DANDI:001051

we recorded activity from over 1,800 units spanning ... 25 brain areas (this includes all non-noise units, as described in STAR Methods; 1,821 ± 341.9 units; 28.1 ± 5.1 areas, mean ± SD

Results, High-quality multi-Neuropixels recordings, paragraph 2; Figure 4B

44 sessions of the 87 have in-area units on exactly six probes; they have 1821.0 non-noise units (SD 341.9) and 28.11 areas (SD 5.12), the paper's values; unchanged without the 6 sessions that differ from the code's table.

high-quality recordings from 25 mice, totaling 87 experiments and 467 probe insertions

Discussion, paragraph 2

Over the 87 sessions 25 mice, 87 experiments and 507 probes with units; 507 is within 467 +- 46.7. Of these 467 are in the code's units table; the other 40 probes carry 2004 units, 1873 on 'Track not annotated' channels, which the paper's tables leave out. Over all 99 sessions 27 mice and 578 probes with units.

How the test differed from the paper’s method

• The dandiset holds 40 probe recordings in the 87 sessions that the paper's tables leave out; nothing in the files marks them, so insertions are counted as probes with units, with the code's 467 shown.

Overall, we achieved a penetration success of 5.46 probes per session

STAR Methods, Neuropixels recordings, Probe insertion, paragraph 2

507 probes with units over 87 sessions give 5.83 per session (a), 578 over 99 give 5.84 (b), within 5.46 +- 0.546; with the code's 467 insertions, 5.37. The code's probes table lists 467 probes for the 87 sessions and 533 for the 99 (5.37 and 5.38 per session), every one marked inserted, so it records no failed attempt. No count gives 5.46.

found it was only slightly higher than that observed in our previous recordings using the 5-mm insertion window method (SHIELD: 25.1 ± 0.1 mm, 5-mm window: 22.82 ± 0.05; median ± SEM

Results, High-quality multi-Neuropixels recordings, paragraph 4; Figure 4D

Median max_drift 25.05 um over 144,030 SHIELD non-noise units (paper 143,828) and 22.82 over 262,177 VBN non-noise units, within 25.1 +- 0.1 and 22.82 +- 0.05; 25.22 without the 6 differing sessions.

How the test differed from the paper’s method

• SHIELD units include 202 more non-noise units than the paper's count, from the 40 probe recordings the paper's tables leave out (143,828 non-noise units on the 467 insertions in the code's table, 202 on the other 40). The code's own table has 144,068 for the 87 sessions; outside the 6 sessions where it describes other units it and the files both give 135,334.

the percentage of units passing our standard quality metrics filter (see STAR Methods) was very similar for the SHIELD versus 5-mm window implants (% units passing quality metrics criteria: SHIELD: 41.1, 5-mm window: 40.7)

Results, High-quality multi-Neuropixels recordings, paragraph 5

41.09% of SHIELD and 40.69% of VBN non-noise units pass the QC filter.

On each recording day, we measured the activity of roughly 300 units from each probe (Figure 4E). ... unit yield was not significantly different across days for SHIELD mice (p = 0.48, Kruskal-Wallis H test)

Results, High-quality multi-Neuropixels recordings, paragraph 6; Figure 4E

What the paper reports

• Kruskal-Wallis H test, p = 0.48 (n = 129, 126, 121 and 91 insertions on days 1 to 4)

Over all 500 insertions with non-noise units, medians 308, 302, 287.5 and 277 per day and Kruskal-Wallis p = 0.333; over the 467 insertions of the code's table (129, 126, 121 and 91, the figure's counts) medians 314, 305.5, 295 and 303 and p = 0.477. Every median is within 300 +- 30.

In contrast to the 5-mm window preparation, for which unit yield decreased on the second recording day (p = 6e–22, Kruskal-Wallis H test)

Results, High-quality multi-Neuropixels recordings, paragraph 6; Figure 4E

What the paper reports

• Kruskal-Wallis H test, p = 6 × 10⁻²² (n = 459 and 446 VBN insertions on days 1 and 2)

VBN medians 321 (459 insertions) and 261 (446) on days 1 and 2; Kruskal-Wallis p = 1.2e-24 for day 1 against day 2 (a) and 1.2e-19 over the code's six groups (b).

Additionally, the average firing rate of recorded units was similar across days (p = 0.21 Kruskal-Wallis H test).

Results, High-quality multi-Neuropixels recordings, paragraph 6; Figure 4E

What the paper reports

• Kruskal-Wallis H test, p = 0.21 (n = 23, 24, 22 and 18 sessions on days 1 to 4)

Kruskal-Wallis p = 0.223 over 23, 24, 22 and 18 sessions (text 0.21, code 0.26).

we measured a consistent fraction of units with well-defined spatial receptive fields (RFs) (Figure 4E and 4F; p = 0.38 Kruskal-Wallis H test)

Results, High-quality multi-Neuropixels recordings, paragraph 7; Figures 4E and 4F

What the paper reports

• Kruskal-Wallis H test, p = 0.38 (n = 12, 11, 12 and 11 sessions on days 1 to 4)

With RF p values recomputed here and p < 0.01, median fractions 0.900, 0.877, 0.913 and 0.900 over 12, 11, 12 and 11 sessions, Kruskal-Wallis p = 0.671; at 0.05, p = 0.368; with the code's p values, 0.864 and 0.813.

How the test differed from the paper’s method

• RF p values are recomputed with the AllenSDK chi-square bootstrap (1000 resamplings) with their own seed; they match the code's table (r = 0.9990) where units agree.

In addition, the stimulusevoked temporal dynamics were highly similar across the 4 recording days (Figure 4G).

Results, High-quality multi-Neuropixels recordings, paragraph 7; Figure 4G

595, 482, 534 and 446 VISp units with significant RFs on days 1 to 4 (12, 11, 12 and 11 sessions and mice); Kruskal-Wallis p = 0.343 for the per-unit response 0 to 250 ms (a) and 0.70 for the time of peak (b); the day population responses correlate at r = 0.977 to 0.991.

we observed rhythmic 3–5 Hz waves of activity in VISp (Figure 7A, left) ... These MOp alpha-like events occurred in a slightly higher frequency band (5–7 Hz)

Results, SHIELD recordings reveal multi-regional subnetworks, paragraphs 3 and 4; Figure 7A

Under frequency reading (b), the autocorrelogram spectrum peak, VISp events at 4.0 Hz and MOp at 4.5 Hz (medians), MOp higher by the one-sided rank-sum test (p 5.3e-7 to 0.0016) in all 8 L, E and set combinations, which the criterion accepts. The 2 s autocorrelogram gives a 0.5 Hz grid on which the MOp median sits on the lower edge of the widened 4.5 to 7.7 Hz range (22.7% of L2 E2 MOp values are exactly 4.5 Hz); with the spectrum zero-padded to 0.049 Hz steps the medians are 4.05 to 4.10 Hz (VISp) and 4.64 to 4.74 Hz (MOp), inside the ranges, with MOp higher at p 7.1e-7 to 0.0038 in all 8. Under reading (a), the inverse of the median interval between maxima, VISp 4.17 to 4.21 Hz and MOp 4.35 to 4.55 Hz, and MOp is not significantly higher (p 0.078 to 0.76). 1469 to 1704 qualifying VISp events and 80 to 122 MOp events. The detector finds synthetic sinusoidal rates of 3 to 5 Hz only (none at 2.5 Hz or below or 5.5 Hz or above), so a 5 to 7 Hz MOp band cannot appear under either reading; this follows from the event rule of the claims file.

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.
• Reading (b) restricts the spectrum peak to 1 to 20 Hz, a standard choice; the zero-padded spectrum (2048 points) is a standard interpolation shown as a variant.

The VISp events were often triggered by the offset of visual stimulation (Figure 7B) ... The timing of MOp events was not related to visual stimuli (Figure 7B).

Results, SHIELD recordings reveal multi-regional subnetworks, paragraph 5; Figure 7B, right

In all combinations VISp events peak 100 to 150 ms after stimulus offset (KS against uniform p 9.2e-233 or less; chi-square p printed as 0) and MOp events do not differ from uniform (chi-square p 0.41 to 0.53, KS p 0.28 to 0.62).

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.
• The stimulus before an event is the last active image presentation, as the code's stimulus-trials table holds active presentations.

We found that CP-RF firing rates were significantly more modulated, particularly during the troughs of the oscillation when the visual thalamo-cortical network is suppressed.

Results, Fine-scale alpha rhythm coupling between visual cortex and striatum, paragraph 4; Figure 8E

What the paper reports

• Wilcoxon rank-sum test per time point, p < 0.05, Benjamini-Hochberg-corrected (n = 439 events)

In every combination (387 to 493 events, 4 or 5 sessions) CP-RF and CP-noRF rates differ in 112 to 222 bins by the rank-sum test, 24 or 25 at the troughs of the VISp rate, and the CP-RF range over 0 to 1 s is larger (1.96 to 2.11 against 1.18 to 1.34 Hz; 0.50 to 0.56 against 0.30 to 0.32 normalised). The code's signed-rank variant gives 98 to 216 bins.

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.

Analysis code used in this paper is available here

STAR Methods, Key resources table and Data and code availability; DynamicGating_ephys_metrics.ipynb and metadata_tables/units_table.csv at commit 3c09a0dc

All 87 listed sessions are dandiset sessions (by the NWB identifier), and session type, mouse and genotype agree for all 99 sessions. In 81 of 87 sessions the code's units table and the dandiset agree unit by unit (no matched unit differs in max_drift, firing_rate or quality). In 6 sessions (1176989662, 1178460518, 1178693650, 1179911454, 1180266229, 1181096406) 55 to 99.8% of matched units differ in max_drift and firing_rate and 14 to 38% in quality; there the files' spike counts fit the files' firing rates (coefficient of variation of n_spikes / firing_rate 0 in five sessions, 0.076 in 1176989662) but not the code's (5.4 to 11.8, against at most 0.0003 elsewhere), and the RF p values recomputed here are uncorrelated with the code's (r = 0.03 over 251 units, against 0.9990 over 7095 units in the other sessions), so the same unit ids name other units. In 5 of them the code's table has 328 units (133 CP) that are not in the dandiset, each session's on one probe_id of the code's (38 to 114 units) that is not among the file's probe_ids and has no unit in the file, and none with a peak channel of a dandiset unit; the dandiset has no unit missing from the code's table. Where the units agree, recomputed RF significance at 0.01 agrees for 98.8% of units pooled; per session the fraction differing has a median of 0.008 and a maximum of 0.038 and exceeds 1% in 33 of 79 sessions, all 83 differing units having p values of 0.003 to 0.020 in both computations, at the threshold (two bootstraps with different random draws). The dependent claims are tested on all 87 sessions; leaving out the 6 changes their numbers little (shown for the unit counts, drift, QC fraction and yield).

How the test differed from the paper’s method

• A link claim added by the author; the code's notebook and tables are matched to the dandiset by session and unit identifiers.

we typically made ... probe insertions (20 ± 4.1, mean ± SD; Figure S4A) targeting ... implant holes (10.4 ± 2.3, mean ± SD; Figure S4B)

Results, High-quality multi-Neuropixels recordings, paragraph 2; Figures S4A and S4B

Insertions per mouse, mean 20.28 (SD 3.34) over the 87 sessions (reading a) and 21.41 (SD 3.33) over all sessions (b), within 20 +- 4.1; with only the 467 insertions of the code's table, 18.68. The implant holes per mouse (10.4 +- 2.3) could not be measured, as no file records the hole of an insertion; that part is untested, so the claim holds only in the part that could be tested.

How the test differed from the paper’s method

• Holes per mouse not tested (no hole per insertion in the dandiset or the code's tables).

we found that mouse behavioral performance on the visual task was stable across the 4 recording days despite repeated probe insertions (median rewards received: day 1, 59.5; day 2, 63.5; day 3, 56.5; day 4, 60.5; Kruskal-Wallis test: p = 0.96)

Results, High-quality multi-Neuropixels recordings, paragraph 8

What the paper reports

• Kruskal-Wallis test, p = 0.96

No reading finds a difference across days (Kruskal-Wallis p 0.731 to 0.817), so the claim of stability holds. The medians do not all match: over the 87 sessions (a) all rewards 73, 65, 61, 53.5 and (b) earned rewards 67, 59, 55, 48 (days 2 and 3 within tolerance in each); over all sessions (a) 68, 63.5, 60, 56.5 (days 2 to 4) and (b) 62.5, 57.5, 54, 50.5 (days 1 to 3). Medians fall from day 1 to day 4 in every reading. Hit trials exceed earned rewards because of the no-reward block. The paper's medians end in .5 on days 1 and 3, which needs an even number of sessions those days; the 87-session set has 23 on day 1 and the 99-session set 25 on day 3, so neither set is the paper's.

VISp and MOp events tended to occur during periods of lower arousal, when the animal was immobile and the pupil was constricted (during alpha events, pupil diameter was ... 25% smaller relative to randomly chosen session times)

Results, SHIELD recordings reveal multi-regional subnetworks, paragraph 5; Figure 7B

At qualifying events the pupil is smaller (Wilcoxon p at most 9.1e-15) and running lower (0.03 to 1.16 against 4.1 to 7.7 cm/s at random times, p at most 8.0e-6) for VISp and for MOp in every combination. The change is not 25%: VISp -12.9 to -15.4% (area, mean), -7.5 to -8.8% (diameter); MOp -30.8 to -36.5% (area), -17.3 to -20.9% (diameter); pooled -14.0 to -16.3% (area). No reading falls within -22.5 to -27.5%. The running and pupil streams share the stimulus clock as far as the data check shows: running slows after hits against misses from a median 0.225 s after the change, and the pupil constricts after white against black flashes from a median 0.35 s, with small contrasts before time 0. Alignment of at least one stream is not shown in 12 of the 99 sessions (2 without a running contrast, as they have no hit or no miss trials, 2 without eye tracking, 5 with a weak running contrast and 3 with a weak pupil contrast).

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.
• Diameter is the square root of window-mean area; two sessions without eye tracking contribute no pupil values.

These units were biased toward the medial dorsal striatum mid-way along the anteriorposterior extent of the CP

Results, Fine-scale alpha rhythm coupling between visual cortex and striatum, paragraph 2; Figures 8B and 8C

688 of 2376 QC CP units (52 sessions, 21 mice) have significant RFs. They are more dorsal than CP-noRF units (median y 3150 against 3500 um, one-sided p = 1.9e-58) but more lateral (median z 4075 against 4350 um, two-sided p = 1.8e-33). Their median AP coordinate, 5388 um, is in the posterior part of the CP sampled (3700 to 6500 um), where the RF fraction rises from 0.10 to 0.45 across AP terciles.

How the test differed from the paper’s method

• RF significance recomputed here (link-code-tables); the CCF z axis orientation is checked against VISp (3050) and midline structures (5300).

CP-RF units showed significantly larger visual responses to this stimulus compared with CPnoRF units (Figure 8D).

Results, Fine-scale alpha rhythm coupling between visual cortex and striatum, paragraph 3; Figure 8D

What the paper reports

• Wilcoxon rank-sum test per time point, p < 0.05, Benjamini-Hochberg-corrected (n = 80 CP-RF and 205 CP-noRF units)

In the five sessions CP-RF units respond more than CP-noRF units in 53 bins of 0 to 250 ms (first at 55 ms), CP-RF higher in all, none before onset, with the Methods' QC, the criterion's units. The unit counts there, 70, 162 and 218, miss the expected 80 +- 8 and 205 +- 20.5 for CP-RF and CP-noRF (VISp 218 is within 239 +- 23.9). With the code's filter (presence ratio above 0.9, noise units kept and split by their RF p values) the counts are 79, 206 and 239, within one unit of the paper's, with 78 significant bins, all CP-RF higher. The effect holds under both; the expected counts only under the variant, so the claim holds in part. Whether the criterion should take the code's filter is asked of a reader (questions/001051.yaml).

How the test differed from the paper’s method

• The unit counts match only with the code's filter (presence ratio above 0.9 and no noise filter), which the claims file names as a variant; the effect holds with both. Noise CP units now have RF p values, so the variant splits them by RF as the code does.

We found that the distribution of QC metrics was comparable with both surgical preparations (Figure 4D).

Results, High-quality multi-Neuropixels recordings, paragraph 5; Figure 4D

Reading (a), the code's Mann-Whitney test, finds all 7 metrics different (p 3.9e-292 to 3.9e-6), as the code's printed output does; reading (b), medians within 10%, holds for 5 of 7 (isi_violations ratio 0.79, amplitude_cutoff 0.89). Both readings give the same outcome. Effect sizes are small (rank-biserial at most 0.07 in absolute value) and the two metrics outside 10% are better for SHIELD.

we find that activity is transiently suppressed directly after probe insertion and gradually recovers with a time constant of 25-30 minutes (Figure S4I)

STAR Methods, Neuropixels recordings, Probe insertion, paragraph 1; Figure S4I

Activity recovers over the session (normalised rate 0.826 in minute 1, 0.999 at minute 30, 1.077 at the end, reading a), but the fitted time constant over the common 77.3 min from the first stimulus is 36.0 min (SE 2.3) under (a) and 35.4 min (SE 2.3) under (b), outside 22.5 to 33 min; per-session fits give a median of 34.8 min (IQR 23.1 to 52.5). Shorter windows give longer time constants, 41.2 to 42.2 min over the shortest active block (64 min), 45.9 to 47.1 at 60 min, 65.2 to 84.0 at 40 to 50 min, and no recovery at 30 min (tau unbounded); per-session fits over each session's active block give a median of 39.8 min (IQR 25.7 to 66.8), 26 of 87 sessions inside the range. No window gives a pooled tau inside the range, so the outcome does not rest on the window.

How the test differed from the paper’s method

• The fit (a - b exp(-t / tau) on 1-minute means, least squares) and the 87 sessions (the paper has 85) are not given by the paper.
• The fit window is not given (the S4I legend plots the change over time from experiment start); the main fit uses the longest window all sessions share, 77.3 min from the first stimulus, which runs past the active block. Windows of 30 to 64 min and per-session active blocks are shown; all give tau above 33 min.

Activity during post-stimulus alpha events was weakly but significantly correlated with activity during the preceding stimulus, indicating that alpha events tend to replay recent sensory stimulation (Figure S6; r = 0.77

Results, SHIELD recordings reveal multi-regional subnetworks, paragraph 5; Figure S6

What the paper reports

• correlation, r = 0.77, p = 1.8 × 10⁻⁶

Across sessions (7 under L1, as in the paper, 8 under L2; four points each) the correlation of evoked and event responses is r = 0.21 to 0.63 over 32 combinations of the readings and windows, below 0.693 in all; p < 0.05 in 18 of 32. Events are assigned to the last presentation before them, as the criterion says, which takes events that start during an image (473 of 1656 L2 E2 events over all sessions) and after the active block (64); keeping only events 0 to 0.5 s after the offset, the post-stimulus events of the S6 legend, gives r = 0.19 to 0.53, p < 0.05 in 14 of 32, none within tolerance. The per-unit variant gives r = 0.46 to 0.67 (0.52 to 0.66 with post-offset events only).

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.
• Rows are normalised by their maximum, a standard choice; sessions enter when each of the four images precedes a qualifying event.

Time lags between VISp and MOp and their associated regions in the 3–7 Hz band straddled zero lag, indicating that these areas did not consistently lead or follow activity in the other network nodes (Figure 7E).

Results, SHIELD recordings reveal multi-regional subnetworks, paragraph 7 (top of page 13); Figure 7E

On the code's test 94 of 102 pair results differ from zero after correction; only VISp to VISl under L1 does not. Median lags are -23 to +14 ms and the fraction of positive lags 0.12 (MOp to PO) to 0.97 (VISp to VISal), over 25 to 1143 events per pair. Whether the claim means that the median lag does not differ from zero (the criterion) or that the distribution of lags spans zero is asked of a reader (questions/001051.yaml).

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.

Neither visual nor motor events appeared to engage the hippocampus.

Results, SHIELD recordings reveal multi-regional subnetworks, paragraph 4; Figures 7A and 7F

For VISp events DG is significantly and positively coherent in every combination (+0.012 to +0.032), with CA1, CA3, ProS, SUB or POST also under R2, and ProS under L1 E2 R1 and SUB under L2 E2 R1; for MOp events CA1 and DG under L2 E1 R2, and CA1 (+0.025, q = 0.044) under L1 E2 R2. 20 of 32 combinations and references have a significant positive hippocampal pair. The hippocampal effects are small next to the visual areas (+0.06 to +0.39).

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.

During VISp events, VISp activity was tightly coupled to the secondary visual thalamus (LP) and higher-order lateral visual cortex (VISl), as evidenced by sharp peaks in the CCG at near-zero lag. VISp was not strongly coupled to somatosensory thalamus (PO) or SSp. Indeed, coherence analysis revealed a prominent peak in the 3–7 Hz band between VISp population spiking activity and LP and VISl, but not PO or SSp (Figure 7D)

Results, SHIELD recordings reveal multi-regional subnetworks, paragraph 6; Figures 7C and 7D

The CCG part holds in every combination (LP and VISl peak at 0 to 20 ms, heights 0.335 to 0.508, above PO and SSp at 0.055 to 0.092). The coherence part holds under R1 in all 8 combinations, but under R2 only with L2 E1: with L1 SSp is also significant (+0.008 to +0.011, q 0.0032 and 2.5e-5), with L2 E2 SSp is (+0.007, q 0.048 and 0.036; PO +0.011 at q 0.051). Every part is tested in every combination. The readings disagree, which the claims file's header makes methods underspecified. Open for the authors is how the random-time coherence is formed (the Methods' wording reads as R1, the code does R2).

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.

Conversely, MOp showed the opposite pattern: strong coupling to PO and SSp, but not to LP or VISl.

Results, SHIELD recordings reveal multi-regional subnetworks, paragraph 6; Figures 7C and 7D

The criterion needs MOp-PO to be significantly coherent. Under E1 (8 of 16 combinations) PO is untested (fewer than 25 events), so coherence_part is False there. Under E2 in all 8 combinations MOp is significantly coherent with PO (+0.15 to +0.20) and SSp (+0.19 to +0.25) and not with LP, and the CCGs of PO and SSp peak at 0 to 20 ms (0.20 to 0.33) against 0.03 to 0.06 for LP and VISl. VISl is untested under L1 and under L2 E1, and coherence_part counts an untested VISl as passing, so every part is tested and holds only under L2 E2 (4 combinations; VISl +0.005 and +0.027, q 0.89 and 0.34). The readings disagree, which the claims file's header makes methods underspecified. Changed from reproduced after verification round 3 agreed with the change proposed after round 2.

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.
• Under E1 the PO part, and under L1 and L2 E1 the VISl part, are untested (fewer than 25 events); all four parts are tested only under L2 E2 (4 of the 16 combinations).

In addition to visual corticothalamic regions, VISp alpha events enhanced coherence between VISp and medial cortical structures anterior cingulate area (ACA) and RSP, as well as CP

Results, SHIELD recordings reveal multi-regional subnetworks, paragraph 8 (page 13; the text cites it as Figure 6F); Figure 7F

What the paper reports

• Wilcoxon signed-rank test, event against random coherence, p < 0.05, Benjamini/Hochberg-corrected (n = 34 CCF regions)

ACA, RSP, CP, LGd, LP and VISl are each significant and positive in 14 of 16 combinations when any subdivision counts and 8 of 16 when all must (22 of 32 combination and mode rows); under L1 E2 R1 CP is not (q = 0.1), and with all subdivisions required ACAv or RSPagl or RSPd fail under R1. SSp has no subdivisions in these files; RSP has RSPagl, RSPd and RSPv, ACA has ACAd and ACAv, under both L1 and L2. The readings disagree.

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.

Conversely, MOp events enhanced coherence between MOp and SSp/PO, as well as motor-related midbrain structures MRN and deep layers of the superior colliculus (SCiw).

Results, SHIELD recordings reveal multi-regional subnetworks, paragraph 8 (page 13; the text cites it as Figure 6F); Figure 7F

What the paper reports

• Wilcoxon signed-rank test, event against random coherence, p < 0.05, Benjamini/Hochberg-corrected (n = 34 CCF regions)

Under E2, SSp, PO, MRN and SCiw are significant and positive with L1 (R1 and R2) and L2 R2; with L2 R1 MRN is not (q = 0.11); under E1 PO, MRN and SCiw have too few events. MRN and SCiw each come from one session. The readings disagree.

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.

CP-RF lagged behind VISp responses by ... 20 ms, whereas CP-noRF had slower responses that lagged behind CP-RF units by an additional 20 ms

Results, Fine-scale alpha rhythm coupling between visual cortex and striatum, paragraph 3; Figure 8D

With latency as the first time the population mean reaches half its peak (b), CP-RF lags VISp by 16 ms and CP-noRF lags CP-RF by 20 ms, within 20 +- 5; with the time of peak (a) the lags are 34 and 9 ms, outside. Both filters (the Methods' QC and the code's) give the same values. The paper says only that CP-RF lagged VISp by about 20 ms and that CP-noRF had slower responses, and gives no latency definition; the claims file lists both readings without preferring either. The readings give different outcomes and nothing in the paper picks one, so the claim is methods underspecified. Half-peak latency had been adopted as an assumption after it was seen to match; changed from reproduced after verification round 3 agreed with the change proposed after round 2. Question for the authors - which latency measure was used.

How the test differed from the paper’s method

• Both latency readings of the claims file were run, half-peak time of the population mean (b) and time of the population peak (a); the paper gives no measure.

As predicted, CP-RF units displayed higher coherence with VISp in the 3–7 Hz band, indicating that alpha events coordinate functionally related cortico-striatal subnetworks.

Results, Fine-scale alpha rhythm coupling between visual cortex and striatum, paragraph 4; Figure 8F

What the paper reports

• Wilcoxon signed-rank test, p = 3.8 × 10⁻¹⁴ (n = 439 events)

Under R2 (the code) CP-RF coherence with VISp exceeds CP-noRF (0.096 to 0.110 against 0.042 to 0.052, paired Wilcoxon p 6.5e-11 to 1.2e-14); under R1 it does not (p 0.36 to 0.75). The readings disagree. The code's L2 E2 readings give 439 events, the paper's count, and p = 1.9e-10 under R2.

How the test differed from the paper’s method

• Run under the claims file's 16 combinations of area membership (L1, L2), event rule (E1, E2), random-time coherence (R1, R2) and session set (a, b); population windows are cut from one 10 ms grid per session. Events are detected on the smoothed rate, and the autocorrelogram threshold, CCGs and coherence use the unsmoothed rate, a detail taken from the code's printed counts quoted in the claims file's header (1653 VISp and 113 MOp qualifying events, code cell 12), which the unsmoothed autocorrelogram matches (1643 to 1704 and 119 to 122 under E2) and the smoothed one does not (3035 to 3426 and 152 to 170). The Benjamini-Hochberg family of the pair test holds brain areas only (not CP-RF, CP-noRF, VIS or unannotated labels); this changes significance for 4 of 734 pair results.