DANDI:000691

The Kiloscope has a resolution of approximately 4 μm at the center of its FOV, which decreases to about 5 μm at the edges and is limited by the sensor pixel size

Main text, page 547, paragraph 1; Fig. 1f

No reading puts both values within tolerance, which the criterion asks for, with either optimizer or with the lower-cost fit for each bead. Under (a) the edge (1500 um group, 11 beads) is 5.441 um, within 5 +/- 0.5, and the centre (11 beads) is 4.467 um, 0.067 um above the bound of 4.4. The mean major axis (c) is 4.537 um at the centre and 7.373 um at the edge, where one bead is fitted at 27.77 um; the minor axis gives 2.722 and 3.306 um (b) and 2.669 and 3.094 um (d). The PSF is narrower than the 5.35 um pixel, and the fits depend on the optimizer; with Jacobian scaling 15 of 118 beads' major FWHM move by more than 0.5 um and (a) gives 3.639 um at the centre (within) and 5.892 um at the edge (0.392 um above the bound). Taking for each bead the lower-cost fit gives 4.467 and 5.892 um under (a), both outside, and 4.685 and 7.944 um under (c); the 27.77 um edge bead keeps its main fit, whose cost is lower (60.092 against 79.942). These are the numbers of the run of 2026-10-21 made for the second revision. The seeds are fixed, but the fits differ between runs on different machines. In the run reviewed in round 2 (as the verifier quotes it) the centre under (a) was 4.422 um (0.022 um above the bound), the mean major axis 4.416 um at the centre, 14 of 118 beads moved with Jacobian scaling, which gave 3.578 um at the centre, the lower-cost fits gave 4.084 um at the centre (within) and 5.892 um at the edge under (a) and 4.222 and 7.952 um under (c), and both fits of the 27.77 um bead had a cost of 60.092. In both runs no reading with any of the three sets of fits has both values within tolerance, and the FWHM is larger at the edge than at the centre under every reading and fit. The outcome is partial because one of the two positions is within tolerance under (a) in each set of fits, the edge with the main fits and the centre with the Jacobian-scaled fits (and, in the run of round 2, with the lower-cost fits). New claim.

How the test differed from the paper’s method

• MATLAB's lsqcurvefit is replaced by scipy's least_squares (trust-region-reflective, the code's start, bounds and tolerances); the fits of sub-pixel beads depend on the optimizer and differ between runs on different machines, by up to 0.4 um in a group's median (4.467 against 4.084 um for the centre's lower-cost fits).
• The code reads the 91 images in order of file date, which the repository does not keep; they are read in file-number order, which the Featherscope's region labels confirm for that stack. For the Kiloscope stack the order is not confirmed; each region's listed slice lies at the median of its brightest slice, but within 5 slices of it for only 72% of regions.

The second design, the Kiloscope, provides cellular resolution imaging across a 4.8 × 3.6 mm FOV in a device weighing only 1.4 g, allowing us to simultaneously record from thousands of neurons in a freely behaving mouse ... For the mouse with the best calcium indicator expression, this extraction yielded 2,948 footprints

Main text, page 546, paragraph 3; Methods, Data analysis, paragraph 1

The session holds 2948 EXTRACT footprints (masks, ids and trace columns agree), within 2948 +/- 295, recorded together in one movie of 36596 frames while the mouse ran 47.7 loop lengths and visited all four quadrants. Same outcome as -01; the criterion no longer asks for duplicate merging or a nonzero trace, and the count is the same.

How the test differed from the paper’s method

• The count is of the EXTRACT output before curation, as the claim states; the curated 2,382 are not marked in the dandiset.

The second design, the Kiloscope, provides cellular resolution imaging across a 4.8 × 3.6 mm FOV

Abstract; main text, page 546, paragraph 3

The file stores no pixel size, so the criterion's two readings from the code apply. Frames of 808 x 608 pixels are 4.800 x 3.612 mm at 5.94 um per pixel and 4.323 x 3.253 mm at 5.35 um per pixel; both sides are within tolerance under both readings, the second only just (4.323 against a lower bound of 4.32). Both pixel sizes come from the paper's code, not the data; with 5.94 um per pixel, 808 pixels give exactly 4.80 mm, so the test under that reading is close to circular. -01 blocked the claim as metadata_insufficient because its criterion admitted only a pixel size from the dandiset and its expected value was 4.0 mm; the rewritten criterion admits the code's pixel sizes.

How the test differed from the paper’s method

• The pixel size comes from the paper's code, not from the dandiset, as the criterion allows.

center of its FOV, which decreases to 7 μm at the edges and is limited by the astigmatism of the GRIN lenses (Fig. 1e and Extended Data Fig. 1a,b)

Main text, page 547, paragraph 1 (continuing from page 546); Fig. 1e

Under reading (a) the median major-axis FWHM is 3.919 um at the centre (11 beads) and 6.433 um in the 500 um group (19 beads), within 10% of 4 and 7 um; reading (c), the mean major axis, gives 3.887 and 6.883 um (6.884 with Jacobian scaling), also within. The minor axis (b, d) is 2.609 to 2.690 um at the centre and 5.203 to 5.244 um at the edge and fails at both. All 155 regions name their image in file-number order. With Jacobian scaling of the optimizer no bead's major FWHM moves by more than 0.5 um, and taking for each bead the lower-cost of the two fits gives the same readings (3.919 and 6.433 um under a). New claim.

How the test differed from the paper’s method

• MATLAB's lsqcurvefit is replaced by scipy's least_squares (trust-region-reflective, the code's start, bounds and tolerances).

Total vignetting (yellow) illustrates the decrease in fluorescent signal caused by the combination of optical vignetting and a Gaussian-shaped excitation light profile. The curve with correction optics (blue) illustrates the flat signal intensity profile that can be achieved with an axicon pair that increases the excitation intensity at the edges of the field of view.

Extended Data Fig. 2b; Methods, Coupling optics, paragraph 3

At both edge readings the corrected profile is higher than total vignetting and falls less from the centre. At 0.5 mm (the code's last bin, 499 um) 0.511 against 0.193, falling by 0.470 against 0.814; at 0.45 mm 0.628 against 0.257, falling by 0.353 against 0.750. It holds for all six centres tried at both radii. The three images are the ones sf3/feather/analysis.m loads in the same roles (its lines 20 to 22, printed in the notebook). -01 found it partially reproduced because its criterion, chosen by its author, asked for the corrected profile to stay above 0.8 out to 0.4 mm; the rewritten criterion compares it with total vignetting, as the legend does. The corrected profile still falls to about half its peak at 0.5 mm.

How the test differed from the paper’s method

• The 0.5 mm reading lies one bin beyond the code's profile (0 to 499 um) and is read at 499 um; the bin at 500 um of an extended profile differs by at most 0.005.

c, same as in (b) but for the Kiloscope.

Extended Data Fig. 2c

At 1.8 mm the corrected profile is 0.557 against 0.180 for total vignetting (falling by 0.437 against 0.843), and at 2.0 mm (the code's last bin, 1999 um) 0.420 against 0.148 (falling by 0.573 against 0.875). It holds at both radii for all six centres tried. The three images are the ones sf3/widefield/analysis.m loads in the same roles (its lines 18 to 20). -01 found it partially reproduced under its own level of 0.8 out to 1.5 mm (0.797 there); the rewritten criterion compares the profile with total vignetting.

How the test differed from the paper’s method

• The 2.0 mm reading is read at the code's last bin, 1999 um; the bin at 2000 um of an extended profile differs by 0.001.

Both microscope designs have sufficiently small field curvature for typical imaging depths (Extended Data Fig. 1c,f).

Main text, page 547, paragraph 1; Extended Data Fig. 1c and 1f

The paper gives no bound on field curvature and no typical imaging depth, as the criterion says. Shown as it asks, the Featherscope's slice of best focus rises from a median of 6 at the centre to 89 in the 500 um group (of 161 slices; Spearman 0.96 with radius) and the Kiloscope's falls from 48 to 29 in the 1500 um group (of 91 slices; Spearman -0.66). If the slices are evenly spaced over the 100 um the Methods give, that is about 52 and 21 um of focal shift. New claim.

Note that the correlated activity visible in each component is driven by footprints from across the FOV (Extended Data Fig. 8).

Main text, page 549, paragraph 1; Extended Data Fig. 8; Methods, Data analysis, paragraph 1

The two event readings give different outcomes. Under (a), each component's largest temporal weight, every event has active footprints in all four quarters (8 of 8, 192 to 326 active). Under (b), the code's times assigned one to one, 7 of 8 do; component 1's event at 1000 s has 151 active footprints and none in the lower-left quarter, which holds 85 of all 2948 footprints. The code's times fall where this NMF's components have 0.04 to 0.68 of their largest weight. With weighted centroids as positions the counts are the same, and with the code's frame window (frames 30 times the times) component 1 again has none in the lower-left quarter. With the quarters split at the image's midlines (row 304.5, column 404.5), the lower-left quarter holds 86 footprints and the counts are again 8 of 8 under (a) and 7 of 8 under (b). Under (b), five components' assigned times fall where their weights are 0.04 to 0.13 of their maximum, so they are not events of this NMF run. -01 found it reproduced under a criterion its author chose (spread of at least half that of all footprints); the rewritten criterion asks for all four quarters under both readings.

How the test differed from the paper’s method

• All 2948 footprints in place of the 2,382 curated ones (the claims file's variant).
• MATLAB's nnmf replaced by scikit-learn's multiplicative-update NMF (100 iterations, 4 random starts).

An examination of individual neural firing rate maps yielded footprints whose activity was localized to subregions of the maze (Extended Data Fig. 10).

Main text, page 549, paragraph 1; Extended Data Fig. 10

The paper selected footprints by eye and neither it nor the code sets a test for localized, as the criterion says. Shown as it asks, over 6411 frames above the 80th percentile of speed the entropy of mean activity over linear position runs from 3.40 to 6.54 bits (median 6.24, largest possible 6.63 over the 99 occupied bins) and the peak of the smoothed map of the fraction of time active from 0.157 to 0.957 (median 0.567). The code's 12 examples index the curated set and cannot be identified. -01 found it reproduced under a shuffle test its author chose; the rewritten criterion has no test.

How the test differed from the paper’s method

• Entropy over the 99 occupied position bins; the code's formula would give 0 for every footprint when a bin is empty.

We trained a support vector machine (SVM) model on half of the data and used the other half to decode the mouse's combined position and movement direction, with the resulting accuracy at least three times chance level (Fig. 2j, P < 0.001, Monte Carlo sampling of n = 1,000 shuffled datasets).

Main text, page 549, paragraph 1; Fig. 2i and 2j; Methods, Data analysis, paragraphs 4 and 5

What the paper reports

• Monte Carlo sampling of shuffled datasets, p < 0.001 (n = 1,000 shuffled datasets)

The criterion judges p under two readings of a null decoder. Under (n1), one fit per shuffle, the claim holds. The average matrix of 200 fits has a median accuracy of 0.437 (3.49 times chance), none of 1000 shuffled decoders reaches it (largest 0.227, p < 0.001), accuracy reading (a) holds and (b) does not (classes 2- and 3- at 0.294 and 0.239). The (n1) null is of single fits; the real single fits have median accuracies of 0.138 to 0.604 (median 0.430), 2 of 200 at or below the largest null value, and the weakest has 78 of 1000 null values at or above it. Reading (n2), 200 fits per shuffle, needs 200,000 fits, measured at 33.3 hours in this run (24.6 hours in the run reviewed in round 2) against a 30 minute limit per notebook. Nine (n2) null decoders were run (one in the decoding notebook, 0.136; eight in decoding_null_n2_quadrant_direction.ipynb, 0.089 to 0.145), none near 0.437, but nine cannot show p < 0.001. The clock check of the main notebook peaks with labels shifted by +2 s (0.519 against 0.495 at 0 s, 3 fits each); an offset would lower the accuracy, not raise it. The claim was dropped from the claims file of -01, which had no outcome for it.

How the test differed from the paper’s method

• All 2948 footprints in place of the 2,382 curated ones (the claims file's variant); the curated set is not marked.
• Position tracked from the behaviour video with the code's rule; linear position unwrapped by the loop length.
• MATLAB's fitcecoc with templateLinear replaced by scikit-learn's SGDClassifier (hinge loss, alpha equal to lambda, 10 passes), as MATLAB's default for more than 100 predictors is 10 passes of dual SGD.

The accuracy of the decoder depended on the number of footprints used (Fig. 2k, P ≤ 2.7 × 10−4 for all decoder sizes, Mann–Whitney U).

Main text, page 549, paragraph 1; Fig. 2k; Methods, Data analysis, paragraph 5

What the paper reports

• printed as P ≤ 2.7 × 10−4 for all decoder sizes; the schema has no "at or below": Mann-Whitney U, p < 2.7 × 10⁻⁴ (n = 64 decoders per size, sizes 2, 4, ..., 2,048 footprints)

It depends on fig2-quadrant-direction-decoding, which is blocked as not_runnable, and it needs 140,800 fits (11 sizes, 64 subsets, 100 fits on the original and 100 on the shuffled data), measured at 4.8 hours on four cores in this run (3.0 hours in the run reviewed in round 2). Not run in full. A reduced run chosen here (16 subsets of 5 fits per size, decoding_reduced.ipynb), which decides nothing, has the median accuracy rising from 0.094 at 2 footprints to 0.442 at 2,048; original and shuffled subsets differ with p of 1.5e-6 to 9.5e-6 from 32 footprints on but not at 2 to 8 (p 0.376 to 0.955) nor, at 2.7e-4, at 16 (p 0.0205), and neither reading has every p at or below 2.7e-4. Its values differ between runs on different machines (in the run of round 2, 0.101 at 2 footprints and p of 0.00213 at 16), with the same pattern. New claim.

Notably, footprints in different regions of the FOV yielded substantially different decoding accuracies (Fig. 2l).

Main text, page 549, paragraph 1; Fig. 2l; Methods, Data analysis, paragraph 5

It depends on fig2-quadrant-direction-decoding, which is blocked as not_runnable. The grid has 5411 used points at 45 pixels and 5905 at 50 (5409 and 5897 with centroids); with 10 fits per point and per null decoder that is 64,090 to 69,050 fits, measured at 4.4 to 4.8 hours on four cores in this run (2.7 to 2.9 hours in the run reviewed in round 2). Not run in full. A reduced run chosen here (a grid every 50 pixels with 56 points used, 500 null decoders of one fit, decoding_reduced.ipynb), which decides nothing, gives point values spanning 0.089 to 0.282 with 32 of 56 points significant (31 with (k + 1) / (n + 1)); in the run of round 2 they spanned 0.097 to 0.282, with 32 significant under both definitions of p. New claim.

We used the same approach to decode even more fine-grained position information (Extended Data Fig. 9, P < 0.001, Monte Carlo sampling as in Fig. 2j).

Main text, page 549, paragraph 1; Extended Data Fig. 9a and 9b

What the paper reports

• Monte Carlo sampling of shuffled datasets, p < 0.001 (n = 1,000 shuffled datasets (as in Fig. 2j))

Under null reading (n1) the claim holds. The average matrix of 200 fits has a median accuracy of 0.326 (2.61 times chance) and none of 1000 shuffled decoders reaches it (largest 0.229, p < 0.001). The real single fits have median accuracies of 0.179 to 0.417 (median 0.320), 11 of 200 at or below the largest (n1) null value, and the weakest has 12 of 1000 null values at or above it. Reading (n2) needs 200,000 fits, measured at 33.7 hours in this run (25.0 hours in the run reviewed in round 2). Nine (n2) null decoders were run (one in the decoding notebook, 0.110; eight in decoding_null_n2_octant.ipynb, 0.109 to 0.149), none near 0.326, but nine cannot show p < 0.001. The criterion asks for both. -01 found it reproduced against decoders with shifted labels; the rewritten claims file uses the paper's null, shuffled activity, with two readings of a null decoder, one of which cannot run here.

How the test differed from the paper’s method

• All 2948 footprints in place of the 2,382 curated ones; position tracked from the video; SGD in place of MATLAB's solver, as for fig2-quadrant-direction-decoding.

Not quoted: The Extended Data Fig. 9c legend and the Methods give the method only. The main text states that accuracy differs across the field of view for the quadrant and direction map (Fig. 2l) and that octant position was decoded with the same approach, but states no result for the octant map. That it is significant in parts of the field of view and varies across it is read from the legend (non-significant pixels plotted in black) and the main text.

It depends on ed9b-octant-decoding, which is blocked as not_runnable. The grid has 5411 used points at 45 pixels and 5905 at 50; with 40 fits per point and per null decoder that is 256,360 to 276,200 fits, measured at 11.9 to 12.8 hours on four cores in this run (6.9 to 7.4 hours in the run reviewed in round 2). Not run in full. A reduced run chosen here (as for fig2l-local-decoding-map, with 10 fits per point in place of 40, decoding_reduced.ipynb), which decides nothing, gives point values spanning 0.087 to 0.195 with 7 of 56 points significant under both definitions of p, so (i) holds in it. In the run reviewed in round 2 the values spanned 0.090 to 0.197 and no point was significant, so (i) failed there; with seeds fixed, the reduced run's result changes between runs on different machines. -01 found it reproduced on a 30 pixel grid with 10 decoders per point and a null of shifted labels; the rewritten criterion uses the code's 5 pixel grid, 40 fits and the paper's null of a thousand decoders on 32 random shuffled footprints.

Lateral shifts of the image due to mouse motion were small (root-mean-square displacement of 0.13 pixels or 0.7 μm (Extended Data Fig. 7).

Main text, page 548, paragraph 2; Extended Data Fig. 7

Rigid registration of all 36596 frames gives an RMS displacement of 0.0367 px under reading (a), 0.0381 px under (b) and 0.0260 px under (c), against 0.13 +/- 0.013 px; under (a) 0.218 um at 5.94 um per pixel and 0.196 um at 5.35, against 0.7 um. No frame moves more than 0.16 px along an axis, and the RMS in each 100 s window is 0.028 to 0.051 px. With a jitter of 0.132 px RMS added to the first 1200 frames, the estimated shifts follow it (correlation 0.996 on each axis) and their RMS rises from 0.033 to 0.136 px, so the registration finds motion of the paper's size. Nothing in the dandiset says whether the movie is the raw video or a motion-corrected one, and it holds no motion-correction module; its values are consistent with the raw video. The histogram of its 10-bit counts has a roughness (mean relative difference of each value's count from its neighbours') of 0.740 to 0.745 in every block, where moving frames by sub-pixel shifts of 0.13 px RMS and writing them back as integers lowers it from 0.745 to 0.052; its zero values are at most 2 single pixels in 243 of 36596 frames, not edge strips. A variant with high-pass filtered frames (chosen here, first 2400 frames) gives 0.0188 px under (a) against 0.0328 px for the main run on the same frames, so a static background pulling the estimate toward zero does not explain the difference. New claim.

How the test differed from the paper’s method

• NoRMCorre (MATLAB) is replaced by a rigid registration written after its algorithm and defaults (dftregistration with 1/50 pixel upsampling, a template of the 100 middle frames, updated every 200 frames); the paper's settings are not given.