Active Touch and Self-Motion Encoding by Merkel Cell-Associated Afferents (DANDI:000226)

Claims

From the paper Active Touch and Self-Motion Encoding by Merkel Cell-Associated Afferents

Reproducing Severson et al. (2017) from dandiset 000226 17

Notebook reproduce_severson_2017.ipynb

Partially reproducedResults, second section; Figure 3A, E, H, Iuses data from outside the dandiset
Merkel and SA afferents responded to contact in a dramatically direction-selective manner, far more strongly to contacts in either the protraction or the retraction direction.
17 of 26 Merkel and SA afferents (65%) have a direction index of 0.5 or more (criterion 80%); 21 of 26 have 0.33 or more and 11 of 26 have 0.67 or more. Median index 0.65 over Merkel and SA, 0.59 over the 14 Merkel afferents, 0.23 over the 7 RA afferents, none of which reaches 0.5. 18 of 26 prefer protraction contacts. Contacts are classified by the sign of their mean M0; in 77% to 93% of contacts per recording that sign matches the whisker velocity at onset.
How the test differed from the paper’s method
- The paper shows direction selectivity in examples and tuning surfaces without a measure or threshold; the index and the 0.5 threshold were chosen before the analysis.
Partially reproducedResults, second section; Figure 3D; Figure S2Cnames its codes from outside the dandiset
M0 and its rate of change together explained spiking (r = 0.63 plus or minus 0.20) better than any other pair of variables, and better than a model of M0, Fax and Flat (r = 0.48 plus or minus 0.19).
M0 and M0' mean r 0.622 (paper 0.63); M0, Fax and Flat 0.475 (paper 0.48), lower than M0 and M0' in 31 of 33 (Wilcoxon p = 4.8e-08). Of the 45 pairs, M0 and omega (velocity) has the highest mean r, 0.633, and M0 and M0' is second; per afferent M0 and M0' is higher than M0 and omega in 19 of 33 (median difference 0.003). Flat and omega (0.620), Flat and M0' (0.606) and M0 and Flat' (0.605) follow. The values are within tolerance, but M0 and M0' is not the single best pair. The rank holds in two variants - with lambda 100 M0 and M0' is second (0.589 against 0.600 for M0 and omega, higher in 17 of 33), and fitted on a random half of the frames and evaluated on the other half it is second (0.620 against 0.630, higher in 20 of 33).
How the test differed from the paper’s method
- Spline models in place of mgcv GAMs; the ranking of close pairs may depend on the smoothing method.
Partially reproducedResults, fourth section; Figure 5D, Enames its codes from outside the dandiset
Across the population of WT* afferents, preferred phase spanned the whisk cycle.
Preferred phases occupy all four quarter cycles (7, 4, 1 and 3 of 15), but the Rayleigh test finds a preferred direction (R = 0.448, p = 0.047, just below 0.05). The outcome sits at the threshold and depends on the estimator of preferred phase, which the criterion fixes as the bin of highest rate of 30 - with the circular mean of the tuning curve p = 0.062 (which would be reproduced), with the highest of 20 or 60 bins, computed on the outlier-free frames kept for figs6-kinematic-gam, p = 0.023 and 0.046; all four quarters are occupied under every estimator. At the preferred phase the mean whisker velocity is negative for 11 of 15. Whether this agrees with the paper's slight abundance in retraction (Figure 7G, on 28 afferents) is not established, since the stored phase is negated and the paper's phase convention for the figure is not known.
How the test differed from the paper’s method
- The paper gives no test for spanning the whisk cycle; the quarter-cycle and Rayleigh criterion was chosen before the analysis.
ReproducedTable S1 (Supplemental Information)names its codes from outside the dandiset
The dandiset's recordings can be matched one to one to the recordings of Table S1, and the matched records agree on mouse, sex, mouse line and date.
All 60 files match a distinct Table S1 recording, 58 by recording id (the NWB identifier without "KS" and leading zeros) and 2 (KS0125_ and KS0128_, matched to 125B and 128D) by number and mouse. All 60 agree on mouse, sex, mouse line and date. Groups - 33 touch, 13 progressive cutting, 5 cut-whisker, 4 EMG, 3 excluded touch, 2 whisking-only. No Table S1 recording lacks a file. The transcription of Table S1 agrees with the table parsed from the supplement on all 60 rows and every column used.
How the test differed from the paper’s method
- The criterion's rule of removing "KS0" was applied as removing "KS" and leading zeros, so that the EMG files KS0007A to KS0011A match 7A to 11A.
ReproducedResults, first section; Quantification and Statistical Analysisnames its codes from outside the dandiset
The initial dataset of 33 touch afferents comprises 54,647,500 video frames with simultaneous single-unit recordings, and the paper analyses 65,257,500 frames (2,175 min) in total.
The 33 touch recordings hold 54,660,000 frames (paper 54,647,500, 5 trials more). The 53 analysed afferents hold 65,710,000 frames, 2,190 min (paper 65,257,500 and 2,175 min, 181 trials more). The other readings are also within 10% - 68,455,000 (53 afferents and EMG), 68,005,000 (56 afferents), 70,750,000 (all files).
ReproducedResults, first section; Quantification and Statistical Analysisuses data from outside the dandiset
Identified Merkel afferents account for 823 minutes of the initial dataset and 846 minutes (25,365,000 frames) overall, against 999 minutes for unidentified afferents in the initial dataset.
Merkel touch recordings 823.8 min (paper 823), unidentified touch recordings 998.2 min (paper 999), the 17 analysed Merkel afferents 846.4 min (paper 846). All 20 identified Merkel afferents hold 922.9 min, also within tolerance. Merkel identity is from Table S1; the minutes are from the dandiset's trials, and they agree with the paper's.
ReproducedResults, first section; Figure 2Cnames its codes from outside the dandiset
During non-whisking periods most afferents spiked at low rates (0.0 plus or minus 0.2 Hz, median and IQR, n = 33).
Median non-whisking rate 0.03 Hz, IQR 0.20 Hz, over the 33 touch afferents (paper 0.0 and 0.2). Range 0.00 to 34.39 Hz; 221B fires at 34.4 Hz without whisking.
ReproducedResults, first section; Figure 2Cnames its codes from outside the dandiset
All afferents responded with increased spike rates during whisker-pole contact (48.0 plus or minus 96.3 Hz, median and IQR, n = 33).
Median contact rate 47.9 Hz, IQR 94.8 Hz (paper 48.0 and 96.3), range 3.6 to 239.2 Hz. All 33 afferents are touch-sensitive by the Glossary rule, with bootstrap CIs over trials and with Poisson CIs. The median is also within 10% of the paper's value, so the outcome does not rest on the wide tolerance.
How the test differed from the paper’s method
- The Methods do not say what the bootstrap resamples; trials were resampled (1,000 times).
ReproducedResults, first section; Figure 2C; Glossary; Table S1names its codes from outside the dandiset
A subset of 13 of the 33 touch afferents (WT*) spiked especially strongly during whisking in air (19.1 plus or minus 58.2 Hz, median and IQR), and the other afferents divide into those sensitive to whisking and touch (WT) and to touch only (T).
Under reading (b), spikes during whisking in air, 13 afferents are WT*, the same set as Table S1's (5 Merkel, 7 other SA, 1 RA), with a median whisking-in-air rate of 19.0 Hz and IQR 57.4 Hz (paper 19.1 and 58.2). Classes agree with Table S1 for 31 of 33; 223B and 243B are WT in Table S1 and T here (whisking-in-air rates 0.23 and 0.38 Hz, CIs overlapping the non-whisking rate). Under reading (a), all spikes, 15 are WT* (median 7.1 Hz) and 29 of 33 agree; 227A and 145B (1.05 and 1.26 Hz) have more than 1,000 spikes in all but 393 and 233 during whisking in air. With Poisson CIs, 15 are WT* and 29 of 33 agree.
How the test differed from the paper’s method
- The Glossary's "at least 1,000 spikes" can be read as spikes in the recording (a) or during whisking in air (b); only (b) matches the paper's count and Table S1's classes.
ReproducedResults, second section; Figure 3Dnames its codes from outside the dandiset
A full statistical model of ten mechanical and kinematic variables strongly predicted ongoing spike rate during touch (r = 0.71 plus or minus 0.16, mean and SD, n = 33).
Mean r 0.708, SD 0.141 (paper 0.71 and 0.16), median 0.726, range 0.363 to 0.919, 33 afferents. With lambda 100 the mean is 0.687. All 1,617 spline fits converged. Fits used a median of 60,000 frames per recording (at least 47,342) of a median 288,418 clean contact frames.
How the test differed from the paper’s method
- Penalized cubic B-splines (10 per variable, lambda 1) fitted as a binomial model with two trials per 2 ms frame replace mgcv GAMs on 1 ms bins with UBRE smoothing.
- At most 60,000 randomly chosen contact frames per recording.
ReproducedResults, second section; Figure 3Dnames its codes from outside the dandiset
Models fitted to moment or its rate of change alone explained spiking relatively poorly (r = 0.45 plus or minus 0.19 and 0.42 plus or minus 0.21).
M0 alone mean r 0.441 (SD 0.178; paper 0.45), M0' alone 0.415 (SD 0.217; paper 0.42). With lambda 100, 0.415 and 0.398.
How the test differed from the paper’s method
- Spline models in place of mgcv GAMs, as for fig3d-gam-full.
ReproducedResults, third section; Figure 4C; STAR Methods, Viscoelastic modeluses data from outside the dandiset
A simple viscoelastic model of moment and its rate of change predicted spike rates of Merkel and SA afferents during touch at levels comparable to the statistical model based on M0 and its rate of change (25 of 26 afferents fitted).
Over 25 Merkel and SA afferents (Table S1's fitted set), mean r 0.656 for the viscoelastic model against 0.686 for the spline model of M0 and M0' on the same frames, difference -0.030 (95% bootstrap CI -0.061 to -0.006); viscoelastic higher in 10 of 25; Merkel afferents alone -0.034. The optimizer reported success for all 25, but 6 of 25 fits end with a parameter at a bound (eps_lim in 4, E in 1, eta in 2), so some fits may be local optima, which would only lower the viscoelastic r. These values differ slightly from the previous run (-0.029, 9 of 25, 7 at a bound) with the same code, data and locked packages; the cause was not established. With three other sets of random starts the mean difference is -0.037, -0.040 and -0.017, so the outcome holds, but single recordings change by up to 0.555 in r. Different parameter sets give nearly the same prediction, so the fitted parameter values are not interpreted.
How the test differed from the paper’s method
- Fitted with L-BFGS-B from 8 starts, with the rate of change of moment scaled by its 80th percentile inside the fit; the Methods used fmincon. During development an unscaled version found worse optima for the same model, so the result depends on the optimizer.
- Fitted and evaluated on the same at most 60,000 contact frames as the spline models, which also exclude frames that are outliers on the other nine variables.
ReproducedResults, fourth section; Figure 5A to C; Figure S3Dnames its codes from outside the dandiset
All 15 WT* afferents (13 touch WT* and 2 whisking-only) were dramatically modulated by whisk phase, with rates at non-preferred and preferred phases of 2.1 plus or minus 3.4 and 51.2 plus or minus 105.7 Hz (median and IQR).
Over the 15 WT* afferents, non-preferred phase median 2.43 Hz, IQR 3.16 (paper 2.1 and 3.4); preferred phase median 51.1 Hz, IQR 96.6 (paper 51.2 and 105.7). All 15 fire at least 9.5 times more at the preferred than the non-preferred phase. Both medians are within the claim's tolerances (the paper's IQRs). Against a 10% band, which is not the criterion, the non-preferred median (16% above the paper's) would be outside and the preferred inside.
ReproducedResults, fourth section; Figure S3names its codes from outside the dandiset
The preferred phase of each WT* afferent was largely invariant across whisk cycles of different amplitudes, frequencies and setpoints.
Median absolute circular difference between the preferred phases of the two halves 0.37 rad (amplitude), 0.59 (frequency), 0.48 (setpoint), against pi/4 = 0.785. 4, 6 and 5 of 15 afferents differ by more than pi/4.
How the test differed from the paper’s method
- The paper shows invariance with tuning surfaces only; the median split and the pi/4 threshold were chosen before the analysis.
ReproducedResults, fifth section; Figure S6A, Bnames its codes from outside the dandiset
Statistical models fitted to whisker position, velocity, acceleration and jerk during whisking in air predicted phase tuning curves nearly perfectly (r = 0.97 plus or minus 0.04, n = 28), while single kinematic variables performed poorly.
On the 15 WT* afferents, the full model's mean r between predicted and observed phase tuning curves is 0.975, SD 0.013 (paper 0.97 and 0.04, on 28). Single-variable models give 0.524 (jerk) to 0.724 (acceleration); velocity and acceleration together give 0.960. No fit failed to converge.
How the test differed from the paper’s method
- Tested on 15 of the paper's 28 afferents; the intact condition of the 13 progressive cutting recordings cannot be separated.
- Spline models in place of mgcv GAMs, fitted on 60,000 frames and evaluated on up to 400,000 whisking-in-air frames per afferent.
ReproducedFigure S4F to H; Figure 7Anames its codes from outside the dandiset
Motor unit activity of m. nasolabialis was phase modulated and correlated with negative whisker velocity (r = 0.93) and anti-correlated with positive velocity (r = -0.75) within whisk cycles.
The spike times of 11A, taken as motor unit spikes, give r = 0.924 with negative velocity (paper 0.93) and -0.716 with positive velocity (paper -0.75) across 30 phase bins of 23.3 min of whisking; the rate ranges 0.7 to 37.6 Hz and peaks where velocity is negative. The raw EMG is not in the file.
How the test differed from the paper’s method
- The file does not say that its spike times are motor units of m. nasolabialis; this is taken from Table S1.
ReproducedSTAR Methods, Video analysis, whisker baseline curvature; Figure S7names its codes from outside the dandiset
The noise level of curvature estimates, curvature change during whisking in air as a percentage of that during contact, was 7.8 plus or minus 5.0% (mean and SD) across 33 sessions.
Mean 10.01%, SD 6.97, range 4.23 to 39.89 over the 33 touch recordings (paper 7.8 and 5.0). The mean is within the paper's stated spread but 28% above its mean value, outside a 10% band; 243B (39.89%) and 265A (28.49%) are high. Both readings of contact frames give the same values.
How the test differed from the paper’s method
- The Methods give no rule for frames with a missing distance to the pole; these are treated as far from the pole.

Not tested in a notebook 11

4 metadata insufficient, 7 data missing from the dandiset

BlockedDiscussion, fourth paragraph; Figure 7G· Metadata insufficientnames its codes from outside the dandiset
Preferred phases of all 28 WT* afferents spanned the whisk cycle, with a slight abundance in the retraction phase.
Needs the intact condition of the 13 progressive cutting recordings, whose trials are all labelled intact (see fig6-cut-eliminated). The 15 other WT* afferents are tested in fig5e-phase-span.
BlockedResults, fifth section; Figure 6B, E· Metadata insufficientnames its codes from outside the dandiset
Progressive whisker cutting eliminated spiking during whisking in air in 6 of 13 afferents (from 50.1 plus or minus 42.3 Hz intact to 0.9 plus or minus 0.7 Hz after the final cut, mean and SD).
Every trial of the 13 progressive cutting recordings has cuttingGroup "intact" and each file has one MOI value, so the intact, sham and cut conditions cannot be told apart. The whisking-in-air rate falls over most sessions (last fifth below half the first in 10 of 13, below 10% in 5 of 13), so the files appear to hold the cut conditions under the intact label. Splitting by trial order would be a guess.
BlockedResults, fifth section; Figure 6C to E· Metadata insufficientnames its codes from outside the dandiset
In 7 of 13 afferents self-motion responses remained after cutting (86.1 plus or minus 71.5 Hz intact against 70.3 plus or minus 58.5 Hz after the final cut).
As fig6-cut-eliminated.
BlockedResults, fifth section; Figure 6A· Data missing from the dandiset
Cutting shortened the whisker to near its base (final length 0.37 plus or minus 0.43 mm, mean and SD, n = 13).
The remaining whisker length after each cut is not in the dandiset; the trial table's MOI has one value per file, the same for all trials.
BlockedResults, sixth section; Figure 7D· Data missing from the dandisetnames its codes from outside the dandiset
After whisker trimming, phase tuning curves of the 7 afferents that kept spiking were correlated with intrinsic protractor or nasolabialis EMG (r = 0.66 plus or minus 0.27).
Needs the intrinsic protractor EMG, which is not in the EMG files 7A, 8A and 9A (they hold tracked whisker series only, with no spike times and nothing in acquisition), and the cut condition of each trial (metadata insufficient, see fig6-cut-eliminated). The EMG is given as the main reason because no labelling of the trials would supply it.
BlockedResults, sixth section; Figure 7E· Metadata insufficientnames its codes from outside the dandiset
The spike rate lost after cutting was correlated with positive acceleration, negative acceleration or jerk (r = 0.71 plus or minus 0.19, n = 13).
Needs the intact and fully cut conditions of the 13 progressive cutting recordings, which the trial tables do not distinguish (see fig6-cut-eliminated).
BlockedResults, sixth section; Figure 7F· Data missing from the dandisetnames its codes from outside the dandiset
In 5 afferents recorded with the whisker already cut near its base, phase tuning was strongly correlated with intrinsic protractor or nasolabialis EMG (r = 0.76 plus or minus 0.06).
The 5 cut-whisker recordings and the nasolabialis motor units (11A) are in the dandiset, but the best match is chosen between the intrinsic protractor and the nasolabialis curves, and the intrinsic protractor EMG is not in the dandiset. The nasolabialis curve alone gives no useful bound - 4 of the 5 afferents are anticorrelated with it (mean r -0.277, SD 0.559), and only 305A correlates (0.695). As a stand-in that is not the criterion, the better of each afferent's correlations with its own positive and negative velocity curves has mean 0.753 (0.680 to 0.848), with positive velocity, which the paper relates to the intrinsic protractor, the better for 4 of 5.
BlockedFigure S4B to D; Figure 7A· Data missing from the dandisetnames its codes from outside the dandiset
Intrinsic protractor EMG was correlated with positive whisker velocity within whisk cycles (r = 0.94, 0.97 and 0.96 in three mice).
Files 7A, 8A and 9A hold the tracked whisker series of the EMG sessions (position, velocity, Mi, dMi and others) but no EMG signal, no spike times and nothing in acquisition, analysis or stimulus. The archive's variableMeasured for them lists only BehavioralTimeSeries and ProcessingModule, and their size per trial (0.31 MB) equals that of 11A and of the progressive cutting files. The undescribed Mi (N*m) and dMi (N*m/s) do not follow positive velocity by phase - r -0.026 to 0.002 for Mi and -0.570 to -0.534 for dMi in 7A, 8A and 9A, close to the medians in every group of afferent files (-0.026 to -0.014 and -0.536 to -0.485) - so neither is the intrinsic protractor EMG.
BlockedSTAR Methods, Optogenetic identification; Figure S1C, D· Data missing from the dandiset
Identified Merkel afferents spiked at short latency (4.9 plus or minus 1.3 ms) and low jitter (SD of first-spike latency 0.91 plus or minus 0.29 ms) after brief light pulses.
The light pulse times and the optogenetic stimulation trials are not in the dandiset; no file has a member in stimulus, acquisition or analysis, or an optogenetic neurodata type.
BlockedSTAR Methods, Histology quantification; Figure S1B· Data missing from the dandiset
In TrkC-CreER;Rosa-Ai9 mice, 99.2% of classified labelled afferent endings in whisker follicles were Merkel cell-associated (1,045 of 1,053).
The histology (confocal stacks and their scoring) is not in the dandiset, and was not found in another public archive.
BlockedSTAR Methods, Video analysis, contact detection· Data missing from the dandiset
Automatic contact classification agreed with manual curation for 99.5 plus or minus 0.2% of frames, with a false positive rate of 0.3 plus or minus 0.1%.
The files hold the final contact classification only, not the manually curated frames the classifier was tested against.

Findings about the data

What attempts found about the dandiset itself, for its owner and the archive.

Description missingThe files do not say which experiment each recording belongs to (touch, progressive cutting, cut-whisker, whisking-only, EMG), the afferent type (RA or SA), which afferents were identified as Merkel afferents by optogenetic tagging, or the whisking and touch class. All of these are in the paper's Table S1, whose recording ids match the NWB identifiers.
Selection not documentedThe paper excludes three touch recordings (227C, 214A, 271A) from all analyses except Figure S1. The dandiset holds them without saying so.
Metadata errorIn the 13 progressive cutting recordings every trial has cuttingGroup "intact" and the trial table's MOI has one value per file, although the whisking-in-air rate falls over most of these sessions (last fifth below half the first in 10 of 13). The cutting conditions (intact, sham, cut 1 to 3) and the moment of inertia after each cut are missing, which blocks Figure 6 and parts of Figure 7.
Data missingThe EMG files of the intrinsic protractor recordings (7A, 8A, 9A) hold the tracked whisker series but no EMG signal; the nasolabialis file (11A) holds spike times but not the EMG. The optogenetic tagging data, the histology, the whisker geometry and the manually curated contact frames are not in the dandiset.
Metadata errorVelocity, acceleration and jerk are stated in degree/s, degree/s^2 and degree/s^3, and dM0, dFaxial and dFlateral in N*m/s and N/s, but the values are per millisecond; each is close to half the change of its variable per 2 ms frame (median ratios 0.467 to 0.502 over files).
Description missingThe stored whisk phase is the negative of the Hilbert phase of the band-passed whisker position; its sine is anticorrelated (median -0.9999 over files) and its cosine correlated (median 0.9999) with those of the Hilbert phase, and position peaks at phase 0 in all 60 files. So phases from -pi to 0 are retraction, the reverse of the convention the paper cites. Neither the files nor the paper describe the convention.
Description missingEvery series has the description "no description". distanceToPoleCenter has unit "unknown"; its median in frames with contact = 1 is between -0.08 and 0.09 in every file but one (209A, a cut-whisker recording with almost no contact), which fits the Methods' distance from whisker to pole in mm, not a distance to the pole's centre. Series not used here (Mi, dMi, Fnorm, thetaAtContact, follicleCoordsX and Y) are not explained; Mi and dMi have units N*m and N*m/s and values below 3e-8 in absolute size, in every file including the EMG files.
Metadata errorTrial stop_time is 6.998 s after start_time in every trial of every file, while each trial's series hold 2,500 frames at 500 Hz (5 s), as the Methods describe.
File format errorThe spike times of file sub-KSm31 (KS0277A) are not sorted.
Metadata errorTwo identifiers, KS0125_ and KS0128_, lack the unit letter of their Table S1 recordings (125B and 128D). All session start times are 12:00, so only the dates carry information.
OtherSome series hold extreme values, likely tracking errors, such as whisker position from -985 to 596 degrees, whisk frequency from -250 to 250 Hz and M0 down to -0.52 N m. The Glossary's percentile outlier removal excludes them, but they are not flagged in the files.
Description missingTwo recordings made with the whisker cut before recording (199B and 209A) hold distance-to-pole series and frames with contact = 1 (5,014 and 32); in 209A the distance at those frames is 5.6 to 6.5. The files do not say whether a pole was presented in these sessions or what these frames are.
Methods not availableThe paper's MATLAB and R analysis code is available only on request; no public repository was found, so the GAM and viscoelastic fits had to be reimplemented.

Notebooks

Every registered entry that uses this dandiset. An entry can hold several notebooks, and can use more than one dandiset.

Reproduction Merkel and other whisker afferents during active touch and whisking (Severson et al. 2017)

Tests the findings of Severson, Xu, Van de Loo, Bai, Ginty and O'Connor (2017) on dandiset 000226, which holds 56 trigeminal afferent recordings with high-speed whisker video and 4 EMG sessions. It checks the recordings against the paper's Table S1, the spike rates during non-whisking, whisking in air and touch, the classes of afferents, direction selectivity, the statistical and viscoelastic models of touch responses, phase tuning during whisking in air, the nasolabialis motor units and the curvature noise. The cutting conditions of the progressive cutting experiment, the intrinsic protractor EMG and the optogenetic tagging data are not in the dandiset. All files

Reproducing Severson et al. (2017) from dandiset 000226
reproduce_severson_2017.ipynbExecuted, with outputs · Source

Verification

Round 1
2026-10-17Agreed with 28 of 28 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: Round 1. The first executed notebook with the conclusion cells and the summary removed, the helper module and the claims file, with the paper (PMC5528144). The verifier queried the archive metadata and tested helper functions on synthetic inputs. Its report gave its own outcome for each claim; all 28 agree with the record. It listed ten problems (B1 to B10), three marked as affecting a conclusion, and outputs it could not check. B7 (contact requires whisking, as the claims file fixes) needed no change. The cell numbers in the reasons below refer to the round-1 notebook.
Round 2
2026-10-17Agreed with 28 of 28 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: Round 2. The full executed notebook after the round-1 revision, the helper module, the claims file and the attempt record, with the paper (PMC5528144). The verifier checked every conclusion cell, the README and the record against the printed outputs and reviewed the round-1 responses. It agreed with all 28 outcomes, found seven statements not fully supported by the outputs, and said the Table S1 transcription had still not been checked by a verifier. The cell numbers in its reasons refer to the round-2 notebook; this revision adds code cells at positions 19, 47 and 64, so later cells move. Outside the claims, the velocity column of the data check (cell 11) was labelled degrees/s although the notebook's own ratio check shows per millisecond; the label now says so.

Method differences

How the tests differed from the paper’s method or data, and the claims each applies to.

The criterion's rule of removing "KS0" was applied as removing "KS" and leading zeros, so that the EMG files KS0007A to KS0011A match 7A to 11A.table-s1-link
The Methods do not say what the bootstrap resamples; trials were resampled (1,000 times).fig2c-touch-rate
The Glossary's "at least 1,000 spikes" can be read as spikes in the recording (a) or during whisking in air (b); only (b) matches the paper's count and Table S1's classes.fig2c-wt-star
The paper shows direction selectivity in examples and tuning surfaces without a measure or threshold; the index and the 0.5 threshold were chosen before the analysis.fig3a-direction-selective
Penalized cubic B-splines (10 per variable, lambda 1) fitted as a binomial model with two trials per 2 ms frame replace mgcv GAMs on 1 ms bins with UBRE smoothing.fig3d-gam-full
At most 60,000 randomly chosen contact frames per recording.fig3d-gam-full
Spline models in place of mgcv GAMs, as for fig3d-gam-full.fig3d-gam-single
Spline models in place of mgcv GAMs; the ranking of close pairs may depend on the smoothing method.fig3d-gam-best-pair
Fitted with L-BFGS-B from 8 starts, with the rate of change of moment scaled by its 80th percentile inside the fit; the Methods used fmincon. During development an unscaled version found worse optima for the same model, so the result depends on the optimizer.fig4c-viscoelastic
Fitted and evaluated on the same at most 60,000 contact frames as the spline models, which also exclude frames that are outliers on the other nine variables.fig4c-viscoelastic
The paper gives no test for spanning the whisk cycle; the quarter-cycle and Rayleigh criterion was chosen before the analysis.fig5e-phase-span
The paper shows invariance with tuning surfaces only; the median split and the pi/4 threshold were chosen before the analysis.figs3-phase-invariance
Tested on 15 of the paper's 28 afferents; the intact condition of the 13 progressive cutting recordings cannot be separated.figs6-kinematic-gam
Spline models in place of mgcv GAMs, fitted on 60,000 frames and evaluated on up to 400,000 whisking-in-air frames per afferent.figs6-kinematic-gam
The file does not say that its spike times are motor units of m. nasolabialis; this is taken from Table S1.figs4-nasolabialis
The Methods give no rule for frames with a missing distance to the pole; these are treated as far from the pole.sigma-kappa

Provenance

Attempt record: 000226-severson-2017-01
Paper: Active Touch and Self-Motion Encoding by Merkel Cell-Associated Afferents
Date: 2026-10-17
Made by: An agent, run by CatalystNeuro
Dandiset version: 0.230607.1747
Harness: Claude Code (GitHub Actions)
Model: claude-opus-5-5
Protocol: pilot-7

Transcripts of the agent runs

Each run’s prompt, messages and tool calls, with long output cut and the text of the paper left out, and its final report.

← Dandiset 000218 Dandiset 000232 →