{
 "dandiset": "000239",
 "page": "dandisets/000239/",
 "generated": "2026-10-19T16:23:39+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": 20,
  "partially_reproduced": 5,
  "not_reproduced": 3,
  "blocked": 11
 },
 "outcome_counts_with_outside_data": {
  "reproduced": 0,
  "partially_reproduced": 0,
  "not_reproduced": 0,
  "blocked": 0
 },
 "outcome_counts_elsewhere": {
  "reproduced": 0,
  "partially_reproduced": 0,
  "not_reproduced": 0,
  "blocked": 0
 },
 "attempts": [
  {
   "id": "000239-xu-2022-01",
   "dandiset": "000239",
   "dandiset_version": "0.230607.1752",
   "paper": "10.1038/s41586-022-04478-7",
   "paper_title": "Cortical processing of flexible and context-dependent sensorimotor sequences",
   "date": "2026-10-17",
   "kind": "agent",
   "accountable": "CatalystNeuro",
   "agent": {
    "harness": "Claude Code (GitHub Actions)",
    "model": "claude-opus-5-5",
    "protocol": "pilot-7"
   },
   "tier": "agent_verified",
   "entry": "000239-xu-2022",
   "entry_title": "Cortical processing of flexible sensorimotor sequences (Xu et al. 2022)",
   "description": "Tests the findings of Xu, Dong, Chen, Delgado, Hughes, Zhang and O'Connor (2022) on dandiset 000239, which holds the trial tables, tongue tracking and sorted spikes of mice performing a sequence licking task. It decodes tongue kinematics, sequence identity and sequence time from S1TJ, M1TJ, ALM and S1BF populations, classifies backtracking and zigzag sequence context, and checks the behavioural measures and unit counts. The optogenetic and control experiments cannot be tested, as the files do not say which region was inhibited or which condition each session was.",
   "notebooks": [
    {
     "path": "reproduce_xu_2022.ipynb",
     "title": "Reproducing Xu et al. (2022) from dandiset 000239",
     "executed": "notebooks/000239-xu-2022/reproduce_xu_2022.html"
    }
   ],
   "outcomes": [
    {
     "claim": "ephys-unit-counts",
     "statement": "Silicon probe recordings gave 1537 single units and 303 multi-units from 57 recording sessions.",
     "source": "Results, \"Sequence tiling of single-unit responses\"; Extended Data Fig. 4a-d",
     "outcome": "partially_reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "The 57 files described as ephys ALM, M1TJ, S1TJ, S1BF, M1B, S1L and ZZ hold 1917 units, within 10% of the paper's 1840. With an ISI violation rate (2.5 ms) of at most 1% and a contamination of at most 15%, 1578 are single units (1537 in the paper, within tolerance) and 339 multi-units (303, outside 272.7 to 333.3). Other thresholds give 1490 to 1732 single units; three variants put the multi-units inside the tolerance, the contamination criterion alone at 15% (331), the RPV criterion alone at 0.5% (309) and both criteria with a 20% contamination threshold (295). Other estimates of the contamination at 15% give 341 multi-units (0.5 ms censored period) and 478 (spikes inside the trial span only, T the whole span). On synthetic units the estimate is 0.163 for a true contamination of 0.15 and 0.233 for 0.20, so 85% of units at 0.15 are called multi-units. Multi-units range from 0 to 21 per session. The 22 \"ephys others\" sessions (742 units) and 14 opto efficiency recordings (414 units) are not among the 57.",
     "deviations": [
      "The contamination estimate (Hill et al. 2011) uses a censored period of 0 and the mean rate over 1 s bins of the trial span holding a spike, with all of a unit's spikes in the count; the Methods give none of these.",
      "Units recorded outside the target structure were excluded in the paper; the files do not say which these are, so all units are counted."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig3-session-counts",
     "statement": "The population analyses of Figure 3 use 8 S1TJ, 9 M1TJ, 13 ALM and 5 S1BF sessions.",
     "source": "Figure 3b legend",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Counted from the session descriptions, the only place the files name the recorded area. The sessions come from 4, 4, 8 and 3 mice, and every one holds units and at least 13 selected standard trials per direction.",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig3-units-per-session",
     "statement": "Regressions used the simultaneously recorded units of each session, 32 plus or minus 13 units (mean and SD).",
     "source": "Results, \"Hierarchical population coding\"",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Reading (b), all 57 recording sessions, gives 33.6 and 13.0 units, both within tolerance. Reading (a), the 35 Figure 3 sessions the claim refers to, gives a mean of 32.0 and an SD of 14.4, just outside 11.7 to 14.3. Single units alone give 26.4 and 13.2. Unit counts do not depend on a random seed, so the SD of reading (a) is 14.4 under every seed.",
     "deviations": [
      "The reading that matches, all 57 sessions, is wider than the sessions the Results describe."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig3a-trial-count",
     "statement": "Figure 3a averages 2684 trials of standard and backtracking sequences.",
     "source": "Figure 3a legend",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "The Methods' trial selection gives 2169 standard and 717 backtracking trials over the 35 sessions; together 2886 (reading a), within 10% of 2684. Standard trials alone are 2169. The count is a third of the candidates (at least 10) and does not depend on the search. The greedy search finds the best subset in 43 of 70 exhaustive checks on 16 candidates (at most 8.4% above the best mean distance); a swap search improves on it by at most 4.9% on the full sets, sharing a median of 97% of the trials.",
     "deviations": [
      "The subset of most similar trials is found by a greedy search; the Methods do not say how it was searched."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig3c-length",
     "statement": "Tongue length L is decoded from population activity best in S1TJ and M1TJ and poorly in ALM and S1BF (R2 0.56, 0.42, 0.11 and 0.04).",
     "source": "Figure 3b,c; Results, \"Hierarchical population coding\"",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Mean R2 by area (S1TJ, M1TJ, ALM, S1BF) are each within the SD the figure prints, and S1TJ and M1TJ exceed ALM and S1BF. With folds that keep each trial in one fold, 0.551, 0.392, 0.087 and 0.016, same order. Decoding L with spikes shifted by 25 to 100 ms lowers R2 in S1TJ.",
     "deviations": [
      "The inter-lick-interval standardisation of the Methods is not applied before the fits.",
      "Fits use scikit-learn's elastic net path (l1 ratio 0.1, 30 penalties, standardised predictors, one-SE rule), not MATLAB's lasso; samples are assigned to folds at random, as MATLAB does by default. Random folds let the autocorrelation within a trial into the test folds; R2 values are higher than with folds that keep trials together, the orderings the same.",
      "The fitting window is -0.5 to 0.5 s from the middle touch, read from the Methods' remark that 0.5 to 0.8 s of the decoding was new data; the paper does not state it.",
      "The expected values were read from the figure image, as the text does not give them."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig3c-velocity",
     "statement": "Tongue velocity L' is decoded best in S1TJ and M1TJ and poorly in ALM and S1BF (R2 0.35, 0.32, 0.07 and 0.01).",
     "source": "Figure 3b,c; Results, \"Hierarchical population coding\"",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "All four means within tolerance of 0.35, 0.32, 0.07 and 0.01, and the order holds with either fold scheme. L' is the derivative of L where the tongue is out.",
     "deviations": [
      "As for fig3c-length."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig3c-angle",
     "statement": "Tongue angle theta is encoded comparably in S1TJ, M1TJ and ALM, but not in S1BF (R2 0.59, 0.57, 0.54 and 0.12).",
     "source": "Figure 3b,c; Results, \"Hierarchical population coding\"",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "All four means within tolerance of 0.59, 0.57, 0.54 and 0.12. S1TJ, M1TJ and ALM differ by 0.049 and S1BF is below each. With trial folds the spread is 0.059. theta is the angle of the tongue base-to-tip vector, used where L exceeds 1 mm.",
     "deviations": [
      "As for fig3c-length. The sign of theta (left or right) cannot be fixed from the files; R2 does not depend on it."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig3c-identity",
     "statement": "Sequence identity I is encoded most strongly in ALM, then progressively less in M1TJ, S1TJ and S1BF (R2 0.67, 0.42, 0.23 and 0.14).",
     "source": "Figure 3b,c; Results, \"Hierarchical population coding\"",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Means (S1TJ, M1TJ, ALM, S1BF) within tolerance of 0.23, 0.42, 0.67 and 0.14, in the order ALM > M1TJ > S1TJ > S1BF. With trial folds 0.188, 0.323, 0.657 and 0.049, same order.",
     "deviations": [
      "As for fig3c-length."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig3c-time",
     "statement": "Relative sequence time tau is encoded most strongly in ALM, then progressively less in M1TJ, S1TJ and S1BF (R2 0.60, 0.45, 0.17 and 0.08).",
     "source": "Figure 3b,c; Results, \"Hierarchical population coding\"",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Means within tolerance of 0.17, 0.45, 0.60 and 0.08, in the stated order; with trial folds 0.145, 0.381, 0.601 and 0.063, same order.",
     "deviations": [
      "As for fig3c-length. tau is the time from the middle touch without lick standardisation."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed5b-variables-uncorrelated",
     "statement": "Any pair of the five behavioural variables (L, L', theta, I, tau) is uncorrelated.",
     "source": "Results, \"Hierarchical population coding\"; Extended Data Fig. 5b",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Mean absolute correlations of the ten pairs range from 0.000 (I-tau, zero by the balanced design) to 0.118 (theta-I); the largest single-session value is 0.368 (L-theta).",
     "deviations": [
      "The bound of 0.3 is a choice; the paper gives no threshold."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed5e-variance-explained",
     "statement": "Across regions, the summed variance explained by the five coding axes reached about half that of the top five principal components.",
     "source": "Results, \"Hierarchical population coding\"; Extended Data Fig. 5e",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "The median ratio over 35 sessions is 0.454 (mean 0.490, IQR 0.387 to 0.544), inside 0.45 to 0.55 but close to its lower bound; per area medians are 0.433 to 0.476. One coding axis of 175 had all weights zero. With seeds 3, 11, 12 and 13 the median is 0.453 to 0.455, inside the range each time.",
     "deviations": [
      "VE is computed on -0.5 to 0.8 s from the middle touch of the selected standard trials; the paper says \"during sequence execution\"."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed5f-axes-orthogonal",
     "statement": "The five coding axes were largely orthogonal to each other.",
     "source": "Results, \"Hierarchical population coding\"; Extended Data Fig. 5f",
     "outcome": "partially_reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Nine pairs have a mean absolute cosine of 0.114 to 0.225; L' and theta have 0.302, just above the bound of 0.3 the criterion fixed. Random directions in the same sessions give 0.154. With seeds 3, 11, 12 and 13, L' and theta give 0.302 to 0.320 and are the only pair at or above 0.3 each time.",
     "deviations": [
      "The bound of 0.3 is a choice; the paper says \"largely orthogonal\"."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig3e-cca",
     "statement": "After canonical correlation, ALM trajectories in the top three PCs correlate best with the theta, I, tau coding subspace (group 2), more than with L, L', theta (group 1) or all five variables with five PCs (group 3); in S1TJ and M1TJ the correlation is high only when the coding subspace includes L and L'.",
     "source": "Figure 3d,e; Results, \"Hierarchical population coding\"",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Mean canonical correlation, groups 1, 2, 3; ALM 0.607, 0.900, 0.743; S1TJ 0.746, 0.589, 0.735; M1TJ 0.712, 0.662, 0.809. All five comparisons hold by exact sign-flip tests (ALM p = 0.0002 twice; S1TJ p = 0.023 and 0.008; M1TJ p = 0.020). The figure marks p < 0.001, which an exact paired permutation cannot reach with 8 or 9 sessions; the criterion uses 0.05.",
     "deviations": [
      "Decoded variables come from the fits of fig3c-length, without lick standardisation."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig3f-backtracking-svm",
     "statement": "ALM and M1TJ activity predicts whether the port backtracked from the initial missed lick on; S1TJ shows only an insignificant trend and S1BF no prediction.",
     "source": "Figure 3f; Results, \"Sequence branching signals in ALM, M1TJ\"",
     "outcome": "partially_reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Sessions with backtracking in both directions are 6 ALM (as in the paper), 9 M1TJ, 8 S1TJ and 5 S1BF. Mean accuracy in 0.2 to 0.4 s after the trigger is 0.740 (ALM), 0.734 (M1TJ), 0.595 (S1TJ) and 0.556 (S1BF), against null 95th percentiles of 0.526, 0.523, 0.516 and 0.515; before 0.1 s all are 0.487 to 0.519. ALM and M1TJ decode as stated; S1TJ and S1BF also exceed the null, which the paper reports as an insignificant trend and no prediction. In this window the classes differ in behaviour (1.854 against 0.113 missed licks, ed2e-missed-licks), which S1TJ and S1BF activity may reflect.",
     "deviations": [
      "The paper's null and confidence intervals come from a hierarchical bootstrap; here the null is 20 label shuffles per session, averaged over sessions.",
      "Trials are balanced across directions by random subsampling.",
      "The PCA before the SVM is fitted on all trials of a session, without labels, before the cross-validation."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig4g-zigzag-context",
     "statement": "In the zigzag task, ALM population activity predicts the sequence identity (context) during subsequences with the same licking movements, across 6 sessions from 3 mice.",
     "source": "Figure 4f,g; Results, \"Context-dependent coding of subsequences\"",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "The 6 ZZ sessions come from 3 mice with 18 to 79 perfectly executed trials per sequence. Mean accuracies of the four comparisons are 0.785 to 0.858 against null 95th percentiles of 0.510 to 0.518; the lowest session is 0.651. By mouse 0.730 to 0.916. The files do not name the recorded area of the ZZ sessions; the paper says ALM.",
     "deviations": [
      "Subsequences are aligned on touches (the seventh touch against the third or eleventh), not on the cross-correlogram of tongue angle, and the extra selection of trials closest to the median angle is not applied.",
      "The null mean over bins uses every fifth 5 ms bin."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed7f-target-position-alm",
     "statement": "During the inter-trial interval, ALM populations encode the upcoming target position more strongly than the other regions.",
     "source": "Extended Data Fig. 7e,f; Results, \"ALM encodes upcoming sequences\"",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Mean R2 of the upcoming direction in the second before the cue is 0.271 (ALM), 0.097 (M1TJ), 0.075 (S1TJ) and 0.051 (S1BF); ALM is higher than each (Mann-Whitney p = 0.0025, 0.0007, 0.0014). With trial folds 0.159 against 0.022 to 0.034 (p = 0.023 to 0.047). Per ALM mouse 0.055 to 0.642. The 13 ALM sessions come from 8 mice; with one value per mouse, p = 0.012 to 0.024.",
     "deviations": [
      "The paper's ITI window and trial set are not fully stated; the second before the cue of every standard trial is used."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed7h-no-time-coding",
     "statement": "None of the regions, including ALM, encoded time or distance to the trial start during the inter-trial interval.",
     "source": "Extended Data Fig. 7h; Results, \"ALM encodes upcoming sequences\"",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Mean R2 of the time to the cue is 0.008 to 0.034 by region (largest session 0.091, ALM), below 0.1 everywhere; 0.004 to 0.018 with trial folds.",
     "deviations": [
      "The paper gives no test; the bound of 0.1 was fixed in the criterion.",
      "Only the last second before the cue is used, where the paper may have used more of the interval."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed4m-active-neuron-counts",
     "statement": "Extended Data Fig. 4m counts 324 ALM, 233 M1TJ and 119 S1TJ neurons.",
     "source": "Extended Data Fig. 4m legend",
     "outcome": "partially_reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Single units with a maximal PETH rate of at least 10 spikes/s are 351 (ALM) and 223 (M1TJ), within 10% of 324 and 233, and 155 (S1TJ) against 119. With multi-units, 387, 261 and 251. At 5 and 15 spikes/s the single-unit counts are 387, 259, 164 and 321, 200, 140.",
     "deviations": [
      "PETHs are not computed on standardised inter-lick intervals.",
      "The unit classification is that of ephys-unit-counts, with its open choices."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed4l-alm-initiation",
     "statement": "More ALM neurons are tuned to sequence initiation than neurons of the other regions.",
     "source": "Results, \"Sequence tiling of single-unit responses\"; Extended Data Fig. 4l",
     "outcome": "partially_reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "91 of 351 active ALM neurons (25.9%) peak in the initiation window, against 48 of 223 in M1TJ (21.5%, Fisher one-sided p = 0.14) and 20 of 155 in S1TJ (12.9%, p = 0.0006). S1BF, outside the criterion, has 17 of 55 (30.9%), above ALM, against the statement's \"the other regions\". The windows differ in length (0.5, 2 and 1.2 s); per second of window ALM (0.519) is still above M1TJ (0.430) and S1TJ (0.258).",
     "deviations": [
      "The paper shows histograms of peak times without a test; the test was chosen in the criterion."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed4o-rhythmicity",
     "statement": "S1TJ and M1TJ contain more neurons modulated by individual licks than ALM.",
     "source": "Results, \"Sequence tiling of single-unit responses\"; Extended Data Fig. 4n,o",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Median r_avg is 0.201 (S1TJ), 0.156 (M1TJ) and 0.093 (ALM); one-sided Mann-Whitney against ALM p = 2.5e-36 and 4.1e-24.",
     "deviations": [
      "Without lick standardisation the rhythm of trial-averaged PETHs is blurred away from the middle touch, which lowers r in every region."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed6-alm-reward-drop",
     "statement": "ALM contains neurons that fire during sequence execution and decrease firing abruptly when the tongue first contacts water, although licking continues.",
     "source": "Results, \"Reward modulation in ALM\"; Extended Data Fig. 6a",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Of 351 active ALM single units, 188 decrease and 92 increase from 0.5 s before to 0.5 s after the first water touch (p < 0.01 each; binomial p = 5e-09). S1TJ 50 against 51, S1BF 16 against 15, M1TJ 96 against 74 (p = 0.054).",
     "deviations": [
      "The paper shows example neurons and decoded tau; the test over units was chosen in the criterion."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed1h-sequence-duration",
     "statement": "Expert mice completed the 7 port positions in about a second.",
     "source": "Results, first section; Extended Data Fig. 1h",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Over 25438 standard trials of 17 mice (the paper's number), the median over mice from the first port-moving touch to the first touch at the last position is 0.992 s (0.890 to 1.138 s by mouse); from the cue, 1.324 s. Recording sessions alone give 1.004 s.",
     "deviations": [
      "The paper does not list the sessions of Extended Data Fig. 1h; recording sessions (except zigzag) and the opto sessions' trials without light are used."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed1l-learning-curves",
     "statement": "Over training, 15 mice reduced their sequence initiation time and increased their sequence speed.",
     "source": "Extended Data Fig. 1l; Results, first section",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "15 mice have learning files. Median over mice of the initiation time falls from 2.573 s (first 100 standard trials) to 0.291 s (last 100), Wilcoxon p = 0.00043, and the sequence speed rises from 4.902 to 6.038 positions/s, p = 0.0043; 14 of 15 mice improve in both. MX181003 does not; its file of 2018-10-02 was left out as a duplicate, so its other learning files may also be mislabelled. Without MX181003, 14 of 14 improve (p = 0.00012 each).",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed1l-proficiency-1500",
     "statement": "Mice reached proficiency in standard sequences after about 1500 trials of training.",
     "source": "Results, first section; Extended Data Fig. 1l-n",
     "outcome": "not_reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Under reading (a) the median over mice of the trial at which sequence speed reaches 90% of its final value is 200 (150 and 250 at 80% and 95%); under reading (b), initiation time within 110% of final, 1100. Both are outside 1350 to 1650. One mouse (MX200101) reaches neither. The paper does not define proficiency, and the learning files may not start at the first day of training. Every mouse's first file holds standard sequences in both directions and almost nothing else (2 other trials in MX170903, 1 in MX180803 and 1 in MX181002), and in 3 of 15 mice (MX170903, MX180201, MX181003) the median initiation time of the first 100 standard trials is already under 1 s; if earlier training is missing, both counts are too low. In the median mouse the speed reaches 90% of its final value within 200 trials, so the result under reading (a) also reflects how proficiency was defined. MX200101 has no standard trials among its last 500 learning trials, so it has no final value.",
     "deviations": [
      "Proficiency was defined in the criterion, as the paper gives no definition."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed2-backtracking-counts",
     "statement": "Backtracking analyses used 7458 standard and 2695 backtracking sequences from 47 sessions of the 9 mice trained on backtracking.",
     "source": "Extended Data Fig. 2c,e legends; Methods, \"Behavioural training\"",
     "outcome": "not_reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "Reading (c) gives 55 sessions of 9 mice with 10365 standard and 3839 backtracking trials; only the number of mice matches. Readings (a) and (b) give 74 and 59 sessions. A variant outside the criterion leaves out the 8 sessions whose touches are mostly a few ms long (5th percentile 3 to 7 ms, against 14 ms or more in the others), which fits the Methods' 8 sessions \"where only touch onsets were correctly registered\"; it gives 47 sessions, the paper's number, of 8 mice, with 8891 and 3207 trials. On these 47 sessions, removing the first and last trial and trials not finished within the video (the Methods' exclusions) leaves 8627 and 3104; both above tolerance. Adding initiation under 1 s, the Figure 3 trial selection and not an exclusion the Methods give for Extended Data Fig. 2, leaves 8058 and 2894, both within tolerance. This chain was built after seeing the results and does not decide the outcome; it has 8 mice where the paper has 9.",
     "deviations": [
      "The paper's own trial exclusions (first and last trial, unfinished videos) are not applied in the readings of the criterion; they are shown as a variant."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed2e-missed-licks",
     "statement": "After the port backtracked, mice took on average 1 to 2 missed licks before relocating it.",
     "source": "Results, first section; Extended Data Fig. 2e",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "In 29 Figure 3 sessions with backtracking (8 mice), backtracking trials have a mean of 1.854 missed licks in the fourth interval, against 0.113 in standard trials (Mann-Whitney p printed as 0). By mouse 0.97 to 5.39. A touch tolerance of 0 or 25 ms in place of 12.5 ms gives the same mean on the same trials.",
     "deviations": [
      "Sessions are the Figure 3 sessions with backtracking, not the paper's 47."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed6b-consummatory-licks",
     "statement": "After the first lick that touched water, mice continued with about 5 consummatory licks.",
     "source": "Results, \"Reward modulation in ALM\"; Extended Data Fig. 6b",
     "outcome": "not_reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "After the first lick touching water, the bout continues for a median of 6 video licks (mean 6.38, 6864 trials), or 4 touches (mean 3.87); counting the first water lick, 7. None is within 4.5 to 5.5. The video ended during the bout in 88 trials, which are left out. By mouse, 4 to 10 video licks and 2 to 7 touches. Under (a) the most common count is 5 (20.2% of trials; 19.6% at 6, 12.8% at 4). With the bout threshold of (b) taken from video lick intervals (238 ms, against 243 ms from touches) the median is still 4 touches.",
     "deviations": [
      "The paper does not say whether licks or touches are counted or where the bout ends; the bout rule (1.5 times the median interval) is the Methods' rule for lick bouts."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed7b-correct-side",
     "statement": "Expert mice usually initiated sequences from the correct side without exploring the other side.",
     "source": "Results, \"ALM encodes upcoming sequences\"; Extended Data Fig. 7b",
     "outcome": "reproduced",
     "blocked_by": null,
     "notebook": "reproduce_xu_2022.ipynb",
     "notes": "In all 11 mice of the Figure 3 sessions, 0.562 to 0.923 of standard trials start with no lick on the other side of the midline before the first touch (0.647 to 0.980 with a 5 degree margin); pooled 0.714. The paper shows 8 mice.",
     "deviations": [
      "The side of the port is taken from the angle of the lick that made the first touch, as the files do not say which side each position is on."
     ],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig2b-s1tj-angle-variability",
     "statement": "Inhibiting S1TJ increased the across-trial variability of lick angle, SD(theta_shoot), without reducing the lateral reach |theta_shoot|, and did not shorten licks.",
     "source": "Figure 2b; Extended Data Fig. 3j",
     "outcome": "blocked",
     "blocked_by": "metadata_insufficient",
     "notebook": null,
     "notes": "The 35 opto sessions at 4 mW (described \"2s 5V\") are 5 per mouse for 7 mice, one per region presumably, but no file names the inhibited region; trial tables mark only light times, and the opto1 and opto2 series give the LED drive. The 25 sessions at 2 mW are the same. Checked in all 74 opto files, in the trial columns (opto takes 0, 1, 2 or NaN, without a description), the session description (cut short in the files themselves), general (only experimenter, institution, lab and subject), acquisition (empty), stimulus (empty presentation and templates groups), and the attributes of the opto1 and opto2 series (unit \"categorical\", no description).",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig2b-almm1tj-angle-length",
     "statement": "Inhibiting ALM/M1TJ reduced the ability to direct licks to the sides and decreased lick length; inhibiting M1B only slightly increased angle variability; inhibiting S1BF or S1Tr changed nothing.",
     "source": "Figure 2b; Extended Data Fig. 3j",
     "outcome": "blocked",
     "blocked_by": "metadata_insufficient",
     "notebook": null,
     "notes": "As fig2b-s1tj-angle-variability; the inhibited region of each session is not stated.",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig2c-almm1tj-initiation",
     "statement": "Inhibiting ALM/M1TJ at sequence initiation strongly suppressed licking (licks largely absent in 4 of 7 mice), and less strongly at mid-sequence.",
     "source": "Figure 2c; Extended Data Fig. 3k",
     "outcome": "blocked",
     "blocked_by": "metadata_insufficient",
     "notebook": null,
     "notes": "As fig2b-s1tj-angle-variability.",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "fig2c-termination",
     "statement": "Inhibiting S1TJ or M1B at water consumption impaired the stopping of ongoing licking.",
     "source": "Figure 2c; Extended Data Fig. 3k",
     "outcome": "blocked",
     "blocked_by": "metadata_insufficient",
     "notebook": null,
     "notes": "As fig2b-s1tj-angle-variability.",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed3lm-half-power",
     "statement": "With half the illumination power (2 mW) the effects were consistent with, though weaker than, those at 4 mW, and the termination deficit of S1TJ and M1B inhibition was similarly strong.",
     "source": "Extended Data Fig. 3l,m; Results, \"Optogenetic inhibition screen\"",
     "outcome": "blocked",
     "blocked_by": "metadata_insufficient",
     "notebook": null,
     "notes": "The 25 sessions at 2 mW (\"2s 2.5V\", 5 mice) do not name the inhibited region either.",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed3f-inhibition-efficiency",
     "statement": "At 4 mW, light within 1 mm reduced the spike rate of putative pyramidal cells by 91%, at about 1.5 mm by 61% and at about 3 mm by 19%; FS neurons within 1 mm increased by 739%.",
     "source": "Methods, \"Photostimulation\"; Extended Data Fig. 3c-f",
     "outcome": "blocked",
     "blocked_by": "data_missing",
     "notebook": null,
     "notes": "The 14 efficiency recordings (4 mice, 414 units) hold spike times and light trials, but no spike waveforms, which the paper uses to separate putative pyramidal and FS cells, and no distance of the fibre from the recording site. The distance is also missing from the metadata (metadata_insufficient).",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed1i-earplug",
     "statement": "Hearing loss with earplugs did not increase missed licks during sequences.",
     "source": "Extended Data Fig. 1i; Methods, \"Hearing loss\"",
     "outcome": "blocked",
     "blocked_by": "metadata_insufficient",
     "notebook": null,
     "notes": "The 10 earplug sessions are two per mouse for 5 mice, as the Methods describe, but nothing in the files says which session had earplugs and which was the sham. Without the labels, the paper's test (paired one-tailed bootstrap of missed licks over mice) finds an effect under 4 of the 32 possible assignments, so the absence cannot be shown label-free. In all 5 mice the earlier session has the longer time to first touch (by 0.40 to 1.49 s), as the paper reports for earplugs; under that inferred assignment missed licks differ by +0.032 per trial (p = 0.36), no effect, as the claim says. The inferred labels do not decide the outcome.",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed1j-odor-masking",
     "statement": "Masking odours with an air flow did not change missed licks.",
     "source": "Extended Data Fig. 1j; Methods, \"Odor masking\"",
     "outcome": "blocked",
     "blocked_by": "metadata_insufficient",
     "notebook": null,
     "notes": "The 6 air sessions (6 mice) switch the air flow after the 100th trial, on first in two mice and off first in four; the files do not say which. Over the 15 possible choices of the two on-first mice, the paper's test finds an effect in one (smallest p = 0.018), so the absence cannot be shown label-free.",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed1k-tongue-numbing",
     "statement": "Numbing the tongue with lidocaine increased missed licks, so mice require tactile feedback from the tongue.",
     "source": "Extended Data Fig. 1k; Methods, \"Tongue numbing\"",
     "outcome": "blocked",
     "blocked_by": "metadata_insufficient",
     "notebook": null,
     "notes": "The 11 numbing sessions (6 mice, one with a single session) do not say which session was lidocaine and which saline. In all 5 mice with two sessions the earlier session has more missed licks in its first 30 min (0.99 to 2.74 against 0.23 to 1.07 per trial), significant in 4; the direction of the claim needs the labels.",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed1f-lick-force",
     "statement": "The maximal force of a lick was on average about 4 mN.",
     "source": "Methods, \"Two-axis optical force sensors\"; Extended Data Fig. 1f",
     "outcome": "blocked",
     "blocked_by": "metadata_insufficient",
     "notebook": null,
     "notes": "The force sensor series (tubeH, tubeV) are in volts; the calibration to mN of Extended Data Fig. 1d is not in the files; their conversion is 1 and offset 0, and the paper gives no slope.",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    },
    {
     "claim": "ed4j-nnmf-clusters",
     "statement": "Clustering the PETHs of all neurons by NNMF was most consistent with 13 clusters.",
     "source": "Methods, \"PETH and NNMF clustering\"; Extended Data Fig. 4j,k",
     "outcome": "blocked",
     "blocked_by": "methods_underspecified",
     "notebook": null,
     "notes": "The consistency measure needs cluster labels matched across bootstrap iterations, and the Methods do not say how; the paper's code could not be found.",
     "deviations": [],
     "disagreed": false,
     "data": "dandiset",
     "where": null,
     "proposed": null
    }
   ],
   "findings": [
    {
     "kind": "metadata_insufficient",
     "statement": "The opto sessions do not name the inhibited region; 35 sessions at 4 mW and 25 at 2 mW, five per mouse, carry only the light times and LED drive. Figure 2 and Extended Data Fig. 3j-m cannot be tested."
    },
    {
     "kind": "metadata_insufficient",
     "statement": "The control sessions do not mark the condition; which earplug session had earplugs, which half of each air session had the air flow on, and which numbing session was lidocaine are not stated."
    },
    {
     "kind": "metadata_insufficient",
     "statement": "The recorded area and the kind of session are given only by the session description (the name of the lab's data folder); there is no electrodes table, unit location, depth or quality label, and the area of the zigzag (ZZ) and \"ephys others\" sessions is not given."
    },
    {
     "kind": "data_missing",
     "statement": "Spike waveforms are not included, so units cannot be classified as putative pyramidal or FS cells (Extended Data Fig. 3c-f), and the fibre distance of the efficiency recordings is not given."
    },
    {
     "kind": "format_error",
     "statement": "1446 spike times in 35 files are duplicated within a unit, away from trial boundaries (median 2.9 s). The Methods say these Phy duplicates were removed before analysis; the deposited spike times keep them."
    },
    {
     "kind": "other",
     "statement": "One training recording is deposited twice; the MX181003 learning file of 2018-10-02 15:30:57 repeats all 4028 touches of the MX181002 file with the same start time and then continues, so it is unclear which mouse it belongs to."
    },
    {
     "kind": "format_error",
     "statement": "In the 574 learning files start_time and stop_time count from 0 while cue, touch, port and water times are on another clock (around 10^6 s), so trial intervals and events do not line up."
    },
    {
     "kind": "format_error",
     "statement": "The seqId column of the learning files holds the text of an unconverted MATLAB categorical object, not a sequence identity."
    },
    {
     "kind": "metadata_insufficient",
     "statement": "Force sensor and lick port series are in volts without a calibration, tongue landmarks are in pixels without a scale (the Methods give 32 um per pixel), and every trial column is described as \"no description\"."
    },
    {
     "kind": "selection_undocumented",
     "statement": "The Methods leave out 8 sessions whose touch offsets were not registered; the files do not mark them. Eight sessions with backtracking have touches of a few ms (5th percentile 3 to 7 ms), and leaving them out gives the paper's 47 sessions."
    },
    {
     "kind": "metadata_error",
     "statement": "The session description is cut short (\"... se\") in the files themselves, so the full name of the lab's data folder, the only place the area and kind of session are given, is lost."
    },
    {
     "kind": "metadata_insufficient",
     "statement": "The trial column opto (values 0, 1 and 2) and the opto1 and opto2 series (unit \"categorical\") have no description, and no file has optogenetics metadata (light source, site, power)."
    },
    {
     "kind": "methods_unavailable",
     "statement": "The paper says its MATLAB analysis code is available at GitHub but gives no address, and no repository was found under the authors' accounts."
    }
   ],
   "verifications": [
    {
     "date": "2026-10-17",
     "kind": "agent",
     "by": "Independent agent run by CatalystNeuro (verifier-1; same harness, separate context, did not write the notebook)",
     "reviewed": "Round 1. The executed notebook with the conclusion cells and the summary removed, the helper module and the claims file; the paper from PMC9109820. The archive's metadata was checked for the blocked claims.",
     "agreed": 39,
     "of": 39,
     "disagreements": []
    },
    {
     "date": "2026-10-17",
     "kind": "agent",
     "by": "Independent agent run by CatalystNeuro (verifier-2; same harness, separate context, did not write the notebook)",
     "reviewed": "Round 2. The full executed notebook with conclusions and summary, the helper module, the claims file, the entry README and this record; the paper from PMC9109820. In reply the entry was corrected and re-run; the README's run time, area count and claim count were also corrected.",
     "agreed": 39,
     "of": 39,
     "disagreements": []
    }
   ]
  }
 ],
 "entries": [
  {
   "name": "000239-xu-2022",
   "type": "reproduction",
   "title": "Cortical processing of flexible sensorimotor sequences (Xu et al. 2022)",
   "description": "Tests the findings of Xu, Dong, Chen, Delgado, Hughes, Zhang and O'Connor (2022) on dandiset 000239, which holds the trial tables, tongue tracking and sorted spikes of mice performing a sequence licking task. It decodes tongue kinematics, sequence identity and sequence time from S1TJ, M1TJ, ALM and S1BF populations, classifies backtracking and zigzag sequence context, and checks the behavioural measures and unit counts. The optogenetic and control experiments cannot be tested, as the files do not say which region was inhibited or which condition each session was.",
   "dandisets": [
    "000239"
   ],
   "papers": [
    "10.1038/s41586-022-04478-7"
   ],
   "derived_from": [],
   "notebooks": [
    {
     "path": "reproduce_xu_2022.ipynb",
     "title": "Reproducing Xu et al. (2022) from dandiset 000239",
     "executed": "notebooks/000239-xu-2022/reproduce_xu_2022.html"
    }
   ]
  }
 ]
}