{ "dataset": { "archive": "dandi", "id": "000447", "label": "DANDI:000447" }, "page": "datasets/dandi/000447/", "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": 7, "not_reproduced": 2, "blocked": 14 }, "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 }, "outcome_counts_links": { "reproduced": 0, "partially_reproduced": 0, "not_reproduced": 0, "blocked": 0 }, "attempts": [ { "id": "000447-tang-2023-01", "dataset": { "archive": "dandi", "id": "000447", "version": "0.250724.2224", "doi": "10.48324/dandi.000447/0.250724.2224" }, "key": "dandi:000447", "label": "DANDI:000447", "dataset_version": "0.250724.2224", "paper": "10.1016/j.celrep.2023.112246", "paper_title": "Geometric transformation of cognitive maps for generalization across hippocampal-prefrontal circuits", "date": "2026-10-30", "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/40994732733" }, "tier": "agent_verified", "deprecated": "", "deprecated_on": null, "entry": "000447-tang-2023", "entry_title": "Hippocampal and prefrontal maps across a novel and a familiar W-maze (Tang, Shin and Jadhav 2023)", "description": "Tests the findings of Tang, Shin and Jadhav (2023) on dandiset 000447, which holds CA1 and prefrontal units, position and trial tables of five rats running a W-maze alternation task in a novel, a familiar and again the novel room. It compares spatial maps across environments, Bayesian decoding with novel-room templates, selectivity and path equivalence of single cells, UMAP manifolds of the pooled populations, and cross-condition generalization. The de novo learning sessions and the second group of rats behind Figure 1 are not in the dandiset.", "notebooks": [ { "path": "reproduce_tang_2023.ipynb", "title": "Reproducing Tang, Shin and Jadhav (2023) from dandiset 000447", "executed": "notebooks/000447-tang-2023/reproduce_tang_2023.html" } ], "outcomes": [ { "claim": "fig1b-faster-novel-learning", "statement": "All five rats acquired the task faster in the novel environment than during de novo learning; proportion correct over the first 50 trials is higher in novel learning.", "source": "Figure 1B; Results, Rats generalize learned rules across contexts, paragraph 2", "quote": "Compared with de novo learning, all animals acquired the inbound and outbound task rules faster in the novel environment", "statistics": [], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "data_missing", "notebook": null, "notes": "The de novo learning sessions are not in the dandiset, which holds only the N-F-N' day (three epochs per file, as its description says). The paper says its raw data are on Brandeis servers, available on request (Data and code availability), so they are not public. The novel sessions are present; the comparison needs the de novo curve.", "deviations": [], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig1-denovo-final-performance", "statement": "Final performance at the end of de novo learning was 84.3% correct (SD 7.7%).", "source": "Results, Rats generalize learned rules across contexts, paragraph 2", "quote": "final performance, 84.3% \u00b1 7.7%, mean \u00b1 SD", "statistics": [], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "data_missing", "notebook": null, "notes": "The de novo learning sessions are not in the dandiset, which holds only the N-F-N' day (three epochs per file, as its description says). The paper says its raw data are on Brandeis servers, available on request (Data and code availability), so they are not public.", "deviations": [], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig1c-session-performance", "statement": "Proportion correct differs across de novo session 1, final de novo session, N, F and N', with performance in N above the first de novo session.", "source": "Figure 1C", "quote": "Proportion correct against session number (data points for individual animals; repeated-measures ANOVA with Tukey\u2019s post hoc, F = 40.48 , n.s., p > 0.75 , ***p = 0.0002, **p = 0.0045)", "statistics": [ { "about": "Tukey post hoc pair (pair not named in the legend)", "test": "Tukey's post hoc", "p": 0.0002 }, { "about": "Tukey post hoc pair (pair not named in the legend)", "test": "Tukey's post hoc", "p": 0.0045 }, { "about": "Tukey post hoc pair (pair not named in the legend)", "test": "Tukey's post hoc", "p": 0.75, "p_relation": ">" } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "data_missing", "notebook": null, "notes": "The de novo learning sessions are not in the dandiset, which holds only the N-F-N' day (three epochs per file, as its description says). The paper says its raw data are on Brandeis servers, available on request (Data and code availability), so they are not public. Session means for N, F and N' are in the notebook under fig1c-familiar-above-novel.", "deviations": [], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig1c-familiar-above-novel", "statement": "In both groups of rats, performance in the familiar environment is higher than in the novel one whatever the order of exposure (N-F-N' group in Figure 1C, F-N-F'-N' group in Figure S1C).", "source": "Results, Rats generalize learned rules across contexts, paragraph 2; Figure 1C; Figure S1C", "quote": "We found that for both animal groups, the performance in the familiar environment was significantly higher than that in the novel one regardless of the order of experience", "statistics": [ { "about": "F-N-F'-N' group, a Tukey pair of familiar against novel sessions (pair not named)", "test": "repeated-measures ANOVA with Tukey's post hoc", "statistic": "F = 48.98", "p": 0.0053, "n": "3 rats" }, { "about": "F-N-F'-N' group, a second Tukey pair (pair not named)", "test": "repeated-measures ANOVA with Tukey's post hoc", "p": 0.0005, "p_relation": "<", "n": "3 rats" } ], "kind": "finding", "inferred": "", "outcome": "not_reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "On N, F and N' of the five rats the repeated-measures ANOVA gives F(2,8) = 1.08, p = 0.383, and Tukey's test does not put F above N (difference of the means 0.060, p = 0.70). F is above N in four rats (by 0.078 to 0.222) and below it in ZT2 (by 0.282), whose F session has 24 trials at 0.50 correct. ZT2's trials table is complete, 22 well-to-well transits counted from position against 24 rows, with the trials tiling the 1554-s epoch; ZT2 spent 0.33 of that session above 5 cm/s, the smallest fraction of any session. Each session estimated alone gives p = 0.46. The F-N-F'-N' rats of Figure S1C, which the criterion tests only if their behaviour is in the dataset, are not.", "deviations": [ "The ANOVA is over N, F and N' only; the de novo sessions of the paper's five-group ANOVA of Figure 1C are not in the dandiset, so Tukey's test uses the error term of a three-group ANOVA.", "The novel curve is estimated on N and N' together, as cal_behavperform.m runs the novel sessions together; the familiar curve on F alone, as the de novo days the code would include are missing. Background probability 1/3 (cal_behavperform.m)." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs1b-first-session-inbound-outbound", "statement": "In the first session, both inbound and outbound performance are higher in novel learning (N) than in de novo learning.", "source": "Figure S1B", "quote": "Note the higher performance for both inbound and outbound during novel learning than de novo learning (****p < 0.0001, rank-sum tests)", "statistics": [ { "test": "rank-sum tests", "p": 0.0001, "p_relation": "<" } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "data_missing", "notebook": null, "notes": "The de novo learning sessions are not in the dandiset, which holds only the N-F-N' day (three epochs per file, as its description says). The paper says its raw data are on Brandeis servers, available on request (Data and code availability), so they are not public.", "deviations": [], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "cells-per-session", "statement": "Cells were tracked across environments, with 30.5 (SD 9.8) CA1 and 37.4 (SD 19.3) PFC cells per session per animal.", "source": "Results, PFC spatial map generalizes, while CA1 completely remaps, across contexts, paragraph 1", "quote": "Cells were stably tracked across environments (mean \u00b1 SD = 30.5 \u00b1 9.8 CA1 and 37.4 \u00b1 19.3 PFC cells per session per animal", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "27.53 (SD 8.68) CA1 and 36.53 (SD 18.60) PFC cells per session per animal over the 15 animal-sessions, within the stated 30.5 (SD 9.8) and 37.4 (SD 19.3). With at least 100 spikes (Methods) 27.67 and 36.53; counting each cell once per animal 34.8 and 38.6. The region of a unit is read from its tetrode number (see the findings).", "deviations": [ "Cells follow the Methods (at least 100 spikes) and the code (poolcells_acrossep.m, more than 100 spikes, CA1 mean rate below 7 Hz); the code's per-animal interneuron lists cannot be applied, as the files carry no cell numbers." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs3a-ca1-global-remapping", "statement": "CA1 spatial similarity between the two environments is no different from that of cell-ID shuffles, that is, CA1 remaps globally.", "source": "Results, PFC spatial map generalizes, while CA1 completely remaps, across contexts, paragraph 1; Figure S3A", "quote": "Notably, the distributions of CA1 spatial-representation similarity between two environments were not different from distributions obtained by shuffling cell identities in the same environment", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "CA1 spatial correlation, F-N' against the cell-ID shuffle, Tukey p = 0.075 (F-N' higher by 0.074; medians -0.059 and -0.064). The margin is thin; over 100 draws of the shuffle p is below 0.05 in 13, and F-N' is always higher (by 0.014 to 0.105). The 4-cm and 2-cm widths give p = 0.21 and 0.40. Rate overlap does not differ either (p = 0.55 and 0.30). JS17 keeps some CA1 spatial correlation across rooms (median r 0.31; the other rats -0.11 to -0.05).", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times.", "The shuffle pairs each cell's F map with another cell's N' map of the same animal (a random derangement within the animal); the Methods say \"a different cell\" without saying whether across animals. One draw decides the outcome; 100 draws are shown." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs3a-pfc-similar-ca1-lower", "statement": "Single-cell rate overlap and spatial correlation of PFC cells across environments (F-N') are as large as across repeated exposures to the novel environment (N-N'), whereas in CA1 they are lower across environments.", "source": "Figure S3A", "quote": "Note that the similarity of PFC firing across familiar-novel environments (F-N\u2019) was as large as with repeated exposures of the novel environment (N-N\u2019; n.s., p > 0.05), in contrast to CA1 (*p = 0.012, ****p < 0.0001, one-way ANOVA with Tukey\u2019s post hoc)", "statistics": [ { "about": "PFC, F-N' against N-N'", "test": "one-way ANOVA with Tukey's post hoc", "p": 0.05, "p_relation": ">" }, { "about": "CA1, F-N' against N-N' (one of the two measures, not named)", "test": "one-way ANOVA with Tukey's post hoc", "p": 0.012 }, { "about": "CA1, F-N' against N-N' (the other measure)", "test": "one-way ANOVA with Tukey's post hoc", "p": 0.0001, "p_relation": "<" } ], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "PFC F-N' does not differ from N-N' in rate overlap (reading a, map mean, p = 0.94; reading b, run rate, p = 0.82) or spatial correlation (p = 0.52). CA1 F-N' is below N-N' in spatial correlation (difference 0.524, p < 0.0001) but not in rate overlap (p = 0.32 and 0.45), so the CA1 part holds for one measure of two. The 4-cm and 2-cm widths give the same pattern (CA1 rate overlap p 0.12 to 0.39).", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig2d-pv-similarity", "statement": "PFC population-vector similarity across environments (F-N') is as high as across repeated exposures to the novel environment (N-N'), while CA1 PV similarity is lower across environments than within the novel environment, where the CA1 map is reinstated.", "source": "Figure 2C and 2D; Results, PFC spatial map generalizes, while CA1 completely remaps, across contexts, paragraphs 1 and 2", "quote": "the PFC spatial-representation similarity between two different environments was just as large as with repeated testing in the same environment", "statistics": [ { "about": "PFC, F-N' against N-N' (the n.s. pair)", "test": "Tukey's post hoc", "p": 0.41, "n": "5 rats" }, { "about": "CA1, F-N' against N-N'", "test": "Tukey's post hoc", "p": 0.01, "n": "5 rats" } ], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "Per animal, CA1 PV similarity 0.419 to 0.652 (N-N') and -0.138 to 0.239 (F-N'); PFC 0.727 to 0.952 (N-N') and 0.647 to 0.943 (F-N'). Repeated-measures ANOVA F(3,12) = 67.53, p = 8.8e-08 (the paper F = 83.24); Tukey CA1 F-N' below N-N' by 0.473, p = 0.0001 (the paper p = 0.01); PFC p = 0.99 (the paper 0.41). The 4-cm and 2-cm widths, the code's \"> 3 Hz\" rule and the peak taken on the native 2-cm maps agree, so the outcome does not rest on the width.", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times.", "Cells follow the Methods (at least 100 spikes) and the code (poolcells_acrossep.m, more than 100 spikes, CA1 mean rate below 7 Hz); the code's per-animal interneuron lists cannot be applied, as the files carry no cell numbers." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig2e-ca1-error-familiar", "statement": "Decoding with N' rate maps, CA1 position error is higher in F than in N, and in F no better than with cell-ID shuffled templates.", "source": "Figure 2E; Results, PFC spatial map generalizes, while CA1 completely remaps, across contexts, paragraph 4", "quote": "the error between actual and reconstructed positions in the familiar environment was significantly higher than that in the novel environment and no better than chance for CA1", "statistics": [ { "about": "CA1, F against N", "test": "Kruskal-Wallis test with Dunn's post hoc", "p": 0.0001, "p_relation": "<" }, { "about": "CA1 in F against its cell-ID shuffle (the n.s. comparison)", "test": "Wilcoxon paired test", "p": 0.14 } ], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "CA1 error is higher in F than in N (median 230 against 172 cm; Dunn p = 1.8e-140 over the four groups, p = 1.0e-139 for N and F alone), in every variant. In F the error is not below the shuffle but above it (median 230 against 226 cm, mean paired difference 8.8 cm); the Wilcoxon test finds the difference under shuffle reading (a), p = 2.4e-08, and (b), p = 0.042, so the claim of no difference fails by the paper's test, though the statement's \"no better than chance\" holds in direction. The widths agree on the outcome, 4 cm (a) p = 3.4e-10 ((b) p = 0.79), 2 cm (a) p = 1.8e-08 and (b) p = 0.0067 with F 3.5 cm below the shuffle. The decoder works, decoding N' with templates from the other half of its passes gives a median error of 12 cm with 64% of bins on the current trajectory (chance 25%). In N 43% of CA1 bins are decoded onto the current trajectory and 65% onto the current arm; errors are large because a decode onto the other trajectory of the same arm counts as far, as the criterion asks. Restricted to the current trajectory, CA1 errors are 15 cm (N) and 57 cm (F).", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times.", "Decoding windows of 200 ms every 150 ms (Methods, \"moving window with 50 ms overlap\"); the code's 50-ms step is a variant. The code's theta restriction cannot be applied (no theta times in the dandiset).", "Positions of N and F are put into N' coordinates by the ratio of trajectory lengths, as each epoch has its own linearization (an assumption).", "Dunn's test is Bonferroni-adjusted over all pairs, as GraphPad Prism (the paper's software) does; a rate floor of 1e-4 Hz keeps log rates finite (a standard choice)." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig2e-pfc-error-stable", "statement": "Decoding with N' rate maps, PFC position error is similar in N and F, and in both well above chance (below the cell-ID shuffles).", "source": "Figure 2E; Results, PFC spatial map generalizes, while CA1 completely remaps, across contexts, paragraph 4", "quote": "whereas PFC decoding was well above chance and similar across environments", "statistics": [ { "about": "PFC, N against F", "test": "Kruskal-Wallis test with Dunn's post hoc", "p": 0.99, "p_relation": ">" }, { "about": "a comparison with cell-ID shuffles (not named)", "test": "Wilcoxon paired test", "p": 0.0023 }, { "about": "comparisons with cell-ID shuffles (not named)", "test": "Wilcoxon paired test", "p": 0.0001, "p_relation": "<" } ], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "PFC error is below the shuffles in N (149 against 218 cm; Wilcoxon p = 1.3e-91 and 6.4e-67) and in F (186 against 223 cm; p = 7.8e-32 and 2.4e-35). It is not similar in N and F, F is higher (Dunn p = 7.0e-06 over four groups, p = 3.9e-06 for N and F alone), in every variant and width (p 6.2e-05 to 1.4e-17). Four of five rats have the higher median in F. With the decode restricted to the current trajectory, F is lower than N (23.5 against 28.5 cm), still different (p = 6.2e-05), so the sign of the difference depends on the decoder. PFC bins fall on the current trajectory in 42% (N) and 38% (F) of bins, 52% in cross-validated N'.", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times.", "Decoder as in fig2e-ca1-error-familiar." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig2f-switch-dynamics", "statement": "After the switch from N to F, CA1 decoding error rises within the first trial, while PFC decoding error stays stable.", "source": "Figure 2F; Results, PFC spatial map generalizes, while CA1 completely remaps, across contexts, paragraph 4", "quote": "Interestingly, the decoding error rapidly increased during the first trial after environment switching for CA1 but remained stable for PFC", "statistics": [ { "about": "CA1, trial -1 against trial 1", "test": "Kruskal-Wallis test with Dunn's post hoc", "p": 0.038 }, { "about": "PFC, trial -1 against trial 1", "test": "Kruskal-Wallis test with Dunn's post hoc", "p": 0.99, "p_relation": ">" } ], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "Reading (a), trial -1 the last N trial, Methods decoder at 8 cm, CA1 error rises from 182 to 223 cm between trial -1 and trial 1 (Dunn p = 0.0021), PFC does not differ (149 and 200 cm, p = 0.12). Reading (b), the code's first ten N trials labelled -10 to -1 and later N trials dropped, CA1 trial 1 is not above trial -1 (223 against 259 cm, p = 1), PFC p = 1. The criterion asks for both readings; the PFC part holds under both and the CA1 part under (a) only. The other widths give no more, 4 cm CA1 rises under (a) (p = 0.0079) but PFC rises too (p = 0.0004), 2 cm CA1 does not rise under either reading (p = 0.30 and 1) and PFC is stable (p = 0.076 and 1). The code's decoder (variant) finds no CA1 rise under (a) (p = 1), and under (b) both regions rise (p = 0.00027 and 0.0013). Calling the outcome partial at every width is a judgement. At 2 cm only the PFC claim of absence holds and CA1 rises under neither reading, which alone would be closer to not reproduced; the outcome is partial or worse at every width and reproduced at none, so the width rule does not block it.", "deviations": [ "Both readings of trial -1 in the criterion are run. Reading (a) is the last N trial; the legend's \"Trial-by-trial decoding error during environment switch (N to F)\" and the Results' \"the first trial after environment switching\" read so, and decoding_position_novelfamiliar_trial_batch.m means it (its comment says \"last 10 trials\") but its condition `ep == 1` never holds for epochs 2 and 4, so as written it takes the first ten, reading (b).", "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times.", "Trials are rows of the dandiset's trials table; Kruskal-Wallis over trials -10 to 10 with Dunn's test Bonferroni-adjusted over all 190 pairs (GraphPad's default for all pairs)." ], "assumed": false, "disagreed": true, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs3c-pfc-assimilation", "statement": "PFC PV similarity across environments grows with experience (F-N' higher than F-N), while CA1 does not change (F-N against F-N' not different) and stays below N-N'.", "source": "Figure S3C; Results, PFC spatial map generalizes, while CA1 completely remaps, across contexts, paragraph 2", "quote": "Furthermore, there was greater assimilation of PFC representations across environments with experience, while CA1 place cells maintained their independent representations", "statistics": [ { "about": "CA1, F-N against F-N'", "test": "one-way ANOVA with Tukey's post hoc", "p": 0.093 }, { "about": "PFC, F-N against F-N'", "test": "one-way ANOVA with Tukey's post hoc", "p": 0.0001, "p_relation": "<" }, { "about": "CA1, F-N' against N-N'", "test": "one-way ANOVA with Tukey's post hoc", "p": 0.0001, "p_relation": "<" }, { "about": "PFC, F-N' against N-N'", "test": "one-way ANOVA with Tukey's post hoc", "p": 0.0013 } ], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "Bins pooled over rats (8 cm), PFC F-N' 0.836 against F-N 0.774 (Tukey p < 0.0001); CA1 F-N 0.062 and F-N' 0.045 do not differ (p = 0.065; the paper p = 0.093); CA1 F-N' below N-N' 0.517 (p < 0.0001). PFC F-N' is below N-N' by 0.019 (p = 0.0001; the paper p = 0.0013). The 4-cm and 2-cm widths agree (CA1 F-N against F-N' p = 0.098 and 0.26).", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs3d-pfc-similarity-over-trials", "statement": "Over trials in the novel environment, PFC similarity to the familiar map rises from the first to the last trial, while CA1 similarity does not change.", "source": "Figure S3D", "quote": "PV similarity between novel and familiar environments over trials in the novel environment", "statistics": [ { "about": "CA1, first against last trial", "test": "one-way ANOVA with Tukey's post hoc", "p": 0.99, "p_relation": ">" }, { "about": "PFC, first against last trial", "test": "one-way ANOVA with Tukey's post hoc", "p": 0.0001, "p_relation": "<" } ], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "The criterion asks for each reading of the novel trials. PFC last trial above the first under every reading at 8 cm, N then N' 0.688 to 0.741 (Tukey p = 0.0018), N 0.684 to 0.757 (p < 0.0001), N' 0.738 to 0.782 (p = 0.0074). CA1 first and last do not differ under N then N' (0.053 and 0.048, p = 1.0) but CA1 falls under N (0.136 to 0.033, p = 0.0014; three rats, as JS34 and ZT2 have no N pass with 80% of CA1 cells active) and under N' (0.114 to 0.002, p < 0.0001), so the CA1 part holds under one reading of three. With each animal's own final kept pass as the last trial, CA1 gives p = 0.53 (N then N'), 0.0001 (N) and < 0.0001 (N'), PFC holds for N then N' and N (p = 0.0046 and 0.0015) but not N' (p = 1). Under N then N', all kept CA1 passes of JS21, JS34 and ZT2 are from N'. The 4-cm and 2-cm widths give the same outcome (CA1 falls under N and N' with at least one definition of the last trial, p <= 0.0098).", "deviations": [ "The paper does not say how trials are numbered across animals. The kept passes (at least 80% of tuned cells active, legend) of each animal are ranked in order; ranks 1 to K enter the test, K the smallest number of kept passes over the animals, and the last trial is (i) rank K or (ii) each animal's own final kept pass. Both are run.", "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times." ], "assumed": false, "disagreed": true, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs4-ramping-fraction", "statement": "Ramping cells make up 16.4% of CA1 cell samples (81 of 495) and 18.1% of PFC cell samples (115 of 636), a cell sample being one cell on one trajectory type.", "source": "Figure S4A and S4B legend", "quote": "ramping cells identified here (CA1 ... 81/495 = 16.4% cell samples; PFC ... 115/636 = 18.1% cell samples across all sessions", "statistics": [], "kind": "finding", "inferred": "", "outcome": "not_reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "Ramping cell samples are 3.0% of CA1 (49 of 1652) and 3.1% of PFC (67 of 2192) over the three sessions (reading a), and 3.7% (22 of 596) and 3.0% (22 of 728) in N' (reading b), against 16.4% (81 of 495) and 18.1% (115 of 636). The 4-cm and 2-cm widths, and denominators of tuned samples only, give at most 7.6%. The N' denominators are above the stated ones beyond tolerance; the tuned N' samples (291 and 484) are below them. Measured values are reading (a).", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times.", "The regression runs over the whole trajectory, as the claims file says; the Methods' \"over traversed positions to reward\" does not say which positions. Question for the paper's authors, which positions enter the regression and which cell samples form the denominator." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs4-ramping-generalization", "statement": "PFC ramping cells generalize their reward representation across environments, whereas CA1 ramping cells remap, and the generalization of ramping cells is similar to that of other cells in each region.", "source": "Results, PFC spatial map generalizes, while CA1 completely remaps, across contexts, paragraph 3; Figure S4A to S4C", "quote": "We found that PFC ramping cells generalized reward representations, but the CA1 representation remapped across environments ... In addition, the generalization for reward ramping cells was seen to a similar degree to other cells within regions", "statistics": [ { "about": "a comparison not named (n.s.)", "test": "Kruskal-Wallis test with Dunn's post hoc", "p": 0.05 }, { "about": "comparisons not named", "test": "Kruskal-Wallis test with Dunn's post hoc", "p": 0.0001, "p_relation": "<" } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "other", "notebook": "reproduce_tang_2023.ipynb", "notes": "Too few ramping cells under the claim's own rule. Ramping samples identified in N' are 4 to 7 per region and trajectory over all animals, and only one animal-trajectory population per region has 3 or more, so each ramping group is a single population of 101 bins. On it, CA1 ramping F-N' is below N-N' (Dunn p = 1.7e-22), PFC ramping does not change (p = 0.34), and ramping and other cells differ only for CA1 F-N' (p = 0.014; CA1 N-N' p = 1, PFC p = 1 and 0.11), which would make it partially reproduced. With the ramping samples pooled over animals into one population per trajectory, the CA1 ramping drop is not found (p = 1) and PFC ramping and other cells differ for F-N' (p = 0.016). The result rests on too few cells to decide, and the ramping selection itself did not reproduce (figs4-ramping-fraction).", "deviations": [ "A PV is computed only over at least 3 cells (an assumption; over 2 cells a correlation is always plus or minus 1).", "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs4d-reward-location-similarity", "statement": "PV similarity between the environments at the reward wells is the same as at other locations, in both regions.", "source": "Results, PFC spatial map generalizes, while CA1 completely remaps, across contexts, paragraph 3; Figure S4D", "quote": "Finally, in both regions, the similarity of population activity across different environments at reward wells was the same as other locations", "statistics": [ { "about": "PV similarity over locations, all tests", "test": "Grubbs' test for outliers", "p": 0.05, "p_relation": ">" } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "methods_underspecified", "notebook": "reproduce_tang_2023.ipynb", "notes": "At 8 cm Grubbs' test finds no outlier among the per-bin PV similarities, mean over rats, for F-N' and F-N in both regions, over all 404 bins (p = 1 in all four) or per trajectory (smallest p 0.40), and the well ends lie inside the range of the other bins (CA1 F-N' -0.100 to 0.224, PFC 0.723 to 0.901). 4 cm agrees (smallest p 0.116). At 2 cm the test finds an outlier among the PFC F-N' bins (all bins p = 0.0070, per trajectory p = 0.0012), so the claim would not reproduce. The outcome changes with the smoothing width, which the text does not fix, and the claims file's rule for that case (fig2d-pv-similarity) makes it methods_underspecified. Question for the paper's authors, the 1D width (see the findings).", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times.", "Grubbs' p-value is the usual Bonferroni bound on the largest studentized deviation (a standard choice)." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs5d-path-equivalence", "statement": "PFC cells show stronger path equivalence (similar firing at the same trajectory phase across trajectory types) than CA1 cells.", "source": "Figure S5D; Results, Distinct geometries of CA1 and PFC representations, paragraph 1", "quote": "Stronger path equivalence in PFC than CA1 (***p = 0.0004, ****p < 0.0001, F = 25.37, one-way ANOVA with Tukey\u2019s post hoc)", "statistics": [ { "about": "a PFC against CA1 pair (not named)", "test": "Tukey's post hoc", "statistic": "F = 25.37 (the one-way ANOVA; the legend gives no p for it)", "p": 0.0004 }, { "about": "PFC against CA1 pairs (not named)", "test": "Tukey's post hoc", "p": 0.0001, "p_relation": "<" } ], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "PFC R exceeds CA1 R in every session (N, PFC minus CA1 0.173, Tukey p = 0.0002; F and N', p < 0.0001), ANOVA F = 22.97 (the paper 25.37), p = 1.3e-21, over 99 CA1 and 171 PFC cells found in all three sessions. Medians CA1 -0.009 to 0.169, PFC 0.304 to 0.405. The 4-cm and 2-cm widths agree (largest p 0.002).", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times.", "Cells follow the Methods (at least 100 spikes) and the code (poolcells_acrossep.m, more than 100 spikes, CA1 mean rate below 7 Hz); the code's per-animal interneuron lists cannot be applied, as the files carry no cell numbers." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs5e-path-equivalence-stable", "statement": "The path-equivalent coefficient of PFC cells is strongly correlated between environments (r = 0.64), more so than that of CA1 cells (r = 0.21).", "source": "Figure S5E", "quote": "Path-equivalent coefficient is more strongly correlated between environments for PFC than CA1 (****p < 0.0001, Z = 4.25, comparing correlation coefficients; PFC, r = 0.64, p = 8.43e-21; CA1, r = 0.21, p = 0.035; Pearson correlation)", "statistics": [ { "about": "PFC against CA1 correlation coefficients", "test": "comparison of correlation coefficients", "statistic": "Z = 4.25", "p": 0.0001, "p_relation": "<" }, { "about": "PFC, between environments", "test": "Pearson correlation", "statistic": "r = 0.64", "p": 8.43e-21 }, { "about": "CA1, between environments", "test": "Pearson correlation", "statistic": "r = 0.21", "p": 0.035 } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "methods_underspecified", "notebook": "reproduce_tang_2023.ipynb", "notes": "F against N' at 8 cm, PFC r = 0.572 (p = 3e-16), CA1 r = 0.231 (p = 0.021), Fisher's z = 3.24, p = 0.0012; CA1 within tolerance and PFC 0.004 below it, partially reproduced at this width. At 4 cm both are within tolerance (0.613 and 0.201, z p = 6.8e-05), reproduced; at 2 cm 0.647 and 0.139, partially. F against N gives PFC 0.516 to 0.552 and CA1 0.095 to 0.153. The outcome changes with the width, and the claims file's rule for that case (fig2d-pv-similarity) makes it methods_underspecified.", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs5f-path-equivalence-vs-selectivity", "statement": "Path equivalence and task-sequence selectivity are uncorrelated across cells.", "source": "Figure S5F", "quote": "Path equivalence and task-sequence selectivity are two uncorrelated measures (n.s., all p\u2019s > 0.05; Pearson correlation)", "statistics": [ { "test": "Pearson correlation", "p": 0.05, "p_relation": ">" } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "methods_underspecified", "notebook": "reproduce_tang_2023.ipynb", "notes": "With the signed SI no correlation of R with SI(IN) or SI(OUT) reaches p < 0.05 in either region (8 cm, smallest p 0.26). With the absolute SI three of four do (CA1 SI(IN) r = -0.174, p = 0.025; PFC SI(IN) r = -0.241, p = 5.2e-06; PFC SI(OUT) r = -0.208, p = 8.7e-05). The criterion makes the claim methods_underspecified when only one reading finds a correlation. At 4 cm the same holds (signed none, smallest p 0.21; absolute four of four, p <= 0.037). At 2 cm both readings find one (signed PFC SI(OUT) r = 0.097, p = 0.047; absolute three of four), so there the claim would not reproduce; the outcome changes with the width, which also makes it methods_underspecified under the claims file's width rule. Question for the paper's authors, whether Figure S5F plots the signed or the absolute selectivity index.", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig3d-selectivity-index", "statement": "Inbound and outbound trajectory selectivity indices are negatively correlated in PFC (r = -0.5), reflecting task-sequence selectivity, and positively correlated in CA1 (r = 0.56), reflecting spatial coverage.", "source": "Figure 3D; Results, Distinct geometries of CA1 and PFC representations, paragraph 1", "quote": "As a result, the selectivity indices of IN (IN L vs. IN R ) and OUT (OUT L vs. OUT R ) trajectories exhibited a positive correlation for CA1 and an opposite, negative correlation for PFC", "statistics": [ { "about": "PFC, SI(IN) against SI(OUT)", "test": "Pearson correlation", "statistic": "r = -0.5", "p": 1.35e-16 }, { "about": "CA1, SI(IN) against SI(OUT)", "test": "Pearson correlation", "statistic": "r = 0.56", "p": 1.39e-11 } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "methods_underspecified", "notebook": "reproduce_tang_2023.ipynb", "notes": "PFC r < 0 under every reading and width, within tolerance in N' at 8 cm (r = -0.466, p = 7.0e-08); pooled -0.335, F -0.417. CA1 r > 0 but weaker than 0.56, at 8 cm pooled 0.284 (p = 0.00021), N' 0.280 (p = 0.040), F 0.127 (p = 0.41). At 8 and 4 cm reading (c) is not significant for CA1 (p = 0.41 and 0.085), so the partial clause fails (not reproduced); at 2 cm every reading has CA1 r 0.249 to 0.336 at p < 0.05 and PFC r < 0 (partially reproduced). The outcome changes with the width, and the claims file's rule for that case (fig2d-pv-similarity) makes it methods_underspecified. Per animal in N' (8 cm), CA1 r runs from -0.40 (JS17, 7 cells) to 0.89 (JS21, 8 cells).", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times.", "The code's \"> 3 Hz\" rule is a variant; it selects the same samples here." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig3e-selectivity-stable", "statement": "Trajectory selectivity is stable across environments for PFC cells (r = 0.49 between familiar and novel) but not for CA1 cells (r = 0.12, not significant).", "source": "Figure 3E; Results, Distinct geometries of CA1 and PFC representations, paragraph 1", "quote": "Finally, consistent with spatial remapping results, these properties were stable across environments for PFC neurons but not CA1", "statistics": [ { "about": "PFC, SI in F against SI in the novel environment", "test": "Pearson correlation", "statistic": "r = 0.49", "p": 2.77e-14 }, { "about": "CA1, SI in F against SI in the novel environment", "test": "Pearson correlation", "statistic": "r = 0.12", "p": 0.22 } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "methods_underspecified", "notebook": "reproduce_tang_2023.ipynb", "notes": "PFC selectivity correlates between F and the novel room at every width (F against N' r = 0.406 to 0.437, p <= 1.4e-10). At 8 and 4 cm CA1 correlates too (F against N' r = 0.318 and 0.231, p = 0.012 and 0.035; F against N p <= 0.0039), not reproduced; at 2 cm CA1 F against N' is not significant (r = 0.128, p = 0.20) and PFC 0.406 is below tolerance, partially reproduced. The outcome changes with the width, and the claims file's rule for that case (fig2d-pv-similarity) makes it methods_underspecified.", "deviations": [ "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Rate maps use run periods above 5 cm/s with occupancy above 20 ms (Methods, Spatial fields and linearization), but SWR times are not excluded, as the dandiset holds no ripple times." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "manifold-cell-counts", "statement": "The pseudo-populations pooled across animals for the manifold analyses hold 98 CA1 and 171 PFC cells.", "source": "STAR Methods, Characterization of neural manifold, paragraph 1; Figure 3F and 3G legend", "quote": "all spatially-tuned units that were active in both environments were pooled across animals to build a pseudo-population of neurons ( n = 98 CA1 and 171 PFC cells, respectively)", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "Reading (b), the code's (units included in N, F and N'), gives 99 CA1 and 171 PFC cells against 98 and 171. Reading (a), the Methods' (also a peak of at least 3 Hz in a session), gives 95 and 151, PFC below tolerance. Reading (b) is used for the manifolds. That the counts match supports the reading of units/electrodes as tetrode numbers.", "deviations": [ "Cells follow the Methods (at least 100 spikes) and the code (poolcells_acrossep.m, more than 100 spikes, CA1 mean rate below 7 Hz); the code's per-animal interneuron lists cannot be applied, as the files carry no cell numbers.", "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig3fg-manifold-shapes", "statement": "UMAP embeddings show that PFC population activity lies on a ring-shaped manifold, whereas the CA1 manifold has a double V shape resembling the W-maze.", "source": "Figure 3F and 3G; Results, Distinct geometries of CA1 and PFC representations, paragraph 2", "quote": "This unsupervised algorithm revealed that the population activity of PFC cells resides on a ring-shaped manifold, making it distinct from the CA1 neural manifold with a double \u201cV\u201d shape that topologically resembles the physical spatial structure of the W-maze", "statistics": [], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "The criterion takes the outcome from fig3h-inseq-outseq-distance and fig3i-phase-location-distance, both reproduced. The embeddings and the mean state per trajectory are shown; no measure of a ring or a double V is computed, and the plotted embeddings are made of one thread per pass, not a clean ring or double V.", "deviations": [ "UMAP is umap-learn with seed 0 (seeds 1 and 2 as variants), not the MATLAB toolbox with the paper's saved templates, which are not public (a standard choice named in the claims file). Numba is compiled for a generic x86-64 target, with the BLAS kernel and numpy's dispatch fixed, so that the embedding is the same on every machine; before this the embeddings differed between machines.", "Assumption. Empty 1-cm bins inside a pass of the single-trial maps are filled linearly, as cal_PVsimilarity.m fills session maps; the Methods say only that bins with spikes from fewer than 5 neurons are discarded. The maps without filling are run as a variant (8 and 2 cm).", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "Passes per trajectory are the smallest number over the animals from the first (Methods), 5, 6, 8 and 6 in N'. Maps at 1-cm bins (Methods) on 201 normalized bins (cal_UMAP.m).", "The Methods give no smoothing for the single-trial maps; they are smoothed as the session maps, at 8 cm, with 4 and 2 cm as variants (the width is open, as in fig2d-pv-similarity)." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig3h-inseq-outseq-distance", "statement": "In the PFC manifold, trajectories of the same task sequence (InSeq) are closer than trajectories out of sequence (OutSeq); in the CA1 manifold, OutSeq pairs with the same spatial coverage are closer than InSeq pairs.", "source": "Figure 3H; Results, Distinct geometries of CA1 and PFC representations, paragraph 3", "quote": "we found that the distance between two InSeq trajectories was significantly smaller than two OutSeq ones in the PFC manifold, whereas OutSeq pairs with same spatial coverages had smaller distances than InSeq pairs in the CA1 manifold", "statistics": [ { "test": "rank-sum test", "p": 0.0001, "p_relation": "<" } ], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "Seed 0, CA1 (by location) OutSeq 13.67 against InSeq 17.27, rank-sum p = 0.0005; PFC (by phase) InSeq 14.04 against OutSeq 17.51, p = 8.9e-06. With the main alignment both directions hold under seeds 1 and 2 (CA1 p = 1.4e-05 and 0.0026, PFC p = 0.00033 and 0.042); of the other alignments PFC by location is weaker under seed 2 (p = 0.069), and the rest give p < 0.01. The 4-cm and 2-cm widths agree (p <= 3.2e-11), and so do the maps without gap filling at 8 cm (CA1 p = 4.2e-27, PFC p = 0.022); without gap filling at 2 cm PFC InSeq and OutSeq do not differ (16.4 and 16.7, p = 0.34), so the outcome rests on the filling. The result moves with the embedding. In the verifier's round-2 copy, embedded on another machine before numba's CPU target was fixed, CA1 with the main alignment gave p = 0.096 under seed 1. Seed 0, the standard choice of the claims file, passed in every execution.", "deviations": [ "UMAP is umap-learn with seed 0 (seeds 1 and 2 as variants), not the MATLAB toolbox with the paper's saved templates, which are not public (a standard choice named in the claims file). Numba is compiled for a generic x86-64 target, with the BLAS kernel and numpy's dispatch fixed, so that the embedding is the same on every machine; before this the embeddings differed between machines.", "Assumption. Empty 1-cm bins inside a pass of the single-trial maps are filled linearly, as cal_PVsimilarity.m fills session maps; the Methods say only that bins with spikes from fewer than 5 neurons are discarded. The maps without filling are run as a variant (8 and 2 cm).", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "The Methods give no smoothing for the single-trial maps; they are smoothed as the session maps, at 8 cm, with 4 and 2 cm as variants (the width is open, as in fig2d-pv-similarity)." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig3i-phase-location-distance", "statement": "PFC neural states at the same trajectory phase are closer than those at the same spatial location, while CA1 shows the opposite.", "source": "Figure 3I; Results, Distinct geometries of CA1 and PFC representations, paragraph 3", "quote": "As a result, the distances between PFC neural states for the same trajectory phases were significantly smaller than those for the same spatial locations, while the opposite is true for CA1", "statistics": [ { "test": "rank-sum test", "p": 0.0001, "p_relation": "<" } ], "kind": "finding", "inferred": "", "outcome": "reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "Methods reading, seed 0, PFC same-phase 14.08 against same-location 17.72 (p near 0); CA1 location 14.33 against phase 17.45 (p near 0). Seeds 1 and 2, the 4-cm and 2-cm widths and the maps without gap filling agree (p <= 9.2e-43). The code's variant (both from same-side pairs) fails under seed 0 (CA1 reversed, p = 2.4e-09; PFC p = 0.68) and for CA1 under seed 1 (reversed, p = 0.03), and holds under seed 2.", "deviations": [ "UMAP is umap-learn with seed 0 (seeds 1 and 2 as variants), not the MATLAB toolbox with the paper's saved templates, which are not public (a standard choice named in the claims file). Numba is compiled for a generic x86-64 target, with the BLAS kernel and numpy's dispatch fixed, so that the embedding is the same on every machine; before this the embeddings differed between machines.", "Assumption. Empty 1-cm bins inside a pass of the single-trial maps are filled linearly, as cal_PVsimilarity.m fills session maps; the Methods say only that bins with spikes from fewer than 5 neurons are discarded. The maps without filling are run as a variant (8 and 2 cm).", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "The Methods give no smoothing for the single-trial maps; they are smoothed as the session maps, at 8 cm, with 4 and 2 cm as variants (the width is open, as in fig2d-pv-similarity)." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig3j-familiar-novel-manifold", "statement": "The PFC manifold is preserved across environments (F states lie closer to N' states than cell-ID shuffles do), while the distortion of the CA1 manifold is similar to that of random shuffles.", "source": "Figure 3J; Results, Distinct geometries of CA1 and PFC representations, paragraph 3", "quote": "We found that the geometric structure of the PFC manifold was largely preserved across environments, whereas the distortion of the CA1 manifold was similar to that from random shuffles", "statistics": [ { "about": "CA1, F-N' against shuffle (the n.s. test of the panels)", "test": "rank-sum test", "p": 0.16 }, { "about": "PFC, F-N' against shuffle", "test": "rank-sum test", "p": 0.0001, "p_relation": "<" } ], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "PFC F states lie closer to N' states than the shuffle (11.86 against 15.28, 22.4% closer, rank-sum p near 0). CA1 F states are also closer than the shuffle, by less (13.44 against 14.05, 4.4% closer, p = 1.4e-30), so the claim that CA1 does not differ from the shuffle fails by the paper's test; seeds 1 and 2 give CA1 6.8% and 4.1% and PFC 31.5% and 11.7% closer. CA1 is closer than the shuffle in every width and gap-filling variant (1.9% to 10.8%, p <= 4.6e-19), PFC in all but one (7.8% to 60.2%). Without gap filling at 2 cm PFC is not closer than the shuffle (0.8% farther, p = 0.029), so the PFC part rests on the filling. In the verifier's round-2 copy, embedded on another machine before numba's CPU target was fixed, that variant gave PFC 4.9% closer (p = 6.2e-12).", "deviations": [ "UMAP is umap-learn with seed 0 (seeds 1 and 2 as variants), not the MATLAB toolbox with the paper's saved templates, which are not public (a standard choice named in the claims file). Numba is compiled for a generic x86-64 target, with the BLAS kernel and numpy's dispatch fixed, so that the embedding is the same on every machine; before this the embeddings differed between machines.", "Assumption. Empty 1-cm bins inside a pass of the single-trial maps are filled linearly, as cal_PVsimilarity.m fills session maps; the Methods say only that bins with spikes from fewer than 5 neurons are discarded. The maps without filling are run as a variant (8 and 2 cm).", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "The Methods give no smoothing for the single-trial maps; they are smoothed as the session maps, at 8 cm, with 4 and 2 cm as variants (the width is open, as in fig2d-pv-similarity)." ], "assumed": true, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs6-high-dimensional-distances", "statement": "The distance relationships of Figure 3H to 3J hold in the original high-dimensional neural space.", "source": "Figure S6; Results, Distinct geometries of CA1 and PFC representations, paragraph 3", "quote": "In all the cases above, the distance relationship of neural states, which determines the representational geometry, was preserved between the low-dimensional manifold space and the original high-dimensional neural space", "statistics": [ { "about": "the n.s. comparison (not named)", "test": "rank-sum test", "p": 0.75 }, { "about": "the other comparisons", "test": "rank-sum test", "p": 0.0001, "p_relation": "<" } ], "kind": "finding", "inferred": "", "outcome": "partially_reproduced", "blocked_by": null, "notebook": "reproduce_tang_2023.ipynb", "notes": "In the neural space 3H holds (CA1 OutSeq 45.04 against InSeq 52.15, p = 2.5e-12; PFC InSeq 64.27 against OutSeq 70.61, p = 4.8e-12) and 3I holds (PFC phase 65.47 against location 86.09; CA1 location 45.02 against phase 48.37, p = 9.3e-129). In 3J PFC F-N' is 56.6% below the shuffle, but CA1 is also below it (9.2%, p = 7.5e-78), so two of the three relationships hold. Every width and gap-filling variant gives the same directions (CA1 3J 7.3% to 9.2% closer, p <= 2.7e-21).", "deviations": [ "Assumption. Empty 1-cm bins inside a pass of the single-trial maps are filled linearly, as cal_PVsimilarity.m fills session maps; the Methods say only that bins with spikes from fewer than 5 neurons are discarded. The maps without filling are run as a variant (8 and 2 cm).", "Linear positions are projections onto W-shaped paths built from the well positions and the run positions (nfn.build_skeleton), as the Methods' \"pre-defined idealized paths along the track\" are not in the dandiset (an assumption; paths are 180 to 220 cm, median 5 cm from the run positions).", "The 1D smoothing width of the linearized maps is garbled in the paper's text (Methods, Spatial fields and linearization, \"s = \u03c3 cm\"). 8, 4 and 2 cm are run; 8 cm is reading (a) of the claims file (fig2d-pv-similarity), the \u03c3 given for the 2D maps in the sentence before. As the claims file's rule for fig2d-pv-similarity has it, the outcome is that of 8 cm when the widths agree, and methods underspecified when they do not.", "The Methods give no smoothing for the single-trial maps; they are smoothed as the session maps, at 8 cm, with 4 and 2 cm as variants (the width is open, as in fig2d-pv-similarity)." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig4d-pfc-ccgp", "statement": "In PFC, the cross-condition generalization performance (CCGP) for trajectory phase, task sequence and environment is high and above trial-label shuffles for every decoder.", "source": "Figure 4D; Results, Distinct abstraction in CA1 and PFC, paragraph 2", "quote": "We found that the CCGP for all three task variables showed high accuracy for PFC", "statistics": [ { "test": "trial-label permutation test", "p": 0.001, "p_relation": "<", "n": "1,000 shuffles" } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "methods_underspecified", "notebook": "reproduce_tang_2023.ipynb", "notes": "With 8-cm single-trial maps 10 of the 16 PFC decoders have p < 0.001 by the code's strict count, but the task- sequence decoder tested on the inbound pair at the late segment predicts one class for every test point (CCGP 0.500, every shuffle ties, p = 1 counting ties). Counting ties, 9 of 16 have p < 0.001; the phase decoders tested on IN_R, OUT_R and OUT_L (CCGP 0.500, p >= 0.54 counting ties), that sequence decoder and three environment phase decoders (0.560, 0.560 and 0.541; p = 0.004, 0.004 and 0.020) fail, so the claim would not reproduce at 8 cm. Above every one of 100 shuffles are 14 of 16 decoders with 4-cm maps and 15 of 16 with 2-cm maps, and 4 of 16 without gap filling at 8 and at 2 cm; in this run no width gives all 16. In the verifier's round-2 copy, embedded on another machine before numba's CPU target was fixed, all 16 were above every one of 200 shuffles at 2 cm, so at the narrow widths the result sits at the edge of the criterion. The single-trial maps' smoothing is not in the Methods and the session width is garbled; at 8 cm the claim fails clearly and at 2 cm it can pass, and the claims file's rule for the width (fig2d-pv-similarity) makes the claim methods_underspecified.", "deviations": [ "UMAP is umap-learn with seed 0 (seeds 1 and 2 as variants), not the MATLAB toolbox with the paper's saved templates, which are not public (a standard choice named in the claims file). Numba is compiled for a generic x86-64 target, with the BLAS kernel and numpy's dispatch fixed, so that the embedding is the same on every machine; before this the embeddings differed between machines.", "Assumption. Empty 1-cm bins inside a pass of the single-trial maps are filled linearly, as cal_PVsimilarity.m fills session maps; the Methods say only that bins with spikes from fewer than 5 neurons are discarded. The maps without filling are run as a variant (8 and 2 cm).", "The task-sequence decoders follow CCGP_taskSeq.m, whose test pairs take the two clusters at different segments (20-30% for one trajectory, 70-80% for the other); taking them at the same segment is a variant and gives CCGP within 0.003 of these.", "Linear C-SVM through scikit-learn's libsvm (C = 1), balanced accuracy, 1,000 label shuffles (decoding_CCGP_linearSVM_simple.m).", "The Methods give no smoothing for the single-trial maps; they are smoothed as the session maps, at 8 cm, with 4 and 2 cm as variants (the width is open, as in fig2d-pv-similarity).", "A decoder that predicts one class for every test point has balanced accuracy 0.5 and ties with every shuffle; the code's p (fraction of shuffles strictly above) gives it 0. Ties are counted here (p_ge), so such decoders are not significant; the strict p is shown too." ], "assumed": false, "disagreed": true, "data": "dataset", "where": null, "proposed": null }, { "claim": "fig4d-ca1-ccgp-chance", "statement": "In CA1, CCGP for trajectory phase, task sequence and environment is at or below chance by the trial-label permutation test.", "source": "Figure 4D; Results, Distinct abstraction in CA1 and PFC, paragraph 2", "quote": "Note that all dichotomies can be accurately decoded in PFC (****p < 0.001) but not in CA1 (n.s., p > 0.05 , trial-label permutation test)", "statistics": [ { "test": "trial-label permutation test", "p": 0.05, "p_relation": ">", "n": "1,000 shuffles" } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "methods_underspecified", "notebook": "reproduce_tang_2023.ipynb", "notes": "Blocked because it depends on fig4d-pfc-ccgp, which is blocked. The claims file lists fig4d-pfc-ccgp under depends_on, and the protocol (Author, step 1.5) blocks a claim that depends on a blocked claim for the same reason. Its own result does not change with the width or the filling. At 8 cm 4 of the 16 CA1 decoders predict one class for every test point (the four phase decoders, CCGP 0.500), and 8 have p < 0.05 counting ties, two sequence decoders (CCGP 0.547 and 0.578) and six environment decoders (0.568 to 0.780), against 95th percentiles of 0.500 to 0.544; the variants give 4 to 11 (100 shuffles). Tested alone, it would not reproduce.", "deviations": [ "UMAP is umap-learn with seed 0 (seeds 1 and 2 as variants), not the MATLAB toolbox with the paper's saved templates, which are not public (a standard choice named in the claims file). Numba is compiled for a generic x86-64 target, with the BLAS kernel and numpy's dispatch fixed, so that the embedding is the same on every machine; before this the embeddings differed between machines.", "Assumption. Empty 1-cm bins inside a pass of the single-trial maps are filled linearly, as cal_PVsimilarity.m fills session maps; the Methods say only that bins with spikes from fewer than 5 neurons are discarded. The maps without filling are run as a variant (8 and 2 cm).", "Decoders as in fig4d-pfc-ccgp.", "The Methods give no smoothing for the single-trial maps; they are smoothed as the session maps, at 8 cm, with 4 and 2 cm as variants (the width is open, as in fig2d-pv-similarity).", "Ties with the shuffles are counted (p_ge), as in fig4d-pfc-ccgp." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs7b-geometric-null", "statement": "Against a geometric random model that keeps within-condition noise but destroys the geometry, PFC CCGP is significant while CA1 CCGP is not.", "source": "Figure S7B; Results, Distinct abstraction in CA1 and PFC, paragraph 2", "quote": "CCGP for trajectory phases (circles), task sequences (triangles), and environments (diamonds; horizontal lines, median) against the null-distributions of the geometric random model (red horizontal line shown the 95th percentile; n.s., p > 0.05, **p = 0.0026, ****p <0.0001)", "statistics": [ { "about": "CA1 (n.s.)", "test": "geometric random model", "p": 0.05, "p_relation": ">" }, { "about": "a comparison not named", "test": "geometric random model", "p": 0.0026 }, { "about": "comparisons not named", "test": "geometric random model", "p": 0.0001, "p_relation": "<" } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "methods_underspecified", "notebook": "reproduce_tang_2023.ipynb", "notes": "Blocked because it depends on fig4d-pfc-ccgp, which is blocked. At 8 cm the PFC CCGP (mean of its decoders) is 0.549 for phase, 0.745 for sequence and 0.681 for environment. Under reading (a), permuted UMAP coordinates, none exceeds the 95th percentile of its null (p 0.08 to 0.59). Under reading (b), permuted neuron axes embedded with the N' fit, sequence does (0.745 against 0.684, no null draw above it) and phase and environment do not (p = 0.76 and 0.48). No CA1 variable exceeds its null (p 0.21 to 1.00 in reading a, 0.26 to 0.56 in reading b). The null was not run at the other widths. Tested alone, it would not reproduce.", "deviations": [ "The CCGP of a variable is the mean over its decoders, compared with the same mean under the null; the legend plots each decoder with the median. Both readings of the space (claims file) are run; the code for the model is not public.", "UMAP is umap-learn with seed 0 (seeds 1 and 2 as variants), not the MATLAB toolbox with the paper's saved templates, which are not public (a standard choice named in the claims file). Numba is compiled for a generic x86-64 target, with the BLAS kernel and numpy's dispatch fixed, so that the embedding is the same on every machine; before this the embeddings differed between machines.", "Assumption. Empty 1-cm bins inside a pass of the single-trial maps are filled linearly, as cal_PVsimilarity.m fills session maps; the Methods say only that bins with spikes from fewer than 5 neurons are discarded. The maps without filling are run as a variant (8 and 2 cm).", "The Methods give no smoothing for the single-trial maps; they are smoothed as the session maps, at 8 cm, with 4 and 2 cm as variants (the width is open, as in fig2d-pv-similarity)." ], "assumed": false, "disagreed": false, "data": "dataset", "where": null, "proposed": null }, { "claim": "figs7d-dichotomies-decodable", "statement": "With traditional cross-validated decoders, most of the 35 dichotomies of the 8 clusters are decoded accurately in both CA1 and PFC, with similar accuracy in the two regions.", "source": "Figure S7D; Results, Distinct abstraction in CA1 and PFC, paragraph 2", "quote": "we found that most of the dichotomies were accurately separated in both CA1 and PFC, suggesting that both regions contain a similar amount of decodable variables", "statistics": [ { "about": "PFC against CA1 decoding accuracy over 35 dichotomies", "test": "rank-sum test", "p": 0.17, "n": "35 dichotomies per region" } ], "kind": "finding", "inferred": "", "outcome": "blocked", "blocked_by": "methods_underspecified", "notebook": "reproduce_tang_2023.ipynb", "notes": "Blocked because it depends on fig4d-pfc-ccgp, which is blocked. At 8 cm most dichotomies are decoded above their 50 shuffles, CA1 35 of 35 (median accuracy 0.830), PFC 34 of 35 (median 0.631), the same counting ties. The accuracies differ, CA1 above PFC (rank-sum p = 8.6e-09, median difference 0.199), so the similarity part would fail. The other widths were not run for this test.", "deviations": [ "The label-permutation test uses 50 shuffles per dichotomy in place of the paper's 1,000, for time (100 took about 3 minutes per region here, and the notebook must run within 30 minutes); p is resolved to 0.02, enough to decide p < 0.05 (at most 2 of 50 shuffles above).", "UMAP is umap-learn with seed 0 (seeds 1 and 2 as variants), not the MATLAB toolbox with the paper's saved templates, which are not public (a standard choice named in the claims file). Numba is compiled for a generic x86-64 target, with the BLAS kernel and numpy's dispatch fixed, so that the embedding is the same on every machine; before this the embeddings differed between machines.", "Assumption. Empty 1-cm bins inside a pass of the single-trial maps are filled linearly, as cal_PVsimilarity.m fills session maps; the Methods say only that bins with spikes from fewer than 5 neurons are discarded. The maps without filling are run as a variant (8 and 2 cm).", "Accuracy is the fraction correct over the 4 folds, as libsvm's -v 4 reports; folds are stratified by class (scikit-learn).", "The Methods give no smoothing for the single-trial maps; they are smoothed as the session maps, at 8 cm, with 4 and 2 cm as variants (the width is open, as in fig2d-pv-similarity)." ], "assumed": false, "disagreed": true, "data": "dataset", "where": null, "proposed": null } ], "findings": [ { "kind": "format_error", "statement": "The units table's `electrodes` column holds tetrode numbers, not row indices of the electrodes table as the NWB format requires, its values run up to 64 in ZT2, whose table has 128 rows (two per tetrode), and up to 30 to 32 in the others. Read as row indices, up to 32 units per animal are put in the wrong region; read as tetrode numbers, the counts match the paper's (98 CA1 and 171 PFC cells against 99 and 171 here).", "found": null, "status": null }, { "kind": "metadata_insufficient", "statement": "Units carry no cell number within a tetrode, no waveform and no putative-interneuron label, so the paper's interneuron exclusion (spike width and rate in the Methods, per-animal (tetrode, cell) lists in its code) cannot be applied. Units clustered only in later epochs are appended to the table with no spike in N (1 to 19 per animal), without a note.", "found": null, "status": null }, { "kind": "other", "statement": "In every file `acquisition/ElectricalSeries` is a byte-identical second copy of the 1.5 kHz LFP in `processing/ecephys/LFP`, stored at a different offset; there is no 30 kHz spike-band recording, and the copy roughly doubles the file sizes.", "found": null, "status": null }, { "kind": "metadata_error", "statement": "The position's reference frame is given as \"center well\", but positions are camera coordinates in which the center well lies at different places in each epoch (for example AM2 (32, 52) cm in N and (73, 86) cm in F). The five files' session start times fall within 16 minutes on 2023-03-14, two days before the first published version, which is the time of conversion, not of recording.", "found": null, "status": null }, { "kind": "metadata_insufficient", "statement": "The files hold no linearized positions, idealized track paths or trajectory labels per position sample, no sharp-wave ripple or theta times, and the epochs carry no names (N, F and N' are known only from the dandiset description's order). The paper's rate maps, decoding and manifolds use all of these; here linear positions had to be rebuilt from the 2D positions and ripple times could not be excluded.", "found": null, "status": null }, { "kind": "metadata_error", "statement": "In AM2 the electrode groups Tetrode26 to Tetrode31 carry the labels tetrode27 to tetrode32, so group names and labels disagree; electrode descriptions are placeholders (\"my description\", \"no description\").", "found": null, "status": null }, { "kind": "data_missing", "statement": "The de novo learning sessions of the five rats and the behaviour of the three F-N-F'-N' rats (Figures 1B, 1C, S1B, S1C) are not in the dandiset; four claims are blocked and one is tested on the N-F-N' day only.", "found": null, "status": null }, { "kind": "code_mismatch", "statement": "decoding_position_novelfamiliar_trial_batch.m (Figure 2F) takes the last ten trials of the pre-switch epoch only `if ep == 1`, but runs on epochs 2 and 4, so it always takes the first ten N trials, against its own comment (\"last 10 trials\") and the legend's switch from N to F. With the last N trials CA1 error rises at the switch (Dunn p = 0.0021); with the first ten it does not (p = 1).", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the authors. What is the width of the 1D Gaussian that smooths the linearized rate maps? The text reads \"s = \u03c3 cm\". Results were run at 8, 4 and 2 cm; figs4d, figs5e, fig3d and fig3e change outcome with the width and are blocked for it.", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the authors. How were the idealized paths for linearization defined, and are the linearized positions available? They are not in the dandiset; here W-shaped paths were built from the well and run positions.", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the authors. In Figure 2F, are trials -10 to -1 the last ten N trials (as the legend and the code's comment suggest) or the first ten (as the code runs)? CA1 error rises at the switch with the last ten (p = 0.0021) but not with the first ten (p = 1).", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the authors. In Figure S3D, which novel trials are shown (N, N' or both), how are trials numbered across animals with different numbers of passes, and which is the last trial? CA1 similarity falls over trials of N or of N' alone but not over N then N'.", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the authors. Does Figure S5F use the signed or the absolute selectivity index? Path equivalence is uncorrelated with the signed SI but negatively correlated with its absolute value in PFC (p < 0.0001).", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the authors. For the ramping cells of Figure S4, which positions enter the regression and which cell samples form the denominator (495 and 636)? The rule as written gives 3 to 8% of cell samples here, against 16.4 and 18.1%.", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the authors. For the geometric random model of Figure S7, in which space are the axes permuted (the 3 UMAP coordinates or the neurons), and what happens to the cluster centres (the legend says rotation-translation)? Its code is not in the repository.", "found": null, "status": null }, { "kind": "methods_unavailable", "statement": "Question for the authors. How were the single-trial rate maps behind the UMAP manifolds smoothed, and were 1-cm bins that hold no position sample within a pass filled or left empty? The Methods say only that bins with spikes from fewer than 5 neurons were discarded. The PFC CCGP of Figure 4D fails clearly with 8-cm smoothing (9 of 16 decoders) and comes within one decoder of passing, or passes, with 2-cm smoothing (15 or 16 of 16, depending on the embedding), and without filling the PFC parts of Figure 3H and 3J fail at 2 cm.", "found": null, "status": null } ], "reconciliation": null, "verifications": [ { "date": "2026-10-30", "kind": "agent", "by": "Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook)", "reviewed": "The handover of the first run, the executed notebook with the conclusions and summary removed, the helper module and the claims file (transcripts/000447/verifier-1-report.md).", "agreed": 28, "of": 32, "disagreements": [ { "claim": "fig2f-switch-dynamics", "outcome": "partially_reproduced", "reason": "Cell 34, reading (a): CA1 trial \u22121 vs 1 p = 0.002 (rise); PFC p = 0.12 (n.s.). Reading (b) cannot be judged because of a labelling bug (B1). The code-decoder variant fails under both readings." }, { "claim": "figs3d-pfc-similarity-over-trials", "outcome": "partially_reproduced", "reason": "Cell 26: PFC rises from first to last trial under all three readings. CA1 stays flat only under reading (a), and that test is cut at K = 9 (B3). CA1 falls under (b) (p = 0.0014) and (c) (p < 1e-4)." }, { "claim": "fig4d-pfc-ccgp", "outcome": "not_reproduced", "reason": "Cell 67: 12 of 16 PFC decoders at p < 0.001; 2 of those 12 are artefacts (B2), so 10 of 16 are real." }, { "claim": "figs7d-dichotomies-decodable", "outcome": "not_reproduced", "reason": "Cell 72: \"most decodable\" holds (CA1 35/35, PFC 32/35). The regions differ: rank-sum p = 1.2e-9, CA1 higher." } ] }, { "date": "2026-10-30", "kind": "agent", "by": "Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook)", "reviewed": "The full executed notebook of the first revision, with conclusions and summary, the attempt record, the helper module and the claims file (transcripts/000447/verifier-2-report.md).", "agreed": 32, "of": 32, "disagreements": [] } ] } ], "entries": [ { "name": "000447-tang-2023", "type": "reproduction", "title": "Hippocampal and prefrontal maps across a novel and a familiar W-maze (Tang, Shin and Jadhav 2023)", "description": "Tests the findings of Tang, Shin and Jadhav (2023) on dandiset 000447, which holds CA1 and prefrontal units, position and trial tables of five rats running a W-maze alternation task in a novel, a familiar and again the novel room. It compares spatial maps across environments, Bayesian decoding with novel-room templates, selectivity and path equivalence of single cells, UMAP manifolds of the pooled populations, and cross-condition generalization. The de novo learning sessions and the second group of rats behind Figure 1 are not in the dandiset.", "datasets": [ "dandi:000447" ], "papers": [ "10.1016/j.celrep.2023.112246" ], "derived_from": [], "notebooks": [ { "path": "reproduce_tang_2023.ipynb", "title": "Reproducing Tang, Shin and Jadhav (2023) from dandiset 000447", "executed": "notebooks/000447-tang-2023/reproduce_tang_2023.html" } ] } ] }