{ "dataset": { "archive": "dandi", "id": "000138", "label": "DANDI:000138" }, "page": "datasets/dandi/000138/", "generated": "2026-11-05T07:41:23+00:00", "caution": "A failure to reproduce is not a finding that the paper is wrong. It may be a failure of the attempt.", "outcome_counts": { "reproduced": 9, "partially_reproduced": 1, "not_reproduced": 2, "blocked": 2 }, "outcome_counts_with_outside_data": { "reproduced": 1, "partially_reproduced": 0, "not_reproduced": 0, "blocked": 0 }, "outcome_counts_elsewhere": { "reproduced": 0, "partially_reproduced": 0, "not_reproduced": 0, "blocked": 0 }, "outcome_counts_links": { "reproduced": 0, "partially_reproduced": 0, "not_reproduced": 0, "blocked": 0 }, "attempts": [ { "id": "000138-churchland-2010-01", "dataset": { "archive": "dandi", "id": "000138", "version": "0.220113.0407", "doi": "10.48324/dandi.000138/0.220113.0407" }, "key": "dandi:000138", "label": "DANDI:000138", "dataset_version": "0.220113.0407", "paper": "10.1016/j.neuron.2010.09.015", "paper_title": "Cortical preparatory activity: representation of movement or first cog in a dynamical machine?", "date": "2026-10-28", "kind": "agent", "accountable": "CatalystNeuro", "agent": { "harness": "Claude Code", "model": "claude-opus-5-5", "protocol": "pilot-12" }, "runner": { "platform": "GitHub Actions", "url": "https://github.com/catalystneuro/encore/actions/runs/40756966342" }, "tier": "agent_verified", "deprecated": "", "deprecated_on": null, "entry": "000138-churchland-2010", "entry_title": "Preparatory tuning in a peri-movement space (Churchland et al. 2010)", "description": "Tests the findings of Churchland, Cunningham, Kaufman, Ryu and Shenoy (2010) on dandiset 000138 (MC_Maze_Large), which holds 500 maze-task trials over 27 conditions and 162 array units from one session of monkey J. It compares preparatory and peri-movement tuning, and how well a preferred direction captures preparatory tuning in peri-movement, kinematic and task spaces, with the controls of Figure 7 and the oscillator of Figure 8. The single-electrode datasets of monkeys A, B and J and the EMG are not in the dandiset.", "notebooks": [ { "path": "reproduce_churchland_2010.ipynb", "title": "Reproducing Churchland et al. (2010) from dandiset 000138", "executed": "notebooks/000138-churchland-2010/reproduce_churchland_2010.html" } ], "outcomes": [ { "claim": "fig4abc-tuning-correlation-single-electrode", "statement": "In the three single-electrode datasets (monkeys A, B, J) the correlation across conditions between a neuron's preparatory and peri-movement firing rates is positive on average but small (mean r 0.21, 0.19 and 0.06; mean Spearman r 0.20, 0.20 and 0.06).", "source": "Results, Correlations between preparatory and peri-movement tuning, paragraph 1; Figure 4A-C", "quote": "Across the four datasets the mean correlation was positive (p=0.01, p=0.01, p=0.12, p=0.03) but only very modestly so (mean r = 0.21, 0.19, 0.06 and 0.13). Non-parametric correlations were similarly modest (mean Spearman's r = 0.20, 0.20, 0.06 and 0.11)", "statistics": [ { "about": "mean correlation above zero, monkey A", "test": "not named", "p": 0.01 }, { "about": "mean correlation above zero, monkey B", "test": "not named", "p": 0.01 }, { "about": "mean correlation above zero, monkey J", "test": "not named", "p": 0.12 } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "data_missing", "notebook": null, "notes": "The dandiset holds only array recordings from one session of monkey J. The single-electrode recordings of monkeys A, B and J (64, 74 and 172 units) are not in it, and no other dandiset or public deposit of them is named by the paper or the archive.", "deviations": [], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig4d-tuning-correlation-j-array", "statement": "In the J-array dataset the mean correlation between preparatory and peri-movement tuning is positive but small (mean r 0.13, p = 0.03; mean Spearman r 0.11).", "source": "Results, Correlations between preparatory and peri-movement tuning, paragraph 1; Figure 4D", "quote": "Across the four datasets the mean correlation was positive (p=0.01, p=0.01, p=0.12, p=0.03) but only very modestly so (mean r = 0.21, 0.19, 0.06 and 0.13)", "statistics": [ { "about": "mean correlation above zero, J-array", "test": "not named", "p": 0.03 } ], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Under the main reading 20 of 162 units pass the inclusion criterion (more than 5 spikes/s of preparatory tuning, as a range) and the 1.5 times peri-movement criterion. Mean Pearson r 0.195 and mean Spearman r 0.212, both above tolerance (0.13 +/- 0.013 and 0.11 +/- 0.011); the mean r is above zero by all three tests (t-test p = 0.023, on Fisher z p = 0.023, Wilcoxon p = 0.024), which the criterion calls a partial reproduction. Over the eight trial-selection and onset readings with a result, mean r is 0.105 to 0.195 (within tolerance in five), mean Spearman r 0.146 to 0.212 (within tolerance in none), and the mean r is above zero at p < 0.05 in two; in the other six the smallest p of the three tests is 0.074 to 0.169, so those readings give not reproduced. With tuning as the SD across conditions no unit passes. With a ratio of 1.0 in place of 1.5, 52 units give mean r 0.137 (p = 0.0041). The correlation is positive and small, as the paper says; its size and significance at 27 conditions depend on the readings.", "deviations": [ "The dandiset holds 27 of the 108 conditions and 500 trials (11 to 18 per condition after selection) where the paper's J-array dataset had a mean 2166 trials per neuron; it has 162 units where the paper reports 146.", "Assumed main reading, as the readings give different outcomes. Trials are those whose concatenated horizontal and vertical velocity profile (movement onset -150 to +800 ms) correlates with the condition's mean profile at r > 0.9 (Supplemental Experimental Procedures, Task design; the window and how the two axes combine are not stated). Requiring each axis to pass keeps 81% of trials and leaves 2 or 3 trials in some conditions, as one velocity component of nearly vertical or horizontal reaches is mostly noise; the concatenated test keeps 96%, and the paper reports \"typically ~90%\" with outliers only occasional. Movement onset is the trial table's `move_onset_time`; the paper refers to Churchland et al. 2006c, not available here, and the speed-threshold reading (10% of peak speed, a median 14 ms later) is also run.", "Tuning in spikes/s is taken as the range of condition means. The SD reading leaves 6 or 7 of 162 units above 5 spikes/s, which does not fit a criterion the paper describes as lower than for single electrodes to suit array data (Methods, Neural recordings and datasets).", "The test of the mean against zero is not named; the three readings of the claims file are run and agree under the main reading." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig4e-correlation-time-course", "statement": "The correlation of preparatory tuning with peri-movement tuning measured at single time points starts to fall about 200 ms before movement onset and is near zero by movement onset.", "source": "Results, Correlations between preparatory and peri-movement tuning, paragraph 2; Figure 4E", "quote": "The correlation with preparatory activity began to decline \u223c200 ms prior to movement onset, and reached values near zero by movement onset ( Figure 4E )", "statistics": [], "kind": "finding", "inferred": "", "outcome": "not_reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Mean correlation over the 20 Figure 4D units is 0.773 at -400 ms, 0.764 at -300 ms, 0.713 at -250 ms, 0.572 at -200 ms and 0.218 at movement onset; it falls below 0.1 only between +150 and +200 ms (0.110, 0.064). The decline starts about 250 ms before onset and the value at onset is below half of that at -250 ms, as the criterion asks, but it is not below 0.1 at onset. All eight readings give 0.129 to 0.218 at onset and meet the other two parts. The measured value is the correlation at movement onset under the main reading. Every time point the criterion uses (-400 ms to onset) lies inside the movement-locked trace, so the splice points of the paper's trace do not enter.", "deviations": [ "The rate at each time point comes from the movement-locked average, the part of the paper's spliced trace locked to movement onset (Supplemental Experimental Procedures, Computing the mean firing rate versus time); the paper chose the splice points to minimise discontinuities and does not give them. The go-locked curve is shown alongside.", "27 of 108 conditions and 20 units, as for fig4d-tuning-correlation-j-array." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "results-ap-location-no-effect", "statement": "The tuning correlation of Figure 4 does not depend on anterior-posterior recording location in any dataset (r = -0.05, -0.03, -0.07 and 0.20; p = 0.82, 0.85, 0.52 and 0.27).", "source": "Results, Comparison of anterior versus posterior sites, paragraph 1", "quote": "Regressing tuning correlation (as in Figure 4 ) against AP location yielded little or no effect: r =-0.05, p =0.82; r =-0.03, p =0.85; r =-0.07, p =0.52; and r =0.20, p =0.27 for the four datasets", "statistics": [ { "about": "monkey A", "test": "regression of tuning correlation on AP location", "statistic": "r = -0.05", "p": 0.82 }, { "about": "monkey B", "test": "regression of tuning correlation on AP location", "statistic": "r = -0.03", "p": 0.85 }, { "about": "monkey J", "test": "regression of tuning correlation on AP location", "statistic": "r = -0.07", "p": 0.52 }, { "about": "J-array", "test": "regression of tuning correlation on AP location", "statistic": "r = 0.20", "p": 0.27 } ], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Only the J-array dataset can be tested. Regressing the per-unit correlation of Figure 4D on array (prefix 1 coded anterior) gives r = 0.047, p = 0.844 under the main reading (12 anterior and 8 posterior units, mean r 0.209 and 0.176), and p from 0.432 to 0.844 over the eight readings, so no reading finds an effect at p < 0.05. The paper reports r = 0.20, p = 0.27 for J-array. The other three datasets are not in the dandiset.", "deviations": [ "Electrode coordinates are NaN, so reading (a), the AP coordinate on the array, cannot be run. Reading (b), array membership, is used.", "Assumed array of each unit. Every unit references an electrode of the PMd group, but the unit ids carry a prefix (1 for 83 units, 2 for 79) and a channel, and the referenced electrode is the channel minus one for both prefixes. The prefix is taken as the array, and prefix 1 as PMd, the group the references name. Prefix 1 units also have more preparatory tuning (median range 6.71 against 4.79 spikes/s), consistent with the paper's remark that preparatory activity is more prevalent at anterior sites; this is the only support for the mapping. Swapping the arrays would only flip the sign of r, and the p-value does not depend on it.", "Monkeys A, B and J (single-electrode) are not tested." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "supp-preparatory-dimensionality", "statement": "Depending on the dataset, 7 to 10 dimensions capture 90% of the variance of preparatory activity across conditions.", "source": "Supplemental Experimental Procedures, Estimating the upper limit of PD performance", "quote": "depending on the dataset, 7-10 dimensions were required to capture 90% of the data variance", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Under the main reading 11 components capture 90% of the variance of the 27 x 162 matrix of preparatory condition means, inside the widened range 6.3 to 11, and 10 with the 92 included units. Under the per-axis trial selection the counts are 12 (outside) and 11; with the 6 or 7 units of the SD tuning reading the count is 4. On each half of the trials the count is 14 and 13, so trial noise adds dimensions at these trial counts. The count of 11 sits at the upper end of the range, so the outcome is fragile. Only J-array is tested, with 27 conditions where the paper had 108.", "deviations": [ "Assumed main reading for trial selection, as in fig4d-tuning-correlation-j-array; the claims file's readings of which units enter (all, or included) both fall in range under it.", "The single-electrode datasets (per 27-condition set for monkey J) are not in the dandiset." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig6-peri-movement-space-best", "statement": "For every dataset, preferred directions capture preparatory tuning best (highest cross-validated performance) when expressed in the peri-movement space, usually at middle dimensionalities.", "source": "Results, PD performance in a variety of spaces, paragraph 2; Figure 6A-D", "quote": "For every dataset, the highest level of performance was achieved ... when PDs were expressed in the peri-movement space", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Under the main reading, over 92 included units, the peri-movement space's best mean performance is 0.351 (k = 10), above the kinematic space (0.103, k = 4) and every task space (best, kitchen sink, 0.250). Over the 2048 combinations of readings it is best in 1468; the criterion asks for one combination. It is best in 95% of combinations with tuning as a range and 49% with the SD (6 or 7 units), and in 94% with all units in the T-matrix against 50% with included units only. Best k is 7 or more in 1776 of 2048 combinations. EMG was not recorded for J-array.", "deviations": [ "Only J-array, on 27 of 108 conditions; the claim says \"for every dataset\" and the other three are not in the dandiset.", "Task spaces are measured on each condition's mean hand trajectory from movement onset to the reach end (standard choice; the paper speaks of the \"average endpoint and peak velocity for each condition\"). Acceleration for the kinematic space is the derivative of the mean velocity (standard choice). Replicating the six kinematic variables to the number of units gives the same PD fits, so they are used once.", "Movement onset, trial selection and tuning as in fig4d-tuning-correlation-j-array; every reading is run." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig6-each-dimensionality", "statement": "At each dimensionality taken alone the peri-movement space almost always performs best; the exception is the speed task at three dimensions, where two task spaces perform very well, while endpoints + maximum speed does not perform well for the maze task.", "source": "Results, PD performance in a variety of spaces, paragraph 2; Figure 6A-D", "quote": "Even if we consider each dimensionality individually, performance was almost always higher for the peri-movement space. An exception was that, among 3-dimensional spaces, the \u2018targs+instr. spd.\u2019 and \u2018endpoints + max. spd.\u2019 performed very well for the datasets collected using the speed task. The former is undefined for the maze task, but the latter did not perform well.", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Under the main reading the peri-movement space wins 20 of 22 comparisons (0.909, at least 0.8 needed) and beats endpoints + max speed at k = 3 (0.126 against 0.0003). Over all combinations the criterion is met in 1074 of 2048 (share won from 0 to 1, median 0.85). It is met in 69% of combinations with the concatenated trial selection and 36% with the per-axis one, 75% and 30% for tuning as range or SD, 67% and 38% for all or included units in the T-matrix, and 37% and 68% with the space computed once or refit. The kinematic space at k = 1 and k = 14 is the one it loses to under the main reading (by 0.005 and 0.004). With the space refit and the main reading otherwise, it wins 21 of 22 comparisons (losing only to the kinematic space at k = 1) and beats endpoints + max speed at k = 3 (0.132 against 0.0003), so the criterion is met under that reading too.", "deviations": [ "Assumed main reading, as the readings give different outcomes. Besides the readings of fig4d-tuning-correlation-j-array, the PCA space is computed once from all conditions, and only the PD fit leaves a condition out. The Methods say \"One condition was left out and we found the PD that best captured tuning for the remaining conditions\", the Results say \"we computed the PD based on all but one condition\", and both describe leaving out a condition only for the PD; Figure 5 plots the left-out conditions in the same space. All units enter the T-matrix (\"neurons with weak peri-movement activity were included\", Methods, PCA-based spaces). Performance is averaged over units (\"Performance (averaged across neurons)\", Results, PD performance in a variety of spaces).", "Only J-array, on 27 of 108 conditions." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig6-dimensionality-trend", "statement": "Performance of the PCA-based spaces typically rises with dimensionality and falls beyond about 10 dimensions, a sign of overfitting.", "source": "Results, PD performance in a variety of spaces, paragraph 1; Figure 6A-D", "quote": "Performance typically improved with dimensionality. ... Performance typically declined at dimensionalities >10, an indication of over-fitting.", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Under the main reading peri-movement performance rises from 0.056 at k = 1 to 0.351 at k = 10 and falls to -0.456 at k = 14; the kinematic space does the same (0.061, best 0.103 at k = 4, -0.451 at k = 14). The criterion is met in 992 of 2048 combinations, 91% of those with the space computed once and 6% of those with the space refit without the left-out condition; refit, performance rises at every k (0.054 at k = 1 to 0.461 at k = 14 under the main reading otherwise) and the best k is 10 to 14. With the space computed once, 14 components are fit from 27 conditions, so the fall may reflect the small number of conditions here (the paper had 108).", "deviations": [ "Assumed main reading, with the space computed once from all conditions, as argued under fig6-each-dimensionality. The outcome rests mainly on this choice.", "Only J-array, on 27 of 108 conditions." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig6-pd-adequate-all-spaces", "statement": "The PD performs adequately (above chance, performance above zero) in most spaces, as expected from correlations among movement parameters.", "source": "Results, Testing preparatory tuning in PCA-based spaces, paragraph 3; Figure 6", "quote": "In fact, the PD generally performed adequately regardless of the space in which it was expressed.", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Under the main reading mean performance is above zero for the peri-movement space (0.351) and kinematic space (0.103) at their best k, and for 7 of 8 task spaces; endpoints + max velocity is -0.125 and endpoints + max speed 0.0003. The criterion is met in 1936 of 2048 combinations; the others have 3 or 4 task spaces above zero.", "deviations": [ "Assumed main reading, as for fig6-each-dimensionality.", "Only J-array, on 27 of 108 conditions." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig6e-summary-rank", "statement": "Combined across monkeys A, B and J, mean performance is highest for the peri-movement space at every maximum dimensionality (3, 6, 12), the advantage grows with dimensionality, and task spaces are outcompeted by the kinematic and EMG spaces, which are outperformed by the peri-movement space; spaces marked with asterisks are significantly worse than the best (p < 0.001).", "source": "Results, PD performance in a variety of spaces, paragraph 3; Figure 6E and legend", "quote": "Performance (averaged across neurons) was highest when the PD employed the space based on peri-movement neural activity. This performance differential increased at higher dimensionalities ... the task spaces were largely outcompeted by the PCA-based kinematic and EMG spaces, which were in turn outperformed by the peri-movement space", "statistics": [ { "about": "spaces marked significantly worse than the best space", "test": "not named", "p": 0.001, "p_relation": "<", "n": "neurons of monkeys A, B and J" } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "data_missing", "notebook": null, "notes": "The summary combines monkeys A, B and J (single-electrode) and uses their EMG spaces; none of these data are in the dandiset, which holds only the J-array session.", "deviations": [], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig7a-shuffled-space-poor", "statement": "A control peri-movement space in which each neuron's peri-movement pattern is inverted on a random half of conditions performs poorly.", "source": "Results, Controls, paragraph 1; Figure 7A and legend", "quote": "The PCA-based \u2018shuffled-space\u2019 performed poorly ( Figure 7A )", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Under the main reading, over 20 draws, the shuffled space's best mean performance is 0.014 (k = 3), below the peri-movement space at every k and below the kinematic space's best (0.103). With 20 draws in every variant, the criterion also holds with proto-trial (b), onset (b), included units in the T-matrix, pooled performance, the kinematic window ending with the reach, and the space refit per fold, the last by a narrow margin (shuffled best 0.159 against a kinematic best of 0.160, peri-movement best 0.461). That margin of 0.001 is within the noise of the draws (standard error of the shuffled best 0.009 over 20 draws). It fails only with the per-axis trial selection, where the peri-movement space peaks at 0.150 (k = 6) and falls to -1.666 at k = 10 and -3.102 at k = 14, and the shuffled space is above it at k = 9 to 14.", "deviations": [ "Assumed main reading, as for fig6-each-dimensionality.", "Judged on J-array alone, where Figure 7A averages all datasets." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig7b-window-start-onset", "statement": "Sliding the start of the peri-movement window from 150 ms before movement onset to movement onset barely lowers performance, which stays at 97% (monkey J) and 94% (monkeys A, B) of its original value.", "source": "Results, Controls, paragraph 1; Figure 7B and legend", "quote": "Even when the window began at movement onset ... performance was still at 97% (monkey J), and 94% (monkeys A, B) of its original value.", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Under the main reading the best performance with the window from movement onset is 0.318 (k = 9) against 0.351 (k = 10) from -150 ms, a ratio of 0.908, inside 0.873 to 1.067. Over all combinations the ratio is within tolerance in 1336 of 2048 (median 0.917, range 0.367 to 2.004), in 47% of combinations with the space computed once and 84% with it refit. Reading (a) of the claims file (J-array against 0.97) is the one run; the 0.94 for monkeys A and B, and reading (b), which needs the J single-electrode dataset, cannot be tested. The measured value is for the first expected value.", "deviations": [ "Assumed main reading, as for fig6-each-dimensionality.", "Monkeys A, B and the J single-electrode dataset are not in the dandiset." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig7c-anterior-posterior", "statement": "The peri-movement space captures preparatory tuning at both anterior and posterior sites, with normalized performance only slightly different from that of random halves of the data, slightly better at anterior sites.", "source": "Results, Comparison of anterior versus posterior sites, paragraph 1; Figure 7C and legend", "quote": "The ability of the peri-movement space to capture tuning also held for both anterior and posterior sites ( Figure 7C ). ... Normalized performance was only slightly different than if we down-selected the data randomly.", "statistics": [], "kind": "finding", "inferred": "", "outcome": "not_reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Under the main reading, with the upper limit from the preparatory PCA space of all units in each subset at k = 6, normalized performance is 0.597 for the anterior array (55 included units) and 0.382 for the posterior array (37), against 0.620 for random halves, ratios of 0.963 and 0.615. Anterior is within the 10% bound and posterior is not; posterior is below the 5th percentile of the 40 random halves (0.512 to 0.751, SD 0.079). Anterior is better than posterior, as the paper reports. The criterion is met in 2 of 16 combinations of readings (upper limit with or without the unit being fit, at k = 6 or best k, space once or refit, trial selection, T-matrix units), both with the space refit and the unit kept in its upper-limit space; with the space computed once it fails under both upper-limit readings. With the unit left out, the ratios are 0.817 and 0.847 under the main reading otherwise, but the random halves exceed the upper limit in 11 of 40 halves (95th percentile 1.252, SD 0.190), and under the per-axis selection at k = 6 the random-half mean is -0.826 with SD 10.8. The outcome rests on the space being computed once, as for fig6-dimensionality-trend.", "deviations": [ "Assumed array of each unit from the id prefix, as for results-ap-location-no-effect; with a binary location, the median split gives the two arrays.", "The upper-limit space is \"derived from the preparatory activity of all neurons (T was c\u00d7n with each row containing the mean preparatory activity for every neuron)\" and \"provides a rough upper limit on performance\" (Supplemental Experimental Procedures, Estimating the upper limit). No exclusion of the unit being fit is stated, where the Methods state one for the peri-movement space (\"all neurons except the neuron whose preparatory activity we were trying to capture\"), so the main reading keeps the unit in its upper-limit space. Left out, the random halves exceed the upper limit. Both readings are run and every row counts. Whether the space is taken at k = 6 or its best k is open in the claims file; both are run.", "Random halves are 20 random splits of all units, 40 halves; the paper does not say how many.", "Judged on J-array alone, where Figure 7C averages the datasets." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig7d-pd-consistency", "statement": "In all four datasets and for every comparison, the PD is more stable across random subsets of conditions when measured in the peri-movement space than in task, kinematic or EMG spaces.", "source": "Results, Stability of the PD, paragraph 1; Figure 7D and legend", "quote": "Yet for every comparison, and for all four datasets, the PD was most stable when measured in the peri-movement space.", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "Under the main reading, over 92 units and 100 draws each, the peri-movement space has the higher mean consistency in all three comparisons, 0.756 against 0.610 (endpoints + max speed, k = 3), 0.684 against 0.570 (segment directions, k = 6) and 0.748 against 0.666 (kinematic, k = 4, its best k). It is higher in every comparison under all eight variants of the readings and with the axes as they are or scaled to unit SD, so the outcome does not rest on these choices. EMG was not recorded for J-array.", "deviations": [ "Only J-array; the claim covers four datasets.", "The spaces are computed once from all 27 conditions; whether axes are scaled before the PD is fitted is not stated, and both are run.", "With 27 conditions, a draw of 25 with replacement holds a mean 16.5 distinct conditions and two draws share a mean 10.1, so the two PDs are fitted to more overlapping data than with the paper's 108 conditions. How much this changes consistency was not measured." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig8-oscillator-model", "statement": "In a simulated two-unit linear oscillator held at 25 initial states and then released, preparatory and peri-movement activity of a unit are nearly uncorrelated (r = 0.08), yet a PD in the two-dimensional peri-movement space captures the preparatory tuning perfectly.", "source": "Results, Tuning from a dynamical systems perspective, paragraphs 1 and 2; Figure 8 and legend", "quote": "The correlation of preparatory activity with peri-movement activity (averaged over the gray interval) was 0.08. ... The PD points towards the pattern recruited by the activity of unit one (panel A ), and perfectly captures preparatory \u2018tuning\u2019.", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_churchland_2010.ipynb", "notes": "A simulation with no recorded data. Over 1000 draws of 25 initial states, the correlation between unit 1's held state and its mean released activity has a median of -0.019 to 0.012 and a central 95% of about -0.41 to 0.40 under each reading (uniform or Gaussian initial states, a full or half cycle), which contains 0.08. A PD in the 2-D PCA space of the released activity of both units captures unit 1's initial state with leave-one-out performance 1.0 in every draw. The measured value is the median r under the first reading.", "deviations": [ "The paper's 0.08 comes from one draw; the criterion judges it against the distribution over draws.", "The hold lasts 50 steps (it cannot change the result, as the held state is constant)." ], "assumed": false, "disagreed": false, "data": "with_inputs", "where": null, "proposed": null } ], "findings": [ { "kind": "data_missing", "statement": "The dandiset holds one session of monkey J's array recordings on the maze task, cut to 500 trials over 27 conditions for the Neural Latents Benchmark. The paper's J-array dataset had 108 conditions and a mean 2166 trials per neuron, and its single-electrode datasets of monkeys A, B and J and all EMG recordings are not in the dandiset, so two of the paper's main findings cannot be tested and the rest only on J-array.", "found": null, "status": null }, { "kind": "metadata_error", "statement": "Every unit's electrode reference points into the PMd group of the electrodes table. The unit ids encode a prefix (1 for 83 units, 2 for 79) and a channel, and the referenced electrode is the channel minus one for both prefixes, so the 79 units of one array reference electrodes of the other. The heldout units split 20 and 20 between the prefixes.", "found": null, "status": null }, { "kind": "metadata_insufficient", "statement": "Which id prefix is the PMd array and which the M1 array is not stated, and all electrode coordinates (x, y, z) are NaN, so neither array membership nor position on the array can be read from the files. Anterior-posterior analyses here rest on an assumption.", "found": null, "status": null }, { "kind": "identifier_missing", "statement": "Nothing in the dandiset or the paper says whether the 2009-10-06 session is the one behind the paper's J-array dataset. The file has 162 units where the paper reports 146 isolations, and 27 of its 108 conditions.", "found": null, "status": null }, { "kind": "other", "statement": "Hand, cursor and eye position are at 1 kHz and only within trials, where the paper tracked the hand at 60 Hz and the eye at 240 Hz; the resampling is not described. Hand velocity agrees with the derivative of hand position (slope 1.000).", "found": null, "status": null }, { "kind": "other", "statement": "Eight trials have a delay of 0 ms or less (down to -2 ms), where the `delay` field is described as the delay between target presentation and go cue.", "found": null, "status": null }, { "kind": "paper_inconsistency", "statement": "The mean trials per neuron of the J-array dataset is 2166 in the Results and 2155 in the Methods.", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the paper's authors. Is the PCA space (peri-movement and kinematic) computed once from all conditions, with only the PD regression leaving a condition out, or recomputed without the left-out condition? Computed once, performance peaks at k = 10 and falls beyond (as the paper reports); recomputed, it keeps rising to k = 14, and the outcomes of fig6-each-dimensionality, fig6-dimensionality-trend, fig7a and fig7c change. This attempt assumed the space is computed once.", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the paper's authors. How was the maze proto-trial defined, over which window, and did the horizontal and vertical profiles each need r > 0.9 or the two together? Requiring each keeps 81% of trials here and leaves 2 or 3 trials in some conditions; the concatenated profile keeps 96%; the paper says about 90%. The fig4d result moves between mean r 0.105 and 0.195 across these readings and the movement onset method. This attempt assumed the condition's mean profile, the concatenated test and the window -150 to +800 ms.", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the paper's authors. What does \"spikes/s of preparatory tuning\" measure (the range of condition means, their SD, or another quantity), and how was movement onset detected? With the range, 92 of 162 units pass the 5 spikes/s criterion; with the SD, 6. This attempt assumed the range and used the dataset's movement onset times.", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the paper's authors. For Figure 7C, did the upper-limit space of a subset include the neuron being fit, and at which dimensionality was it taken? The normalized posterior performance relative to random halves is 0.62 with the neuron at k = 6 and 0.85 without it, and without it random halves exceed the upper limit in 11 of 40 halves; this attempt assumed the neuron included, as the Supplemental Experimental Procedures derive the space from all neurons and state no exclusion.", "found": null, "status": null } ], "reconciliation": null, "verifications": [ { "date": "2026-10-28", "kind": "agent", "by": "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 outcome table in its first cell was left in, and the verifier did not use it), the helper module and the claims file, with the paper. The verifier checked the helper functions against brute-force PCA and regression on synthetic data. Its report gave its own outcome for each claim; all 15 agree with the record. It also noted that the dandiset is probably not the paper's J-array session (27 conditions and about 14 trials per condition against 108 conditions), which the record states, and asked for per-condition trial counts under each reading, now printed in the data check.", "agreed": 15, "of": 15, "disagreements": [] }, { "date": "2026-10-28", "kind": "agent", "by": "Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook)", "reviewed": "Round 2. The full executed notebook after the first revision, with its conclusions and summary, the helper module, the claims file, the README and this record, with the paper. The verifier agreed with all 15 outcomes and found no misreading of the paper in the claims file. It listed sentences the printed numbers do not support (cell 25 Spearman range, the README run time, the summary's test-file sentence, the fig7a reason for the per-axis failure, the fig7d effect of overlapping draws) and noted that the round-1 reviewed text omitted that cell 0 still held the outcome table.", "agreed": 15, "of": 15, "disagreements": [] } ] } ], "entries": [ { "name": "000138-churchland-2010", "type": "reproduction", "title": "Preparatory tuning in a peri-movement space (Churchland et al. 2010)", "description": "Tests the findings of Churchland, Cunningham, Kaufman, Ryu and Shenoy (2010) on dandiset 000138 (MC_Maze_Large), which holds 500 maze-task trials over 27 conditions and 162 array units from one session of monkey J. It compares preparatory and peri-movement tuning, and how well a preferred direction captures preparatory tuning in peri-movement, kinematic and task spaces, with the controls of Figure 7 and the oscillator of Figure 8. The single-electrode datasets of monkeys A, B and J and the EMG are not in the dandiset.", "datasets": [ "dandi:000138" ], "papers": [ "10.1016/j.neuron.2010.09.015" ], "derived_from": [], "notebooks": [ { "path": "reproduce_churchland_2010.ipynb", "title": "Reproducing Churchland et al. (2010) from dandiset 000138", "executed": "notebooks/000138-churchland-2010/reproduce_churchland_2010.html" } ] } ] }