how it is reached | the voxel-wise standardisation runs under warnings.filterwarnings(“error”) whatever DVARS was asked for, so a voxel whose predicted SD is 0 stops the whole computation (RuntimeWarning: divide by zero encountered in divide, or invalid value for 0/0) and all three series are lost. | – | yes |
| PIPE-47 | rule D, measured, then A | an AR(1) estimate above 1 (the float32 FFT’s rounding; the exact Yule-Walker estimate of the biased autocorrelation is below 1) makes the predicted SD NaN and every standardised and voxel-wise value NaN, silently. | – | yes |
| PIPE-48 | rule C, kept, documented; no switch | niworkflows’ RobustACompCor/RobustTCompCor (fMRIPrep’s three CompCor nodes) retry a failed decomposition (a workaround for a random MKL failure): the loop catches LinAlgError only and sleeps randint(10 f - 6, 10 f) seconds before each of 10 retries. | – | – |
| PIPE-49 | rule C, no switch | the crown’s metadata are not filtered (FilterDropped runs on the tCompCor and aCompCor metadata only), so the JSON lists the crown’s components not retained as dropped_ | – | – |
| PIPE-50 | rule A | a voxel whose demeaned series is all zeros (an all-zero voxel, or a constant one whose regress_poly(0) residual rounds to exactly 0; nipype’s function default remove_zerovariance=False keeps it) has 0 · a = 0: with SciPy 1.15 NaN (every standardised and voxel-wise value NaN, silently), with SciPy 1.18.1 LinAlgError | – | yes |
| PIPE-51 | rule A | the maps of a series stored as integers (tSNR, mean, SD, the detrended series) are written in that integer type, quantized by nibabel’s scaling (256 levels for uint8, 65,536 for int16): nipype’s if data.dtype.kind == “i”: header.set_data_dtype(np.float32) never fires, as get_fdata(dtype=np.float32) made the data float32 by then. The replica writes float32 (the check’s intent: the default’s files equal nipype’s code with… | – | yes |
| PIPE-52 | rule A | regress_poly returns X[:, :-1] as the non-constant regressors (the pre-filter file): the constant and every polynomial but the highest | the polynomials of degree 1 to d | yes |
| PIPE-53 | rule B | an axis other than the last takes that axis’ length as the series’ length but reshapes in memory order: the series regressed are runs of consecutive values across voxels and volumes | ValueError | yes |
| PIPE-54 | rule B (A for the float count) | components_criterion 0 ends the loop at the first mask; a negative one other than -1 slices columns off the end; a count given as a float (6.0) fails with TypeError when an SVD fails | 0 and negative criteria refused (ValueError naming PIPE-54); a float count by its integer part | yes |
| PIPE-55 | rule A | an empty last mask with a failure_mode other than ‘error’ returns an empty basis: the pre-filter file loses the basis | the basis kept | yes |
| PIPE-56 | rule A | a header TR in microseconds (xyzt_units usec) is taken as seconds: millions of cosine columns | microseconds converted | yes |
| PIPE-57 | rule B | a filter_type other than ‘cosine’, ‘polynomial’ and False filters nothing, not even the mean, and a float32 series is then decomposed in single precision | ValueError (reproduce on a float32 series: refused, sgesdd not ported) | yes |
| PIPE-58 | rule B | mask_names shorter than the masks silently drops the masks without a name (zip) | ValueError | yes |
| PIPE-59 | rule A | an SVD that fails with failure_mode=‘NaN’ gives T NaN singular values whatever min(T, N), and with ‘all’ one NaN component while all T values are marked retained | the decomposition’s size min(T, N) | yes |
| PIPE-60 | rule A; nipy rotation_vec2mat, through nipype’s public normalize_mc_params | an integer rotation vector’s squares are computed in its own type and wrap around silently (int8 from a component of 12 | – | yes |
| PIPE-61 | rule A; niworkflows AddTSVHeader | np.loadtxt squeezes a one-row table to a vector (which savetxt writes as one column) and a one-value table to a scalar (savetxt’s ValueError | – | yes |
| PIPE-62 | rule B; AddTSVHeader | a header with another number of names than the table’s columns is written as given: the default raises ValueError, reproduce writes the file (notice). | – | yes |
| PIPE-63 | rule A; fMRIPrep GatherConfounds | a shorter table is aligned with the last rows (_adjust_indices, “missing values at the beginning”), but pd.concat(axis=1) joins the row labels with sort=False (the earlier tables’ labels, then the new ones), so when an earlier table is the shorter one (no signals, a DVARS table of T - 1 rows first, then a table of T rows), the first volume’s row goes to the end of the file and every other row sits one volume early (… | – | yes |
| PIPE-64 | rule A | the shorthands (wm, gsr, rps, fd, acc, tcc, dv, dvall, nss, spikes, others) are replaced as text, anywhere and again in every group | – | yes |
| PIPE-65 | rule B for ignored text, A for the rest | parse_expression reads the operators with regular expressions over the whole term | – | yes |
| PIPE-66 | rule A | a variable reached by two terms is repeated | – | yes |
| PIPE-67 | rule A | the lags compound | – | yes |
| PIPE-68 | rule A | minimum_contiguous’s post_final = T + 1 counts the last unflagged run one volume longer (T = 10, a spike at 7, minimum 3: volumes 8 and 9 kept; minimum 4 flags them; no spike, minimum 11: nothing flagged). The replica | – | yes |
| PIPE-69 | rule B | a criterion other than ‘<’ and ‘>’ is skipped; when every one is, TypeError: reduce() of empty iterable with no initial value. The replica | – | yes |
| PIPE-70 | rule A | NormalizeMotionParams on a one-row file | – | yes |
| PIPE-71 | rule A; nitransforms load | a file of one transform is loaded as one Affine (len(xfm) == 1: xfm = xfm[0]), whose 4 × 4 matrix FSLLinearTransformArray.from_ras iterates as four transforms (its rows) | – | yes |
| PIPE-72 | rule B; fMRIPrep FSLMotionParams | a reference whose centre of gravity is undefined (25.2.7: a constant image, every weight value - min zero; 25.2.6: a sum of zero, e.g. +1/-1 voxels; a NaN or infinite voxel in either) gives NaN or infinite translations in every row with numpy’s RuntimeWarning only, which FramewiseDisplacement turns into n/a everywhere (no FD spike). | – | yes |
| PIPE-73 | rule A; nitransforms ITKLinearTransform | every ITK text transform is read with genfromtxt(dtype=“f4”), the double ones too, so their parameters are rounded to float32. | – | yes |
| PIPE-74 | the rule of the corrections by default for undefined conversions | _binary_union’s int64 cast of NaN, ±inf and |x| >= 2^63 (and uint64 from 2^63, through float64) is undefined in C, INT64_MIN on x86-64 (outside the union). | – | yes |
| PIPE-75 | rule B | _binary_union broadcasts masks of different shapes silently; default ValueError (also for shapes numpy cannot broadcast), reproduce numpy’s broadcasting or its ValueError. | – | yes |
| PIPE-76 | the rule of the corrections by default; rule A for the integer wrap | acompcor_masks(is_aseg=True)'s uint8 cast of the GM map: integers wrap (-1 is 255, inside; 256 is 0, outside), floats out of int32’s range or NaN are undefined (x86: 0), in-range floats keep the truncated int32’s low byte (300.0 is 44); default clamped to [0, 255] first (GM = the voxels >= 1), reproduce numpy’s x86-64 cast; the notice counts the voxels the two put differently. | – | yes |
| PIPE-77 | rule B | BinaryDilation/image_binary_dilation with a negative radius (traits.Int takes it) dilates with scikit-image’s arbitrary footprint (-1: the dilated mask empty); default ValueError, reproduce niworkflows’ (with the fixed ball). | – | yes |
| PIPE-78 | rule B | SignalExtraction writes a class label holding a tab, \n or \r as it is (the table breaks); default ValueError, reproduce the file with a notice. | – | yes |
| PIPE-79 | rule D, open; fMRIPrep init_bold_confs_wf / aCompCorMasks | fMRIPrep passes is_aseg=freesurfer (run_reconall, its default), but t1w_tpms are always FAST’s partial-volume maps (sMRIPrep’s fast2bids), so with FreeSurfer the aseg path smooths partial volumes again (mask2vf) and its GM is only the voxels whose GM fraction is exactly 1 (the uint8 truncation), dilated by ball(3). | – | – |
| PIPE-80 | rule A | nipype’s command lines with a blank in a file name | – | yes |
| PIPE-81 | corrected; reproduce keeps it | compound inputs get the previous input’s type | – | yes |
| PIPE-82 | corrected; reproduce keeps it | texts not read as their traits read them | – | yes |
| PIPE-83 | kept, rule C | the command line’s limits | – | – |
| PD-01 | rule A | pandas’ float parser is not correctly rounded | – | yes |
| PD-02 | rule A | ujson’s 10-decimal formatting | – | yes |
| PD-03 | rule A | Integer columns with missing values | – | yes |
| NP-09 | numpy 2.5.4 on Zen 4/5, a crash: never reproduced | with GEMM_Q = 512 the Haswell kernel’s 12-column paths (24 or more columns in one kernel call) overflow their stack buffer (BUFFER1, L_BUFFER_SIZE = 256 · 8 · 12 + 4096 bytes: 298 steps of 12) on depth blocks above 298 steps | – | – |
| NP-10 | numpy 2.3.5 and 2.2.6, fixed in 2.5.4; rule B, a gap of reproduce | numpy’s svd (so pinv and nipype’s fallback_svd) and lstsq hand a matrix holding ±inf to LAPACK, whose scaling turns it into NaN: of 28 svd probe cases (probe_svd_inf.py) 7 hang (an infinite loop), 9 return infinite or NaN values without an error (most after three DLASCL “parameter number 5 had an illegal value” messages on standard output), 12 raise LinAlgError; of 40 lstsq cases (probe_lstsq_inf.py) 6 hang and 34 r… | – | – |
| SCI-41 | rule B; SciPy linalg.solve | SciPy 1.15 solves a diagonal system (every 1 × 1 one) by dividing, so a singular one gives inf or NaN with RuntimeWarnings only (its general path raises LinAlgError | – | yes |
| GLIBC-07 | expected | The correctly rounded cosine on fMRIPrep’s real CompCor files | – | yes |
Each decided by rules A to D (A: corrected by default; B: a clear error; C: kept and documented; D: measured first), with upstream_defects='reproduce' (LARMORX_UPSTREAM_DEFECTS=reproduce) giving the original’s behaviour where the replica has a switch (larmorx_core::defects, the Switch column); crashes (NP-09) and nondeterminism (PIPE-43’s column order) are never reproduced. Each entry links the findings’ section that describes it; the register holds its rows.
Speed
The benchmark (report) runs fMRIPrep’s nodes in-process on their own inputs of three fMRIPrep runs, and the replica on the same inputs, every result checked byte for byte.
- The 26 Python nodes run 1.1 to 1.4 times as fast as fMRIPrep’s at one thread, and 1.5 to 1.6 times at eight.
- numpy’s SVD and least squares, ported in Rust, are still slower than numpy’s own (their kernels without FMA dispatch): a speed item for later.
Seconds per call (the best of three runs) on a log scale; the factor is the original’s time divided by the replica’s, both on one thread.
- the originals, 1 thread (in-process)
- replica, 1 thread
- replica, 8 threads
- the whole workflow (bold_confounds), ds000005replica, 1 thread 5.86 sreplica, 8 threads 4.31 s
- the whole workflow (bold_confounds), ds004958replica, 1 thread 66.3 sreplica, 8 threads 49.7 s
- the whole workflow (bold_confounds), ds005454replica, 1 thread 96.5 sreplica, 8 threads 68.5 s
- dvars, ds000005the originals, 1 thread (in-process) 2.49 sreplica, 1 thread 1.18 sreplica, 8 threads 0.639 s2.1×
- motion_params, ds000005the originals, 1 thread (in-process) 0.0413 sreplica, 1 thread 0.028 sreplica, 8 threads 0.0258 s1.5×
- acc_masks, ds000005the originals, 1 thread (in-process) 0.651 sreplica, 1 thread 0.561 sreplica, 8 threads 0.148 s1.2×
- acompcor, ds000005the originals, 1 thread (in-process) 0.917 sreplica, 1 thread 0.631 sreplica, 8 threads 0.621 s1.5×
- crowncompcor, ds000005the originals, 1 thread (in-process) 1.22 sreplica, 1 thread 1.69 sreplica, 8 threads 1.7 s0.72×
- tcompcor, ds000005the originals, 1 thread (in-process) 1.04 sreplica, 1 thread 0.721 sreplica, 8 threads 0.754 s1.4×
- signals, ds000005the originals, 1 thread (in-process) 0.73 sreplica, 1 thread 0.401 sreplica, 8 threads 0.435 s1.8×
- dvars, ds004958the originals, 1 thread (in-process) 22.3 sreplica, 1 thread 22.6 sreplica, 8 threads 9.76 s0.99×
- motion_params, ds004958the originals, 1 thread (in-process) 0.124 sreplica, 1 thread 0.0741 sreplica, 8 threads 0.0764 s1.7×
- acc_masks, ds004958the originals, 1 thread (in-process) 1.44 sreplica, 1 thread 1.32 sreplica, 8 threads 0.299 s1.1×
- acompcor, ds004958the originals, 1 thread (in-process) 12.8 sreplica, 1 thread 8.91 sreplica, 8 threads 8.52 s1.4×
- crowncompcor, ds004958the originals, 1 thread (in-process) 16.4 sreplica, 1 thread 21 sreplica, 8 threads 16 s0.78×
- tcompcor, ds004958the originals, 1 thread (in-process) 14.2 sreplica, 1 thread 9.98 sreplica, 8 threads 9.33 s1.4×
- signals, ds004958the originals, 1 thread (in-process) 9.34 sreplica, 1 thread 5.74 sreplica, 8 threads 5.51 s1.6×
- dvars, ds005454the originals, 1 thread (in-process) 38.1 sreplica, 1 thread 28.9 sreplica, 8 threads 13.6 s1.3×
- motion_params, ds005454the originals, 1 thread (in-process) 0.107 sreplica, 1 thread 0.0755 sreplica, 8 threads 0.0744 s1.4×
- acc_masks, ds005454the originals, 1 thread (in-process) 1.71 sreplica, 1 thread 1.56 sreplica, 8 threads 0.477 s1.1×
- acompcor, ds005454the originals, 1 thread (in-process) 10 sreplica, 1 thread 11.7 sreplica, 8 threads 11.4 s0.86×
- crowncompcor, ds005454the originals, 1 thread (in-process) 12.4 sreplica, 1 thread 17.3 sreplica, 8 threads 16.7 s0.72×
- tcompcor, ds005454the originals, 1 thread (in-process) 21.1 sreplica, 1 thread 13.1 sreplica, 8 threads 12.2 s1.6×
- signals, ds005454the originals, 1 thread (in-process) 11.9 sreplica, 1 thread 8.06 sreplica, 8 threads 7.59 s1.5×
- numpy svd 280 x 60968the originals, 1 thread (in-process) 3.46 sreplica, 1 thread 11.7 sreplica, 8 threads 10.5 s0.3×
- numpy svd 597 x 30453the originals, 1 thread (in-process) 3.5 sreplica, 1 thread 11 sreplica, 8 threads 8.48 s0.32×
- numpy lstsq 600 x 7, 60968 voxelsthe originals, 1 thread (in-process) 0.573 sreplica, 1 thread 2.4 sreplica, 8 threads 1.98 s0.24×
- numpy lstsq 280 x 8, 60968 voxelsthe originals, 1 thread (in-process) 0.243 sreplica, 1 thread 1.02 sreplica, 8 threads 0.989 s0.24×
- numpy lstsq 600 x 25, 30453 voxelsthe originals, 1 thread (in-process) 0.563 sreplica, 1 thread 2.35 sreplica, 8 threads 2.3 s0.24×
- numpy lstsq 600 x 26, 60968 voxelsthe originals, 1 thread (in-process) 1.13 sreplica, 1 thread 5.15 sreplica, 8 threads 5.05 s0.22×
- SciPy gesvd 280 x 60968the originals, 1 thread (in-process) 3.59 sreplica, 1 thread 11.3 sreplica, 8 threads 11.2 s0.32×
- SciPy gesvd 597 x 30453the originals, 1 thread (in-process) 3.4 sreplica, 1 thread 10.3 sreplica, 8 threads 8.12 s0.33×
From the benchmark report docs/benchmarks/confounds.json (32 logical CPUs, the newest environment, numpy’s OpenBLAS at one thread as fMRIPrep’s nodes; a machine shared with other work; every result identical: each node’s files byte for byte in reproduce). The originals: nipype’s, niworkflows’ and fMRIPrep’s interfaces in-process on fMRIPrep’s own node inputs; the linear algebra: numpy’s and SciPy’s own calls. The whole workflow has no original in-process; fMRIPrep’s runs recorded ds000005 18.3 s, ds004958 191.2 s, ds005454 212.0 s for its nodes (fMRIPrep 25.2.6 on another machine, the old laptop). The linear algebra’s routines are slower than numpy’s and SciPy’s (their level-1 and level-2 kernels without FMA dispatch, one thread): a speed item of the handoff (docs/benchmarks/confounds.md, “Where the time goes”).
Memory
The memory rule allows LARMORX at most 10 % above the original’s peak plus 8 MiB for the same call; here the replica needs at most 0.96 of the originals’ peak.
Peak memory, the replica against the originals (MiB, the memory rule’s calls)
| Call (fMRIPrep's node) | The original | replica, reproduce, 1 thread | replica, fix, 1 thread | replica, fix, 8 threads | replica ÷ the original |
|---|---|---|---|---|---|
| acompcor, ds000005 | 236 | 177 | 177 | 177 | 0.752 |
| tcompcor, ds000005 | 368 | 343 | 343 | 353 | 0.958 |
| dvars, ds000005 | 424 | 194 | 194 | 194 | 0.457 |
| signals, ds000005 | 507 | 171 | 170 | 171 | 0.337 |
| acompcor, ds000228 | 193 | 130 | 130 | 131 | 0.676 |
| tcompcor, ds000228 | 303 | 275 | 275 | 282 | 0.928 |
Each call in a process of its own (validation/investigations/confounds/memory_call.py, the calls of tests/parity/test_memory_peak.py: fMRIPrep’s acompcor, tcompcor, dvars and signals nodes with their own inputs), its peak resident memory from /usr/bin/time -v, in the newest environment. The rule: at most 10 % above the original’s peak plus 8 MiB: 6 of the 6 calls meet it.
Random testing
Random testing runs random valid inputs and option combinations through the originals and the replica and compares every output (how random testing works). A difference is shrunk to the smallest input and fewest options that still show it, then fixed and kept as a permanent case, registered as a defect of the original, or reported. The replica’s run of 2026-11-12 (30 minutes) and two checks of the final build compared 514,694 random calls.
Random testing, the replica against the originals: fMRIPrep 25.2.7, nipype 1.12.0, niworkflows 1.15.4, numpy, SciPy and pandas (the run of 2026-11-12 and the checks of the final build)
| 2026-11-12-confounds-trial | 2026-11-12-confounds-check | 2026-11-12-confounds-check-noavx512 | |
|---|---|---|---|
| Random cases, each compared with the original | 334,124 | 140,003 | 40,567 |
| confounds-linalg | 111,478 cases: 111,478 agree; 76 options drawn | 46,698 cases: 46,698 agree; 76 options drawn | – |
| confounds-nipype | 111,355 cases: 110,988 agree, 28 differ, 339 notice; 98 options drawn | 46,658 cases: 46,658 agree; 98 options drawn | – |
| confounds-fmriprep | 111,291 cases: 109,610 agree, 189 notice, 1,492 skipped; 102 options drawn | 46,647 cases: 46,002 agree, 645 skipped; 103 options drawn | 40,567 cases: 40,567 agree; 102 options drawn |
| length (minutes) | 30 | 10.2 | 5.4 |
| seed, the replica build | confounds-trial, v0.1.0rc1-218-g7ade9fbdd1bd | confounds-check, v0.1.0rc1-221-g9892deaf0f24 | confounds-check-noavx512, v0.1.0rc1-221-g9892deaf0f24 |
Differential random testing (docs/validation/difftest.md): one tool per suite (the linear algebra; nipype’s DVARS, CompCor, TSNR, non-steady-state detector and framewise displacement; fMRIPrep’s and niworkflows’ table, model, motion, mask and signal steps), each case a random call of the original in-process (numpy, SciPy, nitime, nipype 1.12.0, niworkflows, pandas, fMRIPrep 25.2.7’s and 25.2.6’s interfaces) against the replica in reproduce bit for bit, then its default checked by the rule of the corrections by default; 3 GiB a case, one thread. notice: a default that named a correction where the rule expected none (triaged below). skipped: ExpandModel’s powers above 2 while numpy’s pow is SVML’s (AVX-512); the third run turns those kernels off and compares them.
The replica’s differences from the originals in the runs of 2026-11-12, triaged
| Status | Tool | What it was | Register |
|---|---|---|---|
| fixedvalidation/difftest/cases/confounds-nipype/tcompcor-refused-degree-text.json | confounds-nipype | Before the trial (probe 148): CompCor’s refused inputs raised the replica’s own messages (‘regress_poly_degree must be at least 1, …’); every one now raises traits’ TraitError text (as ValueError) or nipype’s OSError of two exclusive inputs and its ValueError of a missing mask_files, checked against nipype for the three interfaces (relay 3a-iii, 2026-11-12). | – |
| harness | confounds-nipype | Trial 2026-11-12-confounds-trial: 20 cases with two refused inputs at once (regress_poly_degree 0 beside a refused criterion or mask_index; findings 2721, 4721, 22868, 70178, …): nipype names the first refused input its caller set, an order a function with keyword parameters does not know (a finding’s recipe, written with its keys sorted, changes nipype’s choice: 10 of the 20 agree on replay). The descriptor draws one refused input at a time since (relay 3a-iii). | – |
| fixedvalidation/difftest/cases/confounds-nipype/compcor-no-volume-left.json | confounds-nipype | Trial 2026-11-12-confounds-trial: 7 cases. (1) No volume left (ignore_initial_volumes at least the series’ length; findings 3005, 31312): nipype’s pre-filters reshape a mask’s rows to (-1, 0), numpy’s ‘cannot reshape array of size 0 into shape (0)’, unless the mask is empty and failure_mode is not ‘error’, in nipype’s mask order (its loop stops at a mask that gives no component); the replica skipped the reshape. (2) A mask of another shape after one that gives no component (find… | – |
| fixedvalidation/difftest/cases/confounds-nipype/regress-poly-fortran-rows.json | confounds-nipype | Before the trial (probe3 485): regress_poly of Fortran-ordered rows (a 2-D array, or a nibabel-like array whose reshape is a view) with degree 0: numpy’s pinv(X).dot(data.T) takes gemv on a C-ordered data.T for the one row of pinv(X), whose sums run in another order; the replica summed as for C rows (one or two ulps). Fixed: regress_poly_layout; probe_regress_layout.py 264 of 264 (relay 3a-iii). | – |
| fixedvalidation/difftest/cases/confounds-nipype/cosine-filter-glibc07-exact.json | confounds-nipype | Trial 2026-11-12-confounds-trial: 128 cosine_filter cases whose glibc and correctly rounded cosines differ only before the product with sqrt(2 / T) rounds them to one value: GLIBC-07 named with the basis unchanged. The witness now compares the basis values (relay 3a-iii); every one agrees on replay. | – |
| fixedvalidation/difftest/cases/confounds-nipype/dvars-constant-series-notices.json | confounds-nipype | Trial 2026-11-12-confounds-trial: 206 compute_dvars and ComputeDVARS cases of constant or all-zero series with remove_zerovariance=False (111 + 95): PIPE-50’s and PIPE-46’s corrections leave every output NaN, as nipype’s (SciPy 1.15), yet were named. fix now names PIPE-46, -47 and -50 only where its series differ from nipype’s (larmorx.confounds.dvars; SciPy 1.18’s LinAlgError and nipype’s RuntimeWarning count as different); every one agrees on replay (relay 3a-iii). | – |
| fixedvalidation/difftest/cases/confounds-nipype/compcor-pipe44-later-masks-add-nothing.json | confounds-nipype | Trial 2026-11-12-confounds-trial: 5 ACompCor cases whose masks after the one that gives no component give none either: fix goes on with them (PIPE-44) to the same result, yet named it. Named now where a later mask adds components or changes the pre-filter’s basis (relay 3a-iii); every one agrees on replay. | – |
| fixedvalidation/difftest/cases/confounds-nipype/cosine-filter-wide-basis-dbdsqr-tolerance.json | confounds-nipype | Trial 2026-11-12-confounds-trial, finding 54131: cosine_filter with a cut-off of 5 s at a TR of 3 s on 370 volumes (a cosine basis of 444 columns, more than the volumes): the replica’s lstsq differed from numpy 2.3.5’s LAPACK by 1 to 4 ulps in a cluster of near-equal singular values (141,010 of 179,080 residuals, at most 6e-14). Found (relay 3a-iv): the whole DBDSDC is one subproblem at DBDSQR’s real epsilon (2^-53; relay 3a-iii’s splits used numpy’s 2^-52), whose DLASD0 differs… | – |
| fixedvalidation/difftest/cases/confounds-fmriprep/rename-acompcor-text-retained.json | confounds-fmriprep | Before the trial (probe 140): RenameACompCor (and FilterDropped) with a text ‘retained’ column: pandas’ KeyError prints the values’ Index with line breaks at width 80, the replica printed one line. Fixed: _index_repr follows pandas’ format_object_summary (both pandas releases); probe_index_repr.py 260 of 260 in pandas 3.0.6 and 2.2.3 (relay 3a-iii). | – |
| fixedvalidation/difftest/cases/confounds-fmriprep/spikes-pipe67-after-filling.json | confounds-fmriprep | Trial 2026-11-12-confounds-trial: 22 SpikeRegressors cases with several lags and minimum_contiguous: the filling of short runs made both lag rules flag the same volumes, yet PIPE-67 was named. Named now where the flagged volumes differ after the filling (larmorx.niworkflows.confounds); every one agrees on replay (relay 3a-iii). | – |
| explained | confounds-fmriprep | Trial 2026-11-12-confounds-trial: 167 cases (155 ExpandModel, 12 fMRIPrep’s FramewiseDisplacement) and the probes’ SpikeRegressors without concatenate: PD-01 names the table read’s correction, whose values change, while the step’s output may not show it (the expansion writes the columns its formula takes, the displacement’s sums of differences round a value read one bit off to the same number, the thresholded flags do not move). The runner counts PD-01 as a condition notice … | PD-01 |
validation/difftest/triage.toml (the confounds, 2026-11-12): each difference of the probes before the trial and of the trial, a the replica mistake fixed with a permanent case (10 in all, validation/difftest/cases/confounds-*/, run by the records), a harness mistake, or a defect of the original explained by the register.
Every case
Each case of the records, with what it tests, its command line, its verdict and the reason. Search, filter by category or verdict (or select a segment of the chart above), sort by a column; #case=<name> links to one.
88 cases
| Case | Category | What it tests | Verdict | Checks |
|---|---|---|---|---|
confounds-fmriprep/tables-quick-e1 | the confounds table and its JSON | pandas' text round trips (precise_xstrtod of pandas 3.0.6 and 2.2.3, ujson), fMRIPrep's FramewiseDisplacement, FilterDropped, RenameACompCor, GatherConfounds, niworkflows' AddTSVHeader, DictMerge, TSV2JSON, fMRIPrep's real nodes (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_tables.py --tier quick | Bit-identical Every one of its 79 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 79/79 |
confounds-fmriprep/tables-quick-e2 | the confounds table and its JSON | pandas' text round trips (precise_xstrtod of pandas 3.0.6 and 2.2.3, ujson), fMRIPrep's FramewiseDisplacement, FilterDropped, RenameACompCor, GatherConfounds, niworkflows' AddTSVHeader, DictMerge, TSV2JSON, fMRIPrep's real nodes (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_tables.py --tier quick | Bit-identical Every one of its 79 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 79/79 |
confounds-fmriprep/tables-quick-e3 | the confounds table and its JSON | pandas' text round trips (precise_xstrtod of pandas 3.0.6 and 2.2.3, ujson), fMRIPrep's FramewiseDisplacement, FilterDropped, RenameACompCor, GatherConfounds, niworkflows' AddTSVHeader, DictMerge, TSV2JSON, fMRIPrep's real nodes (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_tables.py --tier quick | Bit-identical Every one of its 83 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 83/83 |
confounds-fmriprep/model-quick-e1 | the motion model's columns | niworkflows' ExpandModel, parse_formula, SpikeRegressors, NormalizeMotionParams, fMRIPrep's real model_expansion and spike_regressors nodes (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_model.py --tier quick | Bit-identical Every one of its 52 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 52/52 |
confounds-fmriprep/model-quick-e2 | the motion model's columns | niworkflows' ExpandModel, parse_formula, SpikeRegressors, NormalizeMotionParams, fMRIPrep's real model_expansion and spike_regressors nodes (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_model.py --tier quick | Bit-identical Every one of its 52 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 52/52 |
confounds-fmriprep/model-quick-e3 | the motion model's columns | niworkflows' ExpandModel, parse_formula, SpikeRegressors, NormalizeMotionParams, fMRIPrep's real model_expansion and spike_regressors nodes (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_model.py --tier quick | Bit-identical Every one of its 52 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 52/52 |
confounds-fmriprep/motion-quick-e1 | the motion parameters and framewise displacement | SciPy's Rotation, nitransforms' ITK and FSL linear transforms, fMRIPrep's FSLMotionParams (25.2.7 and 25.2.6) and FSLRMSDeviation, fMRIPrep's real nodes (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_motion.py --tier quick | Bit-identical Every one of its 122 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 122/122 |
confounds-fmriprep/motion-quick-e2 | the motion parameters and framewise displacement | SciPy's Rotation, nitransforms' ITK and FSL linear transforms, fMRIPrep's FSLMotionParams (25.2.7 and 25.2.6) and FSLRMSDeviation, fMRIPrep's real nodes (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_motion.py --tier quick | Bit-identical Every one of its 122 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 122/122 |
confounds-fmriprep/motion-quick-e3 | the motion parameters and framewise displacement | SciPy's Rotation, nitransforms' ITK and FSL linear transforms, fMRIPrep's FSLMotionParams (25.2.7 and 25.2.6) and FSLRMSDeviation, fMRIPrep's real nodes (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_motion.py --tier quick | Bit-identical Every one of its 123 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 123/123 |
confounds-fmriprep/masks-quick-e1 | the CompCor masks | fMRIPrep's _binary_union, aCompCorMasks, niworkflows' BinaryDilation, BinarySubtraction and SignalExtraction, fMRIPrep's real mask and signals nodes (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_masks.py --tier quick | Bit-identical Every one of its 42 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 42/42 |
confounds-fmriprep/masks-quick-e2 | the CompCor masks | fMRIPrep's _binary_union, aCompCorMasks, niworkflows' BinaryDilation, BinarySubtraction and SignalExtraction, fMRIPrep's real mask and signals nodes (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_masks.py --tier quick | Bit-identical Every one of its 42 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 42/42 |
confounds-fmriprep/masks-quick-e3 | the CompCor masks | fMRIPrep's _binary_union, aCompCorMasks, niworkflows' BinaryDilation, BinarySubtraction and SignalExtraction, fMRIPrep's real mask and signals nodes (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_masks.py --tier quick | Bit-identical Every one of its 42 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 42/42 |
confounds-fmriprep/workflow-quick-e2 | the whole workflow | init_bold_confs_wf built by nipype for eight option sets against the replica's steps, and the 36 confounds workflows end to end: every node file, fMRIPrep's 36 confounds TSV files column by column and their JSON byte for byte (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_workflow.py --tier quick | Bit-identical Every one of its 48 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 48/48 |
confounds-fmriprep/workflow-quick-e1 | the whole workflow | init_bold_confs_wf built by nipype for eight option sets against the replica's steps, and the 36 confounds workflows end to end: every node file, fMRIPrep's 36 confounds TSV files column by column and their JSON byte for byte (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_workflow.py --tier quick | Bit-identical Every one of its 48 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 48/48 |
confounds-fmriprep/difftest/rename-acompcor-text-retained | random testing | RenameACompCor with a 'retained' column of text: pandas' KeyError prints the values' Index with line breaks at width 80, the replica in one line. Fixed: _index_repr follows pandas' format_object_summary. | Bit-identical Every one of its 2 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A permanent case of random testing (a the replica mistake, fixed). | 2/2 |
confounds-fmriprep/difftest/spikes-pipe67-after-filling | random testing | SpikeRegressors with lags [1, 2] and minimum_contiguous 5 (the trial's notice verdicts): the filling of short runs made both lag rules flag the same volumes, yet PIPE-67 fired. Fixed: PIPE-67 is named where the flagged volumes differ after the filling. | Bit-identical Every one of its 2 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A permanent case of random testing (a the replica mistake, fixed). | 2/2 |
confounds-nipype/dvars-quick-e1 | DVARS | numpy's complex64 product, nitime's crosscov and AR(1), SciPy's 1 x 1 solve, nipype's compute_dvars and ComputeDVARS, fMRIPrep's 36 real dvars nodes (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_dvars.py --tier quick | Bit-identical Every one of its 20 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 20/20 |
confounds-nipype/dvars-quick-e2 | DVARS | numpy's complex64 product, nitime's crosscov and AR(1), SciPy's 1 x 1 solve, nipype's compute_dvars and ComputeDVARS, fMRIPrep's 36 real dvars nodes (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_dvars.py --tier quick | Bit-identical Every one of its 20 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 20/20 |
confounds-nipype/dvars-quick-e3 | DVARS | numpy's complex64 product, nitime's crosscov and AR(1), SciPy's 1 x 1 solve, nipype's compute_dvars and ComputeDVARS, fMRIPrep's 36 real dvars nodes (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_dvars.py --tier quick | Bit-identical Every one of its 20 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 20/20 |
confounds-nipype/dvars-control-quick-e1 | DVARS | the negative control of check_dvars.py: the replica told the other library release must differ (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_dvars.py --control --sections cmul,crosscov,solve,ar,dvars,node --tier quick | Bit-identical Every one of its 16 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 16/16 |
confounds-nipype/dvars-control-quick-e2 | DVARS | the negative control of check_dvars.py: the replica told the other library release must differ (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_dvars.py --control --sections cmul,crosscov,solve,ar,dvars,node --tier quick | Bit-identical Every one of its 16 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 16/16 |
confounds-nipype/dvars-control-quick-e3 | DVARS | the negative control of check_dvars.py: the replica told the other library release must differ (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_dvars.py --control --sections cmul,crosscov,solve,ar,dvars,node --tier quick | Bit-identical Every one of its 16 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 16/16 |
confounds-nipype/compcor-quick-e1 | CompCor | nipype's cosine and Legendre filters, compute_noise_components, CompCor, ACompCor and TCompCor over their options, fMRIPrep's 108 real CompCor nodes (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_compcor.py --tier quick | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 30/30 |
confounds-nipype/compcor-quick-e2 | CompCor | nipype's cosine and Legendre filters, compute_noise_components, CompCor, ACompCor and TCompCor over their options, fMRIPrep's 108 real CompCor nodes (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_compcor.py --tier quick | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 30/30 |
confounds-nipype/compcor-quick-e3 | CompCor | nipype's cosine and Legendre filters, compute_noise_components, CompCor, ACompCor and TCompCor over their options, fMRIPrep's 108 real CompCor nodes (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_compcor.py --tier quick | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 30/30 |
confounds-nipype/compcor-control-quick-e1 | CompCor | the negative control of check_compcor.py: the replica told the other library release must differ (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_compcor.py --control --tier quick | Bit-identical Every one of its 29 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 29/29 |
confounds-nipype/compcor-control-quick-e3 | CompCor | the negative control of check_compcor.py: the replica told the other library release must differ (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_compcor.py --control --tier quick | Bit-identical Every one of its 29 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 29/29 |
confounds-nipype/misc-quick-e1 | nipype's other interfaces | nipype's TSNR, NonSteadyStateDetector, FramewiseDisplacement (every source, NIPY included) and niworkflows' RobustACompCor/RobustTCompCor (PIPE-48) (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_nipype_misc.py --tier quick | Bit-identical Every one of its 68 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 68/68 |
confounds-nipype/misc-quick-e2 | nipype's other interfaces | nipype's TSNR, NonSteadyStateDetector, FramewiseDisplacement (every source, NIPY included) and niworkflows' RobustACompCor/RobustTCompCor (PIPE-48) (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_nipype_misc.py --tier quick | Bit-identical Every one of its 69 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 69/69 |
confounds-nipype/misc-quick-e3 | nipype's other interfaces | nipype's TSNR, NonSteadyStateDetector, FramewiseDisplacement (every source, NIPY included) and niworkflows' RobustACompCor/RobustTCompCor (PIPE-48) (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_nipype_misc.py --tier quick | Bit-identical Every one of its 69 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 69/69 |
confounds-nipype/misc-control-quick-e1 | nipype's other interfaces | the negative control of check_nipype_misc.py: the replica told the other library release must differ (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_nipype_misc.py --control --sections tsnr,nssd,fd,nipy --tier quick | Bit-identical Every one of its 46 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 46/46 |
confounds-nipype/misc-control-quick-e3 | nipype's other interfaces | the negative control of check_nipype_misc.py: the replica told the other library release must differ (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_nipype_misc.py --control --sections tsnr,nssd,fd,nipy --tier quick | Bit-identical Every one of its 46 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 46/46 |
confounds-nipype/cli-quick-e1 | nipype's command line | The replica nipype run against nipype's own nipypecli run: the click layer, the 61 helps, argparse's and traits' errors, the 37 interfaces the replica runs with their options, PIPE-81 and PIPE-82 against nipype's Python API, fMRIPrep's real confounds nodes of three fMRIPrep runs (full comparison) (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_nipype_cli.py --tier quick | Bit-identical Every one of its 18 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 18/18 |
confounds-nipype/difftest/compcor-no-volume-left | random testing | random testing: DN fix pass | Bit-identical Every one of its 2 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A permanent case of random testing (a the replica mistake, fixed). | 2/2 |
confounds-nipype/difftest/compcor-pipe44-later-masks-add-nothing | random testing | ACompCor whose second mask gives no component and whose later masks give none either (the trial's notice verdicts): fix goes on with them (PIPE-44) to the same result, yet named PIPE-44. Fixed: named where a later mask adds components or changes the pre-filter's basis. | Bit-identical Every one of its 2 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A permanent case of random testing (a the replica mistake, fixed). | 2/2 |
confounds-nipype/difftest/compcor-pipe44-reaches-other-shape | random testing | random testing: DN fix pass | Bit-identical Every one of its 2 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A permanent case of random testing (a the replica mistake, fixed). | 2/2 |
confounds-nipype/difftest/cosine-filter-glibc07-exact | random testing | cosine_filter whose glibc and correctly rounded cosines differ only before the product with sqrt(2 / T) rounds them to one value (the trial's notice verdicts): GLIBC-07 fired with the basis unchanged. Fixed: the witness compares the basis values. | Bit-identical Every one of its 2 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A permanent case of random testing (a the replica mistake, fixed). | 2/2 |
confounds-nipype/difftest/cosine-filter-wide-basis-dbdsqr-tolerance | random testing | cosine_filter with a cut-off of 5 s at a TR of 3 s on 370 volumes: a cosine basis of 444 columns, wider than the series, whose lstsq (DGELSD) and svd (DGESDD, DBDSDC's lower-bidiagonal path) reach DLASDQ leaves with nearly equal diagonals and superdiagonals at DBDSQR's convergence tolerance (random-testing finding 54131: 1 to 4 ulps). Fixed: DBDSQR's TOLMUL is glibc's pow(2^-53, -0.125) = 2^6.625 correctly rounded (the constant was 5.8e-6 off). | Bit-identical Every one of its 2 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A permanent case of random testing (a the replica mistake, fixed). | 2/2 |
confounds-nipype/difftest/dvars-constant-series-notices | random testing | compute_dvars of an all-zero series with remove_zerovariance=False (the trial's notice verdicts): PIPE-50's and PIPE-46's corrections leave every output NaN, as nipype's, yet their notices fired. Fixed: named only where the corrected series differ from nipype's. | Bit-identical Every one of its 2 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A permanent case of random testing (a the replica mistake, fixed). | 2/2 |
confounds-nipype/difftest/regress-poly-fortran-rows | random testing | regress_poly of Fortran-ordered rows (387 x 69 float32, degree 0): numpy's pinv(X).dot(data.T) takes gemv on a C-ordered data.T, whose sums run in another order; the replica summed as for C-ordered rows (16,284 of 26,703 residuals one or two ulps off). Fixed: regress_poly_layout follows the rows' memory order. | Bit-identical Every one of its 2 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A permanent case of random testing (a the replica mistake, fixed). | 2/2 |
confounds-nipype/difftest/tcompcor-refused-degree-text | random testing | TCompCor with regress_poly_degree=0: traits' TraitError text ('The 'regress_poly_degree' trait of a TCompCorInputSpec instance must be an integer >= 1, ...') where the replica raised its own message. Fixed: every refused CompCor input raises traits' or nipype's text. | Bit-identical Every one of its 2 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A permanent case of random testing (a the replica mistake, fixed). | 2/2 |
confounds-nipype/compcor-full-e1 | CompCor | nipype's cosine and Legendre filters, compute_noise_components, CompCor, ACompCor and TCompCor over their options, fMRIPrep's 108 real CompCor nodes (full comparison, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_compcor.py --tier full | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 30/30 |
confounds-nipype/compcor-full-e2 | CompCor | nipype's cosine and Legendre filters, compute_noise_components, CompCor, ACompCor and TCompCor over their options, fMRIPrep's 108 real CompCor nodes (full comparison, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_compcor.py --tier full | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 30/30 |
confounds-nipype/compcor-full-e3 | CompCor | nipype's cosine and Legendre filters, compute_noise_components, CompCor, ACompCor and TCompCor over their options, fMRIPrep's 108 real CompCor nodes (full comparison, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_compcor.py --tier full | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 30/30 |
confounds-linalg/blas-quick-e2 | OpenBLAS' kernels | OpenBLAS's dgemm, dsyrk, dgemv, dtrmm, dger, dtrmv, the level-1 routines and numpy's @ and dot against numpy's library (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_blas.py --quick --threads 4 | Bit-identical Every one of its 27 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 27/27 |
confounds-linalg/blas-quick-e1 | OpenBLAS' kernels | OpenBLAS's dgemm, dsyrk, dgemv, dtrmm, dger, dtrmv, the level-1 routines and numpy's @ and dot against numpy's library (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_blas.py --quick --threads 4 | Bit-identical Every one of its 27 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 27/27 |
confounds-linalg/lapack-quick-e2 | LAPACK | LAPACK's blocked QR and LQ (DGEQRF, DGELQF, DLARFT, DLARFB), DORGQR/DORGLQ/DORMQR/DORMLQ, DORMBR/DORGBR and the bidiagonal reduction (DGEBRD) against numpy's library (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_lapack.py --quick --threads 4 | Bit-identical Every one of its 48 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 48/48 |
confounds-linalg/lapack-quick-e1 | LAPACK | LAPACK's blocked QR and LQ (DGEQRF, DGELQF, DLARFT, DLARFB), DORGQR/DORGLQ/DORMQR/DORMLQ, DORMBR/DORGBR and the bidiagonal reduction (DGEBRD) against numpy's library (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_lapack.py --quick --threads 4 | Bit-identical Every one of its 48 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 48/48 |
confounds-linalg/svd-quick-e2 | numpy's and SciPy's SVD | the divide-and-conquer bidiagonal SVD (DBDSDC, DLASD0 to DLASD5, DLASDQ, DLAED6) against numpy's library (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_svd.py --quick --threads 4 | Bit-identical Every one of its 27 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 27/27 |
confounds-linalg/svd-quick-e1 | numpy's and SciPy's SVD | the divide-and-conquer bidiagonal SVD (DBDSDC, DLASD0 to DLASD5, DLASDQ, DLAED6) against numpy's library (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_svd.py --quick --threads 4 | Bit-identical Every one of its 27 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 27/27 |
confounds-linalg/svd-control-quick-e2 | numpy's and SciPy's SVD | the negative control of check_svd.py: the replica told the other library release must differ (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_svd.py --control --quick --threads 4 | Bit-identical Every one of its 17 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 17/17 |
confounds-linalg/svd-control-quick-e1 | numpy's and SciPy's SVD | the negative control of check_svd.py: the replica told the other library release must differ (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_svd.py --control --quick --threads 4 | Bit-identical Every one of its 17 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 17/17 |
confounds-linalg/gesdd-quick-e2 | LAPACK's divide-and-conquer SVD | DGESDD's four thin paths, numpy's svd(full_matrices=False) and pinv up to CompCor's largest real shapes (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesdd.py --quick --threads 4 | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 30/30 |
confounds-linalg/gesdd-quick-e1 | LAPACK's divide-and-conquer SVD | DGESDD's four thin paths, numpy's svd(full_matrices=False) and pinv up to CompCor's largest real shapes (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesdd.py --quick --threads 4 | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 30/30 |
confounds-linalg/gesdd-control-quick-e2 | LAPACK's divide-and-conquer SVD | the negative control of check_gesdd.py: the replica told the other library release must differ (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesdd.py --control --quick --threads 4 | Bit-identical Every one of its 18 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 18/18 |
confounds-linalg/gesdd-control-quick-e1 | LAPACK's divide-and-conquer SVD | the negative control of check_gesdd.py: the replica told the other library release must differ (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesdd.py --control --quick --threads 4 | Bit-identical Every one of its 18 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 18/18 |
confounds-linalg/gelsd-quick-e2 | LAPACK's least squares | DGELSD's three paths, DLALSD's two (DLASDA, DLALS0, DLALSA), DLASRT and numpy's lstsq, the cosine filter's real calls included (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gelsd.py --quick --threads 4 | Bit-identical Every one of its 106 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 106/106 |
confounds-linalg/gelsd-quick-e1 | LAPACK's least squares | DGELSD's three paths, DLALSD's two (DLASDA, DLALS0, DLALSA), DLASRT and numpy's lstsq, the cosine filter's real calls included (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gelsd.py --quick --threads 4 | Bit-identical Every one of its 106 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 106/106 |
confounds-linalg/gelsd-control-quick-e2 | LAPACK's least squares | the negative control of check_gelsd.py: the replica told the other library release must differ (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gelsd.py --control --quick --threads 4 | Bit-identical Every one of its 80 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 80/80 |
confounds-linalg/gelsd-control-quick-e1 | LAPACK's least squares | the negative control of check_gelsd.py: the replica told the other library release must differ (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gelsd.py --control --quick --threads 4 | Bit-identical Every one of its 80 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 80/80 |
confounds-linalg/gesvd-quick-e2 | LAPACK's SVD | SciPy's DGESVD fallback (DBDSQR, OpenBLAS 0.3.28's DLARFT) and nipype's fallback_svd and _full_rank against SciPy's own library (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesvd.py --quick --threads 4 | Bit-identical Every one of its 83 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 83/83 |
confounds-linalg/gesvd-quick-e1 | LAPACK's SVD | SciPy's DGESVD fallback (DBDSQR, OpenBLAS 0.3.28's DLARFT) and nipype's fallback_svd and _full_rank against SciPy's own library (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesvd.py --quick --threads 4 | Bit-identical Every one of its 83 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 83/83 |
confounds-linalg/gesvd-control-quick-e2 | LAPACK's SVD | the negative control of check_gesvd.py: the replica told the other library release must differ (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesvd.py --control --quick --threads 4 | Bit-identical Every one of its 87 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 87/87 |
confounds-linalg/gesvd-control-quick-e1 | LAPACK's SVD | the negative control of check_gesvd.py: the replica told the other library release must differ (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesvd.py --control --quick --threads 4 | Bit-identical Every one of its 87 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 87/87 |
confounds-linalg/reductions-quick-e1 | numpy's reductions | numpy's ordered sums, means, variances, cumsum, extremes, percentiles, searchsorted, nan_to_num, diff and the Legendre basis at the confounds' call sites (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_numpy_reductions.py --tier quick | Bit-identical Every one of its 18 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 18/18 |
confounds-linalg/reductions-quick-e2 | numpy's reductions | numpy's ordered sums, means, variances, cumsum, extremes, percentiles, searchsorted, nan_to_num, diff and the Legendre basis at the confounds' call sites (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_numpy_reductions.py --tier quick | Bit-identical Every one of its 18 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 18/18 |
confounds-linalg/reductions-quick-e3 | numpy's reductions | numpy's ordered sums, means, variances, cumsum, extremes, percentiles, searchsorted, nan_to_num, diff and the Legendre basis at the confounds' call sites (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_numpy_reductions.py --tier quick | Bit-identical Every one of its 18 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 18/18 |
confounds-linalg/reductions-control-quick-e1 | numpy's reductions | the negative control of check_numpy_reductions.py: the replica told the other library release must differ (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_numpy_reductions.py --control --tier quick | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 30/30 |
confounds-linalg/reductions-control-quick-e2 | numpy's reductions | the negative control of check_numpy_reductions.py: the replica told the other library release must differ (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_numpy_reductions.py --control --tier quick | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 30/30 |
confounds-linalg/reductions-control-quick-e3 | numpy's reductions | the negative control of check_numpy_reductions.py: the replica told the other library release must differ (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_numpy_reductions.py --control --tier quick | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 30/30 |
confounds-linalg/fft-quick-e1 | SciPy's FFTs | SciPy's float32 FFTs (1.15's pocketfft, 1.18.1's ducc0) and nitime's fftconvolve under nipype's DVARS (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_fft.py --tier quick | Bit-identical Every one of its 81 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 81/81 |
confounds-linalg/fft-quick-e2 | SciPy's FFTs | SciPy's float32 FFTs (1.15's pocketfft, 1.18.1's ducc0) and nitime's fftconvolve under nipype's DVARS (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_fft.py --tier quick | Bit-identical Every one of its 81 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 81/81 |
confounds-linalg/fft-quick-e3 | SciPy's FFTs | SciPy's float32 FFTs (1.15's pocketfft, 1.18.1's ducc0) and nitime's fftconvolve under nipype's DVARS (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_fft.py --tier quick | Bit-identical Every one of its 81 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 81/81 |
confounds-linalg/fft-control-quick-e1 | SciPy's FFTs | the negative control of check_fft.py: the replica told the other library release must differ (quick tier, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_fft.py --control --tier quick | Bit-identical Every one of its 80 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 80/80 |
confounds-linalg/fft-control-quick-e2 | SciPy's FFTs | the negative control of check_fft.py: the replica told the other library release must differ (quick tier, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_fft.py --control --tier quick | Bit-identical Every one of its 80 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 80/80 |
confounds-linalg/fft-control-quick-e3 | SciPy's FFTs | the negative control of check_fft.py: the replica told the other library release must differ (quick tier, fMRIPrep 25.2.6's image: numpy 2.2.6, SciPy 1.15.2, pandas 2.2.3)command linepython -P validation/investigations/confounds/check_fft.py --control --tier quick | Bit-identical Every one of its 80 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. A negative control: told the other library release, the replica’s results differ from the original’s, as they must. | 80/80 |
confounds-linalg/svd-full-e2 | numpy's and SciPy's SVD | the divide-and-conquer bidiagonal SVD (DBDSDC, DLASD0 to DLASD5, DLASDQ, DLAED6) against numpy's library (full comparison, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_svd.py --threads 4 | Bit-identical Every one of its 27 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 27/27 |
confounds-linalg/svd-full-e1 | numpy's and SciPy's SVD | the divide-and-conquer bidiagonal SVD (DBDSDC, DLASD0 to DLASD5, DLASDQ, DLAED6) against numpy's library (full comparison, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_svd.py --threads 4 | Bit-identical Every one of its 27 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 27/27 |
confounds-linalg/svd-full-e1-zen | numpy's and SciPy's SVD | the divide-and-conquer bidiagonal SVD (DBDSDC, DLASD0 to DLASD5, DLASDQ, DLAED6) against numpy's library (full comparison, numpy 2.5.4 / OpenBLAS 0.3.34 at this machine's GEMM_Q 512)command linepython -P validation/investigations/confounds/check_svd.py --threads 4 | Bit-identical Every one of its 32 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 32/32 |
confounds-linalg/gesdd-full-e2 | LAPACK's divide-and-conquer SVD | DGESDD's four thin paths, numpy's svd(full_matrices=False) and pinv up to CompCor's largest real shapes (full comparison, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesdd.py --threads 4 | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 30/30 |
confounds-linalg/gesdd-full-e1 | LAPACK's divide-and-conquer SVD | DGESDD's four thin paths, numpy's svd(full_matrices=False) and pinv up to CompCor's largest real shapes (full comparison, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesdd.py --threads 4 | Bit-identical Every one of its 30 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 30/30 |
confounds-linalg/gesdd-full-e1-zen | LAPACK's divide-and-conquer SVD | DGESDD's four thin paths, numpy's svd(full_matrices=False) and pinv up to CompCor's largest real shapes (full comparison, numpy 2.5.4 / OpenBLAS 0.3.34 at this machine's GEMM_Q 512)command linepython -P validation/investigations/confounds/check_gesdd.py --threads 4 | Bit-identical Every one of its 45 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 45/45 |
confounds-linalg/gelsd-full-e2 | LAPACK's least squares | DGELSD's three paths, DLALSD's two (DLASDA, DLALS0, DLALSA), DLASRT and numpy's lstsq, the cosine filter's real calls included (full comparison, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gelsd.py --threads 4 | Bit-identical Every one of its 126 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 126/126 |
confounds-linalg/gelsd-full-e1 | LAPACK's least squares | DGELSD's three paths, DLALSD's two (DLASDA, DLALS0, DLALSA), DLASRT and numpy's lstsq, the cosine filter's real calls included (full comparison, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gelsd.py --threads 4 | Bit-identical Every one of its 126 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 126/126 |
confounds-linalg/gelsd-full-e1-zen | LAPACK's least squares | DGELSD's three paths, DLALSD's two (DLASDA, DLALS0, DLALSA), DLASRT and numpy's lstsq, the cosine filter's real calls included (full comparison, numpy 2.5.4 / OpenBLAS 0.3.34 at this machine's GEMM_Q 512)command linepython -P validation/investigations/confounds/check_gelsd.py --threads 4 | Bit-identical Every one of its 136 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 136/136 |
confounds-linalg/gesvd-full-e2 | LAPACK's SVD | SciPy's DGESVD fallback (DBDSQR, OpenBLAS 0.3.28's DLARFT) and nipype's fallback_svd and _full_rank against SciPy's own library (full comparison, numpy 2.3.5 / OpenBLAS 0.3.30, SciPy 1.15.3, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesvd.py --threads 4 | Bit-identical Every one of its 83 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 83/83 |
confounds-linalg/gesvd-full-e1 | LAPACK's SVD | SciPy's DGESVD fallback (DBDSQR, OpenBLAS 0.3.28's DLARFT) and nipype's fallback_svd and _full_rank against SciPy's own library (full comparison, numpy 2.5.4 / OpenBLAS 0.3.34 (OPENBLAS_L2_SIZE=512), SciPy 1.18.1, pandas 3.0.6)command linepython -P validation/investigations/confounds/check_gesvd.py --threads 4 | Bit-identical Every one of its 83 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 83/83 |
confounds-linalg/gesvd-full-e1-zen | LAPACK's SVD | SciPy's DGESVD fallback (DBDSQR, OpenBLAS 0.3.28's DLARFT) and nipype's fallback_svd and _full_rank against SciPy's own library (full comparison, numpy 2.5.4 / OpenBLAS 0.3.34 at this machine's GEMM_Q 512)command linepython -P validation/investigations/confounds/check_gesvd.py --threads 4 | Bit-identical Every one of its 95 checks: the originals’ outputs bit for bit (numbers, files, messages and exit statuses), or fMRIPrep’s own node files and confounds files of its runs. | 95/95 |
Known limits
- fMRIPrep’s confounds with FreeSurfer’s segmentation (its runs with
recon-all) are not compared yet: PIPE-79 waits for fMRIPrep’s runs with FreeSurfer. - fMRIPrep’s reports of the confounds (its reportlets) come with larmorprepx’s reports.
- The comparisons with fMRIPrep stage by stage come once every program is ported.
History
- 2026-11-12: the replica’s random-testing run and the checks of the final build.
- 2026-11-13: the replica of fMRIPrep’s confound estimation complete, with nipype’s command line and numpy’s and SciPy’s linear algebra (the roadmap).
Data credits
The data this page shows or names: each dataset with its version, DOI and licence, from LARMORX's test-data catalog. Public-domain dedications (CC0, PDDL) ask nothing; the datasets are cited all the same.
- Mixed-gambles task (OpenNeuro ds000005, version 00001), Open Data Commons Public Domain Dedication and Licence 1.0.
- Multi-echo fMRI replication sample (OpenNeuro ds000210, version 00002), Open Data Commons Public Domain Dedication and Licence 1.0.
- Children and adults watching a short film (OpenNeuro ds000228, version 1.1.1, doi:10.18112/openneuro.ds000228.v1.1.1), CC0 1.0 public-domain dedication.
- Behavioral interventions for reducing head motion in children (OpenNeuro ds000256, version 00002), Open Data Commons Public Domain Dedication and Licence 1.0.
- Multi-echo Cambridge (OpenNeuro ds000258, version 1.0.1, doi:10.18112/openneuro.ds000258.v1.0.1), CC0 1.0 public-domain dedication.
- fMRI: resting state and arithmetic task (OpenNeuro ds002422, version 1.1.0, doi:10.18112/openneuro.ds002422.v1.1.0), CC0 1.0 public-domain dedication.
- MB-EPI-comparison (OpenNeuro ds002603, version 1.0.0, doi:10.18112/openneuro.ds002603.v1.0.0), CC0 1.0 public-domain dedication.
- Neurocognitive aging data release (OpenNeuro ds003592, version 1.0.13, doi:10.18112/openneuro.ds003592.v1.0.13), CC0 1.0 public-domain dedication.
- Aphasia Recovery Cohort (OpenNeuro ds004884, version 1.0.2, doi:10.18112/openneuro.ds004884.v1.0.2), CC0 1.0 public-domain dedication.
- The FreeSurfer Maintenance Dataset (OpenNeuro ds004958, version 1.0.0, doi:10.18112/openneuro.ds004958.v1.0.0), CC0 1.0 public-domain dedication.
- Retrospective Task/Rest fMRI Data from Tumor Patients (OpenNeuro ds005003, version 2.0.0, doi:10.18112/openneuro.ds005003.v2.0.0), CC0 1.0 public-domain dedication.
- Political Moralization (OpenNeuro ds005040, version 1.2.0, doi:10.18112/openneuro.ds005040.v1.2.0), CC0 1.0 public-domain dedication.
- fMRIPrep boot camp 7T training dataset (OpenNeuro ds005454, version 1.0.0, doi:10.18112/openneuro.ds005454.v1.0.0), Creative Commons Attribution 4.0 (CC BY 4.0). Attribution: Sleight EC, Esteban O, Grouiller F. fMRIPrep boot camp -- 7T training dataset. OpenNeuro ds005454, version 1.0.0. doi:10.18112/openneuro.ds005454.v1.0.0 The images shown are adapted from it: reoriented to RAS, cropped to the region shown, masked outside it, and shown as the outputs of the programs compared.
- NIMH Healthy Research Volunteer Dataset (OpenNeuro ds005752, version 2.1.0, doi:10.18112/openneuro.ds005752.v2.1.0), CC0 1.0 public-domain dedication.
- CS_fmri (OpenNeuro ds006736, version 1.0.0, doi:10.18112/openneuro.ds006736.v1.0.0), CC0 1.0 public-domain dedication.