Internally organized mechanisms of the head direction sense (DANDI:000056)

Claims

From the paper Internally organized mechanisms of the head direction sense

Reproducing Peyrache et al. (2015) from dandiset 000056 8

Notebook reproduce_peyrache_2015.ipynb

Partially reproducedMethods, Head direction cells, paragraph 2; Results, first paragraph
Of the recorded cells, 242 of 720 in the anterior thalamus and 111 of 357 in the PoS were classified as HD cells, 353 of 1077 in all.
242 out of 720 recorded cells (41% on average) in the anterior thalamus, and 111 out of 357 (31%) cells in the PoS were classified as HD cells.
Methods, Head direction cells, paragraph 2; Results, first paragraph
By the Methods' criteria 378 units are HD cells in the 38 sessions with position, within 353 plus or minus 35.3. They are 0.218 of the 1731 units with spikes in the open-field epoch (reading a) and 0.205 of all 1848 unit rows of those sessions (reading b), both outside 0.328 plus or minus 0.033. The files hold 1959 unit rows in 40 sessions against the paper's 1077 recorded cells in 42 sessions. By mouse the fraction is 0.108 to 0.386; in the mice that appear to be thalamus only it is 0.133 (133 of 1000) against the paper's thalamic 0.336, and in the mice that appear dual-site 0.335 (245 of 731). With a concentration threshold of 0.8 or 1.2 the count is 424 or 357. None of the 34 units with unsorted spike times in these sessions is an HD cell. The count passes only with Mouse12-120808 as deposited, where position and spikes do not share a clock and no unit passes the criteria. With position realigned by the best shift of each quarter of the epoch, that session has 13 HD cells (12 with one constant shift of -2.6 s), and the count is 391 (390), above the upper bound of 388.3, so the match of the count is fragile. The unit rows cannot be brought down to the paper's 1077 cells from the files (1826 rows fire at 0.05 Hz or more and 1526 at 0.5 Hz or more; unit ids are 0 to N-1 and carry no shank or cluster number). The agreement of the count may therefore also be coincidental. The split into 242 thalamic and 111 PoS HD cells cannot be tested.
How the test differed from the paper’s method
- Regions pooled, as the units carry no region.
- The von Mises fit (maximum likelihood fit to the tuning curve as a distribution of angles) and the Rayleigh test (resultant length of the tuning curve with the number of waking spikes) are choices made here; the Methods do not specify them.
- The open-field epoch is found from position, as the files do not mark it.
- Mouse17-130201 and Mouse32-140820 have no processed position and are left out.
- Mouse12-120808 is counted as deposited, with position and spikes on different clocks; the realigned session is a variant.
Partially reproducedResults, second paragraph; Figure 2e,f; Abstract ("accelerated tenfold")
The angular velocity of the internal HD signal, estimated from the temporal profile of pairwise cross-correlograms, is about tenfold faster during SWS than during waking.
The angular velocity of the internal HD signal, either estimated by a Bayesian decoder or from the temporal profile of pairwise cross-correlograms ... was approximately tenfold faster during SWS than during waking (Fig. 2e,f)
Results, second paragraph; Figure 2e,f; Abstract ("accelerated tenfold")
Over 1509 pairs (64 of 1573 dropped for a fit at the edge of the speed range), the median ratio of SWS to waking angular velocity is 8.33 (quartiles 5.91 to 11.31), below 9 to 11; SWS is faster in 1420 pairs (sign test p = 4.1e-309). Median speeds are 36.9 deg/s in waking, 33.5 in REM and 324.4 in SWS. By mouse the ratio is 7.89 to 13.52. The Bayesian decoder variant gives a median of 11.53 over 28 sessions. Under reading (b) the median is 8.02 over 2129 pairs; with pairs whose fits reach R2 of 0.25 or 0.5 in every state, 8.46 (1063 pairs) and 8.56 (748 pairs). Median R2 is 0.84 in waking, 0.55 in REM and 0.78 in SWS. The waking estimates are 0.60 times the measured mean angular speed of the head (median over sessions) and correlate with it at r = 0.20 across 33 sessions. On Poisson surrogates driven by the measured head direction and the waking tuning curves, the estimator returns 0.37 of the true mean speed (median 37.0 deg/s, as on the data, against a true 99.9), and for the same trajectory played ten times faster it gives a median ratio of 4.99 (quartiles 2.76 to 7.02), not 10. Scaled by that compression the data's 8.33 becomes 16.69, so the data are consistent with a speed-up below, at or above tenfold. The estimator shows that SWS is faster but does not determine the fold; how far its bias in SWS matches that of the surrogates is not known.
How the test differed from the paper’s method
- The paper does not say how the expected and observed cross-correlograms are put on one scale; a free scale and offset are fitted here.
- The lag range (4 s) and speed grid are chosen here.
- The expectation over the chi-squared speed distribution is computed exactly from its characteristic function.
- Pairs within and across structures are pooled; the paper shows ADn-ADn and PoS-PoS pairs.
ReproducedFigure 2a (ADn n = 215, PoS n = 62 cells)
The firing rates of HD neurons are strongly correlated between waking and SWS and between waking and REM sleep.
First, the firing rates of HD neurons remained strongly correlated across brain states (Fig. 2a).
Figure 2a (ADn n = 215, PoS n = 62 cells)
Over 378 HD cells in 37 sessions with sleep states, Pearson r of waking against SWS rates is 0.838 (p = 3.7e-101) and against REM rates 0.891 (p = 4.5e-131); with log rates 0.806 and 0.883. By mouse r is 0.718 to 0.889 for SWS and 0.866 to 0.942 for REM. The cell count (378 against the paper's 277) does not decide the outcome; the paper does not say which HD cells were in its figure.
How the test differed from the paper’s method
- ADn and PoS cells pooled; the paper plots them separately.
ReproducedResults, second paragraph; Figure 2b (ADn pairs n = 970, PoS pairs n = 92)
The pairwise correlations between HD neurons, within and across structures, are similar across brain states.
Second, the pairwise correlations between HD neuron pairs both within and across structures were robustly similar across states (Fig. 2b)
Results, second paragraph; Figure 2b (ADn pairs n = 970, PoS pairs n = 92)
Reading (a), 1573 pairs in 33 sessions, waking against SWS pair correlations r = 0.938 and against REM r = 0.963. Reading (b), 2246 pairs, r = 0.925 and 0.952. Every mouse gives r of at least 0.913. 888 of 2461 pairs fail reading (a) for low rates.
How the test differed from the paper’s method
- Pairs within and across structures are pooled, and pairs on the same shank cannot be removed.
- Bin sizes are those of Figure 3a; Figure 2b does not state them.
ReproducedResults, third paragraph; Figure 3a
The zero-lag correlation between HD cell pairs falls with the offset of their preferred directions, positive below 60 degrees, negative above it and near zero at about 60 degrees, and this holds in waking, REM and SWS.
Pairs of neurons with <60° offset showed positive correlations, pairs with >60° offset showed negative correlations, and pairs with approximately 60° offset were uncorrelated (Fig. 3a). The relationship between the 0-lag correlation and the preferred angular direction was preserved across brain states
Results, third paragraph; Figure 3a
What the paper reports
- Pearson's test, p < 10⁻⁶ (n = 907 ADn-ADn, 512 ADn-PoS and 92 PoS-PoS pairs)
The mean zero-lag correlation, in 10 degree bins of the offset of preferred directions, crosses zero at 54.2 degrees in waking, 56.9 in REM and 55.9 in SWS (measured gives the waking value, the lowest), all within 54 to 66. The Pearson correlation of pair correlation with offset is -0.751, -0.741 and -0.671 (p below 1e-200). With 15 degree bins the crossings are 55.4, 57.9 and 57.7 degrees; under reading (b) 54.1, 57.1 and 56.5, the waking value just outside the tolerance. The waking crossing of reading (a) is 0.2 degrees inside the tolerance, so the pass is close to the boundary.
How the test differed from the paper’s method
- ADn-ADn, ADn-PoS and PoS-PoS pairs are pooled.
ReproducedResults, second paragraph; Figure 2e,f
The angular velocity of the internal HD signal is similar during REM sleep and waking.
Angular velocity of the internal HD signal was similar during waking and REM (Fig. 2e,f).
Results, second paragraph; Figure 2e,f
The median REM to waking ratio is 0.87 over 1509 pairs (quartiles 0.60 to 1.26), inside 0.67 to 1.5, and 0.85 over 2129 pairs under reading (b); by mouse 0.69 to 1.07; the decoder variant gives 1.05. By a Wilcoxon test of the log ratio the median REM speed is below the waking one (p = 6.3e-18, median log10 ratio -0.060, about 13%), a test the paper does not make, but REM is faster in 581 of 1509 pairs.
How the test differed from the paper’s method
- The paper names no test for "similar"; the range 0.67 to 1.5 is chosen here.
ReproducedResults, fourth paragraph; Figure 4a,b
During waking, the observed correlations of HD cell pairs exceed those expected from independent rate-coded neurons for pairs with overlapping fields (0 plus or minus 15 degrees), and fall below them for pairs at about 60 degrees.
However, correlations expected from independent rate-coded neurons could not fully account for the observed pairwise correlations of HD cells (Fig. 4a). ... were observed for neuron pairs with overlapping fields (0±15°), with a marked negative signal correlation at 60±15° offset (Fig. 4b,c)
Results, fourth paragraph; Figure 4a,b
With the Methods' formula (mean cross-products over bins where the head direction is known at both times), the noise correlation of pairs at 0 plus or minus 15 degrees is positive in 110 of 118 pairs (median 7.84 Hz2, 0.81 of the chance product, Wilcoxon p = 5.3e-19) and of pairs at 60 plus or minus 15 degrees positive in only 35 of 256 (median -4.08 Hz2, p = 4.2e-32). Under reading (b) the medians are 4.22 and -1.63 Hz2 (p = 2e-24 and 6.8e-42). On independent Poisson surrogates with tuning curves estimated from their own spikes, the measure is biased in the same direction, median +0.64 Hz2 at 0 degrees and -0.37 Hz2 at 60, about a tenth of the observed values; the data exceed their surrogates in 110 of 118 pairs at 0 degrees (p = 5e-19) and fall below them in 223 of 256 at 60 degrees (p = 1.6e-31). It holds in the mice that appear thalamus only. Under the first criterion, with both cross-correlations on a Pearson scale, the 0 degree pairs are negative (median -0.78), which reflects the scale and not the data.
How the test differed from the paper’s method
- The criterion was revised after a first run, before this outcome was recorded, from a Pearson scaling that contradicted the Methods' stated formula to that formula; the first version is shown as a variant.
- Regions pooled.
ReproducedMethods, Bayesian reconstruction of head-direction; Figure 1c
The HD signal decoded with a Bayesian decoder from the HD cells faithfully tracks the head direction of the waking mouse.
These estimates faithfully tracked the head position in the waking mouse.
Methods, Bayesian reconstruction of head-direction; Figure 1c
In 28 sessions with at least 6 HD cells, two-fold cross-validated decoding has a pooled median absolute error of 11.7 degrees (chance 90); session medians 8.1 to 27.2, all at or below 30. With HD cells selected on the training half alone, the same 28 sessions give a pooled median of 11.7 degrees (session medians 8.1 to 24.7).
How the test differed from the paper’s method
- HD cells of both structures are decoded together; the paper decodes one structure at a time.
- The threshold of 30 degrees for "faithfully" and the 6 degree angle bins are chosen here.

Not tested in a notebook 13

13 metadata insufficient

BlockedResults, first paragraph; Supplementary Figure 2b· Metadata insufficient
ADn HD cells have peak rates almost three times those of PoS HD cells (p < 1e-10, Mann-Whitney, 242 ADn and 111 PoS HD cells).
ADn neurons exhibited peak rates almost three times higher than PoS cells (Supplementary Fig. 2b
Results, first paragraph; Supplementary Figure 2b
What the paper reports
- Mann-Whitney U-test, p < 10⁻¹⁰ (n = 242 ADn and 111 PoS HD cells)
The units table holds only spike times, with no electrode, shank or region, and every electrode's location is "unknown", so units cannot be assigned to the ADn or the PoS.
BlockedResults, first paragraph; Supplementary Figure 2c· Metadata insufficient
ADn HD cells convey 30% more head-direction information than PoS HD cells (1.44 against 1.11 bits per spike, p < 1e-6).
conveyed 30% more head-direction information than PoS cells (1.44 bit per spike vs. 1.11 bit per spike
Results, first paragraph; Supplementary Figure 2c
What the paper reports
- Mann-Whitney U-test, p < 10⁻⁶
No region per unit (see adn-pos-peak-rate).
BlockedResults, first paragraph· Metadata insufficient
Ensembles of 8.4 plus or minus 5.1 (s.d.) HD units per session were recorded in the ADn and 5 plus or minus 2.8 in the PoS.
Ensembles of HD neurons from ADn (8.4 ± 5.1 s.d. units per session) and PoS (5 ± 2.8 s.d.) were recorded by multi-site silicon probes
Results, first paragraph
No region per unit. Pooled, the sessions hold 0 to 26 HD cells (median 10). In the mice that appear thalamus only there are 7.4 plus or minus 6.2 (s.d.) HD cells per session over 18 sessions (7.8 plus or minus 6.1 without Mouse12-120808), against the ADn's 8.4 plus or minus 5.1; in the mice that appear dual-site 12.2 plus or minus 5.9 over 20 sessions, both structures together. The assignment of mice is consistent with the files but not shown by them, so these numbers do not decide the claim.
BlockedResults, third paragraph; Supplementary Figure 5· Metadata insufficient
Pairwise correlations of HD cells closer than 60 degrees are stronger in the ADn than in the PoS in every brain state (p < 1e-5, 345 ADn-ADn and 55 PoS-PoS pairs).
HD signal was more strongly preserved in the ADn than in the PoS as shown by the stronger pairwise correlations in any brain state
Results, third paragraph; Supplementary Figure 5
What the paper reports
- Mann-Whitney U-test, p < 10⁻⁵ (n = 345 ADn-ADn and 55 PoS-PoS pairs closer than 60 degrees)
No region per unit.
BlockedResults, third paragraph; Figure 3b· Metadata insufficient
The activity packets in the ADn and the PoS are coherent with the packet expected from the Bayesian decoded HD in waking, REM and SWS.
The actual activity packets, both in the ADn and the PoS, and across brain states, were highly similar to the predictions given by the signal decoded from a Bayesian estimator (Fig. 3b) in waking, REM and SWS
Results, third paragraph; Figure 3b
The coherence is shown per structure, and units have no region. The measure is also described only by reference to earlier work (the bootstrap of population vectors is not fully specified), which would block it as methods_underspecified as well. A version pooled over regions could be run, but the Methods do not say how the actual and expected packets are compared or how the probability is formed from the 5,000 random waking population vectors, so it is not run here.
BlockedResults, fourth paragraph; Figure 4c· Metadata insufficient
Noise correlations of HD cell pairs are stronger in the ADn than in the PoS (p < 1e-9, Kruskal-Wallis).
These signal correlations were stronger for ADn than PoS neurons
Results, fourth paragraph; Figure 4c
What the paper reports
- Kruskal-Wallis one-way analysis of variance, p < 10⁻⁹ (n = ADn-ADn pairs 96 and 200, ADn-PoS pairs 77 and 157, PoS-PoS pairs 29 and 57, at 0 and 60 degrees)
No region per unit.
BlockedResults, fourth paragraph; Figure 4d,e· Metadata insufficient
In the ADn the covariance-based optimal linear estimator of HD is about 50% better than the signal-based one, and this gain is larger than in the PoS (p < 0.05, 20 ADn and 8 PoS ensembles).
In the ADn, where noise correlations were highest for overlapping HD cells, the covariance-based OLE was approximately 50% better than the signal-based OLE (Fig. 4d,e), and significantly stronger than in the PoS
Results, fourth paragraph; Figure 4d,e
What the paper reports
- Mann-Whitney's U-test, p < 0.05 (n = 20 ADn and 8 PoS cell ensembles)
The claim compares ADn and PoS ensembles, and units have no region.
BlockedResults, fifth paragraph; Figure 5a,b· Metadata insufficient
The HD signal decoded from ADn assemblies during sleep gives tuning curves of PoS cells more similar to their waking tuning than the signal decoded from PoS assemblies does for ADn cells, in all brain states.
However, the HD signal extracted from ADn assemblies resulted in tuning curves more similar to actual HD fields than the HD signal decoded from PoS assemblies, in all brain states
Results, fifth paragraph; Figure 5a,b
What the paper reports
- wake: Mann-Whitney U-test, p < 10⁻¹⁰ (n = 86 PoS and 92 ADn HD cells)
- SWS: Mann-Whitney U-test, p = 0.018 (n = 86 PoS and 92 ADn HD cells)
- REM: Mann-Whitney U-test, p = 3 × 10⁻⁶ (n = 86 PoS and 92 ADn HD cells)
Decoding from one structure to compute tuning in the other needs the region of each unit.
BlockedResults, fifth paragraph; Figure 5c to e· Metadata insufficient
ADn assemblies predict PoS spike trains in all brain states better than PoS assemblies predict ADn spike trains, with optimal readout windows of 123 ms (waking), 133 ms (REM) and 79 ms (SWS).
Optimal readout windows were similar between waking and REM (median: 123 ms and 133 ms respectively) and shorter during SWS (median: 79 ms). In the reverse direction, PoS cell assemblies predicted spike trains of ADn neurons more weakly, especially during SWS
Results, fifth paragraph; Figure 5c to e
What the paper reports
- wake: Mann-Whitney U-test, p = 3 × 10⁻⁵ (n = 86 PoS and 62 ADn neurons)
- SWS: Mann-Whitney U-test, p < 10⁻¹⁰ (n = 86 PoS and 62 ADn neurons)
- REM: Mann-Whitney U-test, p = 1.1 × 10⁻⁶ (n = 86 PoS and 62 ADn neurons)
Predicting one structure from the other needs the region of each unit.
BlockedResults, sixth paragraph; Figure 6· Metadata insufficient
Fast synchrony between ADn HD cell pairs decreases with the difference of their preferred directions in all brain states (p < 1e-10, Kruskal-Wallis, 970 pairs).
The degree of synchrony decreased with the magnitude of difference between the preferred head directions of the neurons (Fig. 6b ... in all brain states (Fig. 6c and Supplementary Fig. 7a).
Results, sixth paragraph; Figure 6
What the paper reports
- Kruskal-Wallis test, p < 10⁻¹⁰ (n = 970 ADn HD cell pairs)
The claim is about ADn pairs on different shanks; units have neither region nor shank. The mice that appear thalamus only could stand in for the ADn, but that assignment is not shown by the files, and without shanks the pairs on the same shank, whose fast synchrony at 1 to 5 ms is distorted by spike sorting, cannot be removed.
BlockedResults, sixth paragraph; Figure 7a,b· Metadata insufficient
Putative monosynaptic connections were found in 4.45% of 6960 ADn to PoS pairs and 0.35% of PoS to ADn pairs; among pyramidal PoS cells, 0.41% against 0.18% (p = 0.014).
revealed putative synaptic connections in a fraction of the pairs (4.45% of 6960 pairs from ADn to PoS, 0.35% from PoS to ADn ... (0.41% ADn-PoS pairs versus 0.18% PoS-ADn pairs
Results, sixth paragraph; Figure 7a,b
What the paper reports
- binomial test, p = 0.014 (n = 3394 ADn and PoS pyramidal cell pairs)
Needs the region of each unit, and for the pyramidal subset waveforms, which the dandiset does not hold for units (raw broadband data are in only 3 of the 40 files).
BlockedResults, sixth paragraph; Figure 7d,e· Metadata insufficient
PoS HD cells postsynaptic to ADn HD cells share the preferred direction of their presynaptic cell (12 pairs), and such PoS pyramidal cells convey more HD information than those without an identified presynaptic cell (p = 0.011).
PoS HD cells that were post-synaptic to ADn HD cells shared the same preferred direction as their presynaptic neuron (Fig. 7d ... Pyramidal neurons in PoS that were synaptically driven by ADn HD cells conveyed more HD information than PoS neurons for which a presynaptic cell could not be identified
Results, sixth paragraph; Figure 7d,e
What the paper reports
- preferred direction of postsynaptic PoS against presynaptic ADn HD cells: Rayleigh test followed by Wilcoxon's signed rank test, p = 0.68 (n = 12 pairs)
- HD information of driven against other PoS pyramidal neurons: Mann-Whitney U-test, p = 0.011
Depends on the ADn to PoS connections, which need the region of each unit.
BlockedResults, sixth paragraph; Figure 7c· Metadata insufficient
ADn HD cell spikes are associated with high-frequency oscillations of the PoS LFP.
ADn cell spiking was associated with high frequency LFP oscillations in the PoS (Fig. 7c)
Results, sixth paragraph; Figure 7c
Needs the region of the units and of the LFP channels; electrode locations are "unknown".

Findings about the data

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

Description missingThe units tables hold only spike times. No unit has an electrode, shank, brain region, waveform or quality measure, and every electrode's location is "unknown". Units cannot be assigned to the antero-dorsal thalamus or the post-subiculum, which blocks 13 of the paper's 21 claims. The electrode groups and ripple annotations are consistent with Mouse24, Mouse25 and Mouse28 being the three dual-site mice, but the files do not say so.
Selection not documentedThe files hold 1959 unit rows (103 with no spikes) in 40 sessions, against the paper's 1077 recorded cells in 42 sessions. Mouse12-120809 alone has 292 units, 167 of them firing below 0.5 Hz. Unit ids run from 0 to N-1 and carry no shank or cluster number, the units table has no quality measure, and 1526 rows fire at 0.5 Hz or more, so the paper's 1077 cells cannot be recovered. Nothing marks which units or sessions the paper analysed (its Methods discard some sessions).
Description missingThe open-field exploration (the paper's waking epoch) is not marked. The "Awake" state also covers waking in the home cage, and some sessions hold a second exploration or a larger second environment. The epoch had to be found from position.
File format errorIn 36 of 40 files one unit's spike times are not sorted. Each such unit holds exactly two sorted runs, and in all 36 the first run ends at an index of the form k x 65536 - 1. The second run starts within 36 s of the start of the recording and the first between 24 and 13704 s, and the two end within 15 s of each other. In 26 units no time of the second run appears in the first; in 10, 821 to 22281 times do. This suggests an error in the conversion.
OtherIn Mouse12-120808 the position and the spikes do not share a clock. Head-direction tuning is best with position shifted by -1.5 s at the start of the exploration and -3.4 s at its end, and no unit passes the HD criteria as deposited; with position realigned by the best shift of each quarter, 13 units do. In all other sessions the best shift is -0.2 to 0 s.
Metadata errorThe raw LED positions in `acquisition` are stated in meters with a conversion of NaN but hold camera pixels (values in the hundreds).
Data missingMouse32-140820 has no sleep-state table and no processed position, and Mouse17-130201 has no processed position; both have raw LED series only. In three Mouse20 sessions the position series ends 11 to 109 minutes before the recording.
Description missingThe annotation series `dow`, `hdD` and `rip` (labels such as "DOWN ON 5", "HD UP ON", "Ripple peak 1") have no description.
Metadata errorThe dandiset's related resource for the paper gives its name and identifier as "nternally organized mechanisms of the head direction sense" (first letter missing, and the title in place of a DOI); the URL is correct.
Methods not availableNeither the paper nor the dandiset links analysis code, so the von Mises fit, the Rayleigh test, the scaling in the cross-correlogram estimate of angular velocity and the coherence measure had to be chosen or could not be repeated.

Notebooks

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

Reproduction Head-direction cells across waking and sleep (Peyrache et al. 2015)

Tests the findings of Peyrache, Lacroix, Petersen and Buzsaki (2015) on dandiset 000056, which holds spike times, LED positions and sleep states from 40 sessions of 7 mice recorded in the anterior thalamus and post-subiculum. It classifies head-direction cells, compares their rates and pairwise correlations across waking, REM and slow-wave sleep, estimates the speed of the internal head-direction signal from cross-correlograms, and tests noise correlations and decoding. The units carry no brain region, so the comparisons between the two structures cannot be made. All files

Reproducing Peyrache et al. (2015) from dandiset 000056
reproduce_peyrache_2015.ipynbExecuted, with outputs · Source

Verification

Round 1
2026-10-16Agreed with 21 of 21 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: Round 1. The first executed notebook with the conclusion cells and the summary removed, the helper module and the claims file, with the paper. The verifier checked the archive metadata for the blocked claims and tested helper functions and two measures on synthetic inputs. Its report had no attempt record and gave its own outcome for each claim; all 21 agree with the record.
Round 2
2026-10-16Agreed with 21 of 21 outcomes. Independent agent run by CatalystNeuro (same harness, separate context, did not write the notebook). Reviewed: Round 2. The full executed notebook after the round-1 revision, with its conclusions and summary, the helper module, the claims file, the attempt record and the README, with the paper. The verifier agreed with all 21 outcomes. Besides the points below, it found that cell 18 gave the shift of the largest contrast as 10 s in 10 sessions where cell 17 shows -10 s in 9 and +10 s in one, and that the README's run time (nine minutes, five streaming) did not match cells 53 and 1; both are corrected, the README from a new run (10.1 min, 322 s loading). It also noted that the record's finding on the unsorted units had no printed output; the new data check prints it, and the finding now gives what it shows (two runs in all 36 units, all broken at k x 65536 - 1, with repeated times in 10).

Method differences

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

Regions pooled, as the units carry no region.hd-cell-fraction
The von Mises fit (maximum likelihood fit to the tuning curve as a distribution of angles) and the Rayleigh test (resultant length of the tuning curve with the number of waking spikes) are choices made here; the Methods do not specify them.hd-cell-fraction
The open-field epoch is found from position, as the files do not mark it.hd-cell-fraction
Mouse17-130201 and Mouse32-140820 have no processed position and are left out.hd-cell-fraction
Mouse12-120808 is counted as deposited, with position and spikes on different clocks; the realigned session is a variant.hd-cell-fraction
ADn and PoS cells pooled; the paper plots them separately.fig2a-rates-preserved
Pairs within and across structures are pooled, and pairs on the same shank cannot be removed.fig2b-pair-correlations-preserved
Bin sizes are those of Figure 3a; Figure 2b does not state them.fig2b-pair-correlations-preserved
ADn-ADn, ADn-PoS and PoS-PoS pairs are pooled.fig3a-packet-60-degrees
The paper does not say how the expected and observed cross-correlograms are put on one scale; a free scale and offset are fitted here.fig2f-sws-tenfold-faster
The lag range (4 s) and speed grid are chosen here.fig2f-sws-tenfold-faster
The expectation over the chi-squared speed distribution is computed exactly from its characteristic function.fig2f-sws-tenfold-faster
Pairs within and across structures are pooled; the paper shows ADn-ADn and PoS-PoS pairs.fig2f-sws-tenfold-faster
The paper names no test for "similar"; the range 0.67 to 1.5 is chosen here.fig2f-rem-similar-to-wake
The criterion was revised after a first run, before this outcome was recorded, from a Pearson scaling that contradicted the Methods' stated formula to that formula; the first version is shown as a variant.fig4b-noise-correlations
Regions pooled.fig4b-noise-correlations
HD cells of both structures are decoded together; the paper decodes one structure at a time.decoding-tracks-hd
The threshold of 30 degrees for "faithfully" and the 6 degree angle bins are chosen here.decoding-tracks-hd

Provenance

Attempt record: 000056-peyrache-2015-01
Paper: Internally organized mechanisms of the head direction sense
Date: 2026-10-16
Made by: An agent, run by CatalystNeuro
Dataset: DANDI:000056
Version: 0.250624.0430 (doi:10.48324/dandi.000056/0.250624.0430)
Run on: GitHub Actions
Harness: Claude Code (GitHub Actions)
Model: claude-opus-5-5
Protocol: pilot-7

Transcripts of the agent runs

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

← DANDI:000044 DANDI:000059 →