Bayesian computation through cortical latent dynamics (DANDI:000130)

Claims

From the paper Bayesian computation through cortical latent dynamics

Reproducing Sohn et al. (2019) from dandiset 000130 20

Notebook reproduce_sohn_2019.ipynb

Not reproducedResults, Task and behavior; Figure 1E and S1; Table S1
Regression slopes of tp on ts are positive and less than one in every condition, so the animals estimated ts and their responses were biased toward the mean of the prior.
Six of 8 weighted slopes of tp on ts lie between 0 and 1 (0.50 to 0.90) and 5 of 8 have 95% intervals excluding 0 and 1. The Short prior slopes with the eye are 1.11 (Eye Right, interval 1.01 to 1.21) and 1.01 (Eye Left). All 8 are positive, so the animal estimated ts; the part of the statement about bias toward the mean fails in the two eye Short conditions.
How the test differed from the paper’s method
- One session of monkey H; the paper's slopes come from 17 sessions of H and 12 of G.
- Outliers removed with the file's `is_outlier` flag, which a refit of the Methods' mixture model does not reproduce (see findings).
Not reproducedResults, Task and behavior; Figure 1E; Table S2
At the overlapping ts of 800 ms, biases were in opposite directions for the two priors (rank-sum test, p < 1e-43 in animal H over 17 sessions).
At ts = 800 ms the median bias is +28.3 ms for Short (93 trials) and +47.6 ms for Long (107 trials), both positive; rank-sum p = 0.073. The Long tp is the longer in 2 of 4 datasets (Hand Right p = 0.00099, Hand Left p = 0.069, where the Short bias is -61.6 ms); in the eye datasets the Short tp is the longer (Eye Left +79.1 against +18.6 ms, p = 0.051).
How the test differed from the paper’s method
- One session; the paper's p < 1e-43 pools 17 sessions of monkey H.
Not reproducedResults, Single-neuron response profiles; Figure 3E; Methods, Generalized linear model
About 30% of neurons were modulated by elapsed time (27% in monkey H) and more than 60% were sensitive to the prior (65% in monkey H), by a Poisson GLM of spike counts before Set compared by BIC.
36 of 48 units are prior-dependent (75.0%, 95% interval 61.2 to 85.1%) and 22 of 48 are ts-dependent (45.8%, interval 32.6 to 59.7%), against 65% and 27% for monkey H. Both are outside tolerance and higher; the interval of the prior fraction includes 65%, that of the ts fraction excludes 27%. A likelihood-ratio test gives higher fractions still (prior 77.1 to 87.5%, ts 66.7 to 77.1% at p < 0.01 and 0.05). Chance levels by the same BIC rule are far lower; prior labels shuffled within effector and direction give 0 to 8.3% and labels shifted along the session by 50 to 200 trials, which keeps the prior blocks, 2.1 to 14.6%; time points shuffled within trials give 0 to 2.1% ts-dependent. The ts test in the Short prior alone gives 31.2% and in the Long prior alone 33.3%, both above 27 +/- 2.7%.
How the test differed from the paper’s method
- 48 units from one session, where the paper's fractions use 536 units of monkey H over 17 sessions.
- The GLM design with attrition (rows for every trial whose Set comes after each time point) is read from the Methods; the paper's code was not available.
- The Long prior's ts test uses the Long ts as time points (800 to 1200 ms after Ready); Eq. 3 of the paper does not say which time points were used, and a unit counts as ts-dependent when either prior's test finds it.
Not reproducedResults, Geometry and dynamics of population activity; Figure S3
Population activity is low dimensional; 3 to 4 principal components in the estimation epoch and 5 to 10 in the production epoch explain about 75% of the variance.
Two components reach 75% of the variance in the estimation epoch in all 4 datasets (3 explain 86 to 90%), and 2, 2, 5 and 4 in the production epoch; 1 of 8 values is in range. The count depends on normalisation, which the Methods do not mention; with rates z-scored per unit (a variant) the counts are 3 or 4 and 6 to 10, all 8 in range, and with soft normalisation (range plus 5 spikes/s) 6 of 8 are in range. The outcome is judged on unnormalised rates, as the claims file fixed.
How the test differed from the paper’s method
- 48 simultaneously recorded units, where the paper's trial-averaged PCA pools units over 17 sessions; more units can add dimensions.
- Rates not normalised before PCA; the Methods do not say.
Not reproducedResults, Alternative mechanisms (transient hypothesis); Figure 5D
In the first 200 ms after Set, trajectories for different ts run parallel; their distance to the middle ts trajectory does not change (signed-rank test for zero slopes, p = 0.92 for the shortest ts, p = 0.13 for the longest).
For the longest ts the distance to the middle trajectory falls in the first 200 ms after Set in 6 of 8 dataset-priors (median -15.3% of the mean distance per 100 ms), signed-rank p = 0.039; for the shortest ts p = 0.078 (6 of 8 rising, median +26.7%). The PCA kept 1 to 4 components, one in Eye Left. For the longest ts p = 0.016 with 60% of the variance, 0.039 with 90%, 0.023 in the full space and 0.039 with z-scored rates; with soft-normalised rates 0.055.
How the test differed from the paper’s method
- 8 dataset-priors of one animal, where the paper's test used 16.
- Rates smoothed over 100 ms before to 300 ms after Set and cut to the first 200 ms, a choice made here.
Not reproducedResults, Alternative mechanisms (threshold hypothesis); Figure 5F
Aligned to the response, distances between trajectories for different ts decrease steadily up to movement initiation, without the converging then diverging pattern a ts-dependent threshold would produce.
With the components reaching 60% of the variance (1 to 3), the median time of the minimum distance is 210 ms before Go; 6 of 32 curves have their minimum in the last 40 ms and 14 in the last 200 ms. Distances do fall overall (the distance at Go is below that 800 ms earlier in 26 of 32). With 75% of the variance the median minimum is 80 ms before Go and in the full space 40 ms, which would meet the criterion, so the outcome depends on the number of components. With one component (eye datasets, 16 curves) the signed difference changes sign at the minimum in 12, so those minima are crossings; 4 have the minimum at Go (3 of them not crossings), and the distance at Go is more than 10 times the minimum in 5. With at least 2 components the median minimum is 230 ms before Go, with at least 3 it is 70 ms before Go.
How the test differed from the paper’s method
- The paper's test (one-tailed signed-rank, p = 1.5e-4) cannot be read unambiguously; the criterion on the median time of the minimum was fixed before the analysis.
- The 60% of variance follows the figure legend; the 75% of the Methods is shown as a variant.
Not reproducedResults, The curved manifold hypothesis captures the variance; Figure 6C
The variance across single trials of the GPFA neural state at Set projected on the encoding axis has an inverted-U profile over ts, lower at the extremes of each prior (one-tailed signed-rank test on quadratic coefficients, p = 0.0273, monkey H).
The quadratic coefficient is negative in 5 of 8 dataset-priors; one-tailed signed-rank p = 0.125 (paper 0.0273). Cross-validation chose 8 latent dimensions in all 4 datasets, the larger candidate. 12 latents, scored on the same folds as a variant, beat 8 in all 4 datasets; with them the coefficient is negative in 5 of 8, p = 0.125. After 100 EM iterations the training log-likelihood still rose by 0.064 to 0.085% over the last 25 iterations.
How the test differed from the paper’s method
- Latent dimensionality chosen only from 4 and 8, with 100 EM iterations, to keep the run short; the paper's cross-validation range is not stated.
- One GPFA per effector and direction with both priors; the paper does not say how trials were grouped.
Partially reproducedResults, Task and behavior; Figure 2C and 2D; Table S3
A Bayesian observer model with scalar measurement and production noise fit to the behaviour captures the bias and the variance of tp in all conditions.
The predicted SD of tp correlates positively with the observed one in all 4 datasets (r = 0.28 to 0.91), but the bias R2 reaches 0.8 in 2 of 4 (0.12, 0.67, 0.87, 0.98). The poorly fit datasets are the eye ones, whose Short slopes exceed 1, with wm = 0.013 and 0.025. All fits converged. Table S3 was not available, so the fitted wm (0.013 to 0.117) and wp (0.090 to 0.106) are not compared.
How the test differed from the paper’s method
- Fit to one session per effector and direction; the paper pools sessions.
- The prior is taken as continuous uniform over each range, as the Results describe; the task used 5 discrete values.
- The bias R2 threshold of 0.8 is a choice made here; the paper shows the fit graphically.
Partially reproducedResults, Encoding Bayesian estimate along curved trajectory; Figure 4D top
Neural projections are better explained by the Bayesian model than by a linear function of ts with the same number of parameters, for both priors (signed-rank test on RMSE, p = 0.008 for each prior).
The Bayesian RMSE is lower in 4 of 4 datasets for Long and 3 of 4 for Short (Hand Right Short 0.99 against 0.80). Where the Bayesian model wins it is better by 0.045 to 0.573 spikes/s; in Hand Right Short it is worse by 0.186. With the RMSE weighted by trial counts, as in the fit, the counts are the same. In the matched control period of the Long prior (short-in-long-control) the Bayesian model is also better in 3 of 4 datasets, so in this session the advantage is not specific to the support of the prior.
How the test differed from the paper’s method
- 4 datasets, where the paper had 8; a signed-rank test over 4 cannot reach p < 0.05, so the criterion counts datasets.
Partially reproducedResults, Controlling speed based on Bayesian estimate; Figure 4E
Larger projections on the encoding axis at Set are followed by slower neural trajectories in the production epoch (Pearson correlation, rho Short = -0.74, rho Long = -0.51).
Both correlations are negative and significant (Short p = 6.2e-10, Long p = 3.7e-07) and stronger than reported, outside tolerance. Within datasets -0.75 to -1.00; with Xu z-scored per dataset -0.759 and -0.779, and the speed divided by mean tp gives -0.917 and -0.923. Both quantities change with ts; with the residuals of each on ts within dataset and prior the correlation is -0.195 (Short) and +0.069 (Long), so the pooled relation is carried by the common dependence on ts. The paper's control (z-scoring within ts) is on single-trial GPFA states of the production epoch, not computed here.
How the test differed from the paper’s method
- 20 points per prior (4 datasets), where the paper had 40.
- Speed is the mean step between 20 ms states up to the shortest tp; the Methods' division by the Set to Go duration is shown as a variant.
Partially reproducedResults, Controlling speed based on Bayesian estimate; Figure 4G
The speed of neural trajectories in the production epoch predicts the mean tp (Pearson correlation, rho Short = -0.58, rho Long = -0.52).
Faster trajectories go with shorter mean tp in both priors (Short p = 0.00024, Long p = 0.00017), more strongly than reported and outside tolerance. Within datasets -0.80 to -0.99. The Neural Latents Benchmark's trial-by-trial variant gives r = -0.195 to -0.740 (mean -0.519), 7 of 8 conditions with p < 0.05. With speed and tp z-scored within each ts trial by trial, as in the paper's control, r = -0.317 (Short, 481 trials, p = 1.1e-12) and -0.201 (Long, 502 trials, p = 5.5e-06), negative in 8 of 8 dataset-priors and p < 0.05 in 6. A Poisson null at constant rates with each trial's own duration gives mean null r of -0.071 (Short) and -0.033 (Long) over 20 simulations, and the observed values are below every null value, so the shorter averaging of short trials accounts for a small part. Residuals on ts of the 20 condition means give +0.084 and -0.283.
How the test differed from the paper’s method
- 20 points per prior, where the paper had 40.
ReproducedResults, Task and behavior; Figure 2E and 2F
Increments of tp between neighbouring ts are smaller at the extremes of the prior than near its mean, as the sigmoidal Bayesian estimator predicts (signed-rank test, p < 1e-11 in animal H).
The middle increment of mean tp exceeds the extreme one in 7 of 8 conditions (median difference 31.3 ms), two-sided signed-rank p = 0.0156. Hand Left, Short is the exception (-10.7 ms).
How the test differed from the paper’s method
- The test is over the 8 conditions of one session; the paper's is over sessions and conditions of both animals.
ReproducedResults, Task and behavior; Figure 2G and 2H
Biases are larger for the Long prior than for the Short prior, as scalar variability predicts; regression slopes are smaller for Long.
The Long slope is below the Short slope in all 4 datasets (mean 0.716 against 0.898), weighted or unweighted.
How the test differed from the paper’s method
- One session.
ReproducedResults, Single-neuron response profiles; Figure 3C and 3D
At the overlapping ts of 800 ms, firing rates differ between priors, with the largest difference during the support of the prior, between 480 and 800 ms after Ready.
The mean RMS rate difference at ts = 800 ms is largest at 480 to 640 ms after Ready (6.86 spikes/s), above the 95th percentile of the shuffled maximum (3.11). Before Ready it is at the shuffle level. Per dataset the largest bin is 480-640 (2 datasets), 640-800 and 800-960 ms, and the mean is also high just after Set (6.07 at 800 to 960 ms). Over 200 bootstrap samples of trials the peak is in 480 to 640 ms in 192 and in 800 to 960 ms in 8 (96% within 480 to 800 ms).
How the test differed from the paper’s method
- The time range (320 ms before Ready to 640 ms after Set) and the shuffle reference are chosen here.
ReproducedResults, Encoding Bayesian estimate along curved trajectory; Figure 4C
Projections of estimation-epoch neural states onto the encoding axis match the Bayesian estimates from the model fitted to the behaviour (R2 = 0.993 for the Short prior, 0.996 for the Long prior, monkey H, Hand Left).
Both readings match. (a) Hand Left R2 = 0.998 (Short) and 0.996 (Long), given as measured; (b) the mean of 4 datasets 0.959 and 0.989. With only 2 PCs in the estimation epoch, random unit axes also fit te well (median R2 0.73 to 0.98); the R2 on u exceeds 95% of random axes in 3 of 8 dataset-priors (Hand Right Long and both Hand Left), so the paper's statement that the match is specific to u (Figure S5) holds only for those. With normalised rates (3 or 4 PCs) R2 stays above 0.94 in every dataset-prior, and u beats 95% of random axes in 2 of 8 (soft normalisation) and 0 of 8 (z-scored).
How the test differed from the paper’s method
- te comes from the observer model fitted to this session.
- The encoding axis averages difference vectors up to the middle of the support, a choice made here.
ReproducedResults, Encoding Bayesian estimate along curved trajectory; Figure 4D top
In the Long prior during the period matched to the support of the Short prior ("Short In Long"), there is no sigmoidal representation; the Bayesian model is not better than the linear one (signed-rank test, p = 0.74).
Judged on the paper's test as the claims file fixes; signed-rank p = 0.25 over 4 datasets, where the smallest attainable p is 0.125, so the test could not have found an effect. The effect sizes do not show the control's contrast; the Bayesian model is better in 3 of 4 datasets with a median relative RMSE difference of -0.089, at least as large as in the Short prior (-0.060). This outcome carries no evidence for the absence the paper reports.
How the test differed from the paper’s method
- 4 datasets, where the paper had 8.
ReproducedResults, Encoding Bayesian estimate along curved trajectory; Figure 4D bottom
Neural projections show more bias for the Long prior than for the Short prior, as a smaller regression slope of Xu on ts (signed-rank test, p = 0.008).
The Long slope of Xu on ts is below the Short slope in 4 of 4 datasets (ratios 0.55, 0.63, 0.34, 0.97).
How the test differed from the paper’s method
- 4 datasets, where the paper had 8.
ReproducedResults, Controlling speed based on Bayesian estimate; Figure 4F
The initial speed of production trajectories (Set to Set + 400 ms) is biased toward the middle of the prior; its dependence on ts is shallower than the unbiased profile proportional to 1/ts (signed-rank test, p < 1e-3 over conditions and animals).
The initial speed falls with ts in 8 of 8 dataset-priors and is shallower than the 1/ts profile in 7 of 8 (Eye Right Long is the exception), two-sided signed-rank p = 0.039. No trial had tp below 400 ms.
How the test differed from the paper’s method
- 8 dataset-priors of one animal, where the paper's test combines 16 over two animals.
ReproducedResults, Alternative mechanisms (speed hypothesis); Figure 5B
Instantaneous speed of neural trajectories in the full neural space is not modulated during the support of each prior (signed-rank test for zero regression slopes, p = 0.64 for Short, p = 0.08 for Long).
Judged on the paper's test as the claims file fixes; signed-rank p = 0.125 for each prior, the smallest a test over 4 datasets can give, so the outcome carries no evidence for the absence of modulation. All 8 slopes are positive, but noise produces slopes of this size, as later bins average fewer trials. For Poisson units at constant rates with the same trials and Set times, the mean null slope is above the observed one in 8 of 8 and no observed slope exceeds the null's 95th percentile. With only the trials of the longest ts, which last the whole support, slopes are -0.106 to +0.529 per 100 ms and outside the null's 95% range in 1 of 8 (below it). The bootstrap intervals over trials are biased upward (the point estimate is below the interval in 2 of 8) and are not used. The observed slopes are below the null mean in 8 of 8 with all trials and in 8 of 8 with fixed trials, while the null's mean speed over the support is below the observed one in 7 of 8 (null / observed 0.65 to 1.18), so the null is not noisier than the data. This points to a speed that falls during the support, if the units' noise is no larger than Poisson; it was not tested further. The session shows no speed increase beyond what noise produces.
How the test differed from the paper’s method
- 4 datasets, where the paper had 8.
- Speed from smoothed trial-averaged rates with attrition in 20 ms bins of 48 units; the paper's smoothing for this measure is not stated.
ReproducedResults, The curved manifold hypothesis captures the variance; Figure 6D
The mean of the single-trial GPFA projections is sigmoidal (signed-rank test on increments, p = 0.0078) and has a higher slope for the Short prior (one-tailed signed-rank, p = 0.0625).
Middle increments of the mean projection exceed the extreme ones in 8 of 8 (two-sided signed-rank p = 0.0078) and the Short slope is the higher in 4 of 4 datasets (one-tailed p = 0.0625), the paper's values; the mean rises with ts in 8 of 8. With 12 latents, 8 of 8 (p = 0.0078) and 4 of 4.
How the test differed from the paper’s method
- As for gpfa-variance-inverted-u.

Not tested in a notebook 3

3 data missing from the dandiset

BlockedResults, Recurrent network models; Figure 7A to 7C and S7· Data missing from the dandiset
Recurrent neural networks trained on the two-prior task with scalar measurement noise behave like a Bayesian observer and show curved trajectories during the support of the prior, like DMFC.
The claim is about recurrent networks trained in silico. The dandiset holds only the two NWB files of spikes and trials, and the trained networks and the training code are not public (the code is available from the corresponding author on request). Training new networks from the Methods would be a new study, not a test against this data.
BlockedResults, Recurrent network models; Figure 7D and 7E· Data missing from the dandiset
In trained networks, compressing states along the encoding axis increases the regression to the mean of tp, and translating them shifts tp.
As for rnn-curved-manifold; the perturbation experiments are on the trained networks.
BlockedResults throughout; Figures 1 to 6 and S1 to S6· Data missing from the dandiset
The behavioural and neural results hold in monkey G as in monkey H.
The dandiset holds one session of monkey H. A search of the DANDI archive for Jazayeri, Sohn, DMFC and Ready-Set-Go finds no other dandiset with these recordings (only 000620, 000897 and 001275 of the same lab, on other tasks). The paper says the data are in the article and the code is available on request.

Findings about the data

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

Data missingThe dandiset holds one recording session (2016-12-11) of monkey H, with 54 units, where the paper analyses 17 sessions of H and 12 of monkey G (617 and 741 units recorded, 536 and 636 kept for analysis). The dandiset's description does not say that it holds one session of the paper's 29 or which one. The 48 units that pass the paper's rate rule match its count for the single session of H used for GPFA.
Data missingFollowing the Neural Latents Benchmark design, 283 trials of the session are in the test file without ts, tp, prior, effector or direction and without 14 of the 54 units, on a separate clock. These trials cannot be used for any analysis of the paper.
Metadata errorThe description of the `iti` column says 500 ms after Short prior trials and 1220 ms after Long ones; the values are the reverse (Short 1220, Long 500), as the paper states.
Description missing`tp` is described as the time between Set and Go, but is a median 17.9 ms longer than go_time minus set_time (5th to 95th percentile 1.1 to 29.8 ms, up to 200.6 ms in used trials). `ts` is the nominal interval, within 23.5 ms of ready_time minus set_time. Neither difference is documented.
Selection not documentedThe `is_outlier` flag is described only as a probabilistic mixture model. It flags 46 completed trials (4.47%); a refit of the Methods' Gaussian and uniform mixture per prior, ts, effector and direction flags 10 of these and no others, and the other 36 have tp from 0.53 to 1.60 times ts.
Description missingUnits carry no single-unit or multi-unit label and no electrode or probe; the electrode table has no positions or impedances and gives the filtering as unknown. The paper distinguishes 456 single units from 902 multi-units.
Methods not availableThe paper's analysis code is available only on request. The code the dandiset links (nlb_tools) loads the data and scores benchmark models, and holds none of the paper's analyses.
Paper inconsistentFor the single high-yield session of each monkey used for GPFA, the Results give the number of neurons as N = 48 for monkey H and N = 107 for monkey G, the Methods as N = 48 in H and 96 in G.

Notebooks

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

Reproduction Bayesian computation through cortical latent dynamics (Sohn et al. 2019)

Tests the findings of Sohn, Narain, Meirhaeghe and Jazayeri (2019) on dandiset 000130, the Neural Latents Benchmark's DMFC_RSG set, which holds one session of monkey H performing the Ready-Set-Go task with 54 units from dorsomedial frontal cortex. It tests the behaviour and the Bayesian observer model, single-unit sensitivity to the prior and elapsed time, the geometry and speed of population trajectories, the alternative mechanisms, and single-trial GPFA trajectories. Monkey G and the recurrent network models are not in the dandiset. All files

Reproducing Sohn et al. (2019) from dandiset 000130
reproduce_sohn_2019.ipynbExecuted, with outputs · Source

Verification

Round 1
2026-10-14Agreed with 23 of 23 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: The first executed notebook with the conclusion cells and the summary removed, the helper module and the claims file, with the paper's accepted manuscript (PMC6805134). The verifier checked the archive's asset list and metadata for the blocked claims and ran synthetic checks of weighted_slope, PCA, encoding_axis, increments, BayesObserver and attrition_rates.
Round 2
2026-10-14Agreed with 23 of 23 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: The second executed notebook in full, with its conclusion cells and summary, the README and this attempt record. The verifier could not get the paper's text (PMC's BioC service returned 429 and Europe PMC 500), so the figures taken from the paper were not checked against it.

Method differences

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

4 datasets, where the paper had 8.short-in-long-control neural-long-more-bias alt-speed-unmodulated
One session of monkey H; the paper's slopes come from 17 sessions of H and 12 of G.beh-slopes
Outliers removed with the file's `is_outlier` flag, which a refit of the Methods' mixture model does not reproduce (see findings).beh-slopes
One session; the paper's p < 1e-43 pools 17 sessions of monkey H.beh-overlap-prior
The test is over the 8 conditions of one session; the paper's is over sessions and conditions of both animals.beh-sigmoid-increments
One session.beh-long-more-bias
Fit to one session per effector and direction; the paper pools sessions.beh-bayes-model
The prior is taken as continuous uniform over each range, as the Results describe; the task used 5 discrete values.beh-bayes-model
The bias R2 threshold of 0.8 is a choice made here; the paper shows the fit graphically.beh-bayes-model
48 units from one session, where the paper's fractions use 536 units of monkey H over 17 sessions.glm-prior-ts
The GLM design with attrition (rows for every trial whose Set comes after each time point) is read from the Methods; the paper's code was not available.glm-prior-ts
The Long prior's ts test uses the Long ts as time points (800 to 1200 ms after Ready); Eq. 3 of the paper does not say which time points were used, and a unit counts as ts-dependent when either prior's test finds it.glm-prior-ts
The time range (320 ms before Ready to 640 ms after Set) and the shuffle reference are chosen here.dfr-overlap-support
48 simultaneously recorded units, where the paper's trial-averaged PCA pools units over 17 sessions; more units can add dimensions.pca-dimensionality
Rates not normalised before PCA; the Methods do not say.pca-dimensionality
te comes from the observer model fitted to this session.encoding-r2
The encoding axis averages difference vectors up to the middle of the support, a choice made here.encoding-r2
4 datasets, where the paper had 8; a signed-rank test over 4 cannot reach p < 0.05, so the criterion counts datasets.encoding-bayes-beats-linear
20 points per prior (4 datasets), where the paper had 40.speed-vs-projection
Speed is the mean step between 20 ms states up to the shortest tp; the Methods' division by the Set to Go duration is shown as a variant.speed-vs-projection
8 dataset-priors of one animal, where the paper's test combines 16 over two animals.initial-speed-biased
20 points per prior, where the paper had 40.speed-predicts-tp
Speed from smoothed trial-averaged rates with attrition in 20 ms bins of 48 units; the paper's smoothing for this measure is not stated.alt-speed-unmodulated
8 dataset-priors of one animal, where the paper's test used 16.alt-transient-parallel
Rates smoothed over 100 ms before to 300 ms after Set and cut to the first 200 ms, a choice made here.alt-transient-parallel
The paper's test (one-tailed signed-rank, p = 1.5e-4) cannot be read unambiguously; the criterion on the median time of the minimum was fixed before the analysis.alt-threshold-monotonic
The 60% of variance follows the figure legend; the 75% of the Methods is shown as a variant.alt-threshold-monotonic
Latent dimensionality chosen only from 4 and 8, with 100 EM iterations, to keep the run short; the paper's cross-validation range is not stated.gpfa-variance-inverted-u
One GPFA per effector and direction with both priors; the paper does not say how trials were grouped.gpfa-variance-inverted-u
As for gpfa-variance-inverted-u.gpfa-mean-sigmoid

Provenance

Attempt record: 000130-sohn-2019-01
Paper: Bayesian computation through cortical latent dynamics
Date: 2026-10-14
Made by: An agent, run by CatalystNeuro
Dandiset version: 0.241017.1448
Harness: Claude Code (GitHub Actions)
Model: claude-opus-5-5
Protocol: pilot-6

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 000127 Dandiset 000147 →