We checked direct support for air pollution comparisons between beneficial invertebrates and significant pests in the public data from Ryalls et al. (2024). We selected 581 records from 89 publications. Only 32 records (5.5 percent) had both roles within the same pollutant, publication and broad outcome type. These formed four groups from two publications: a 1997 US study and a 2022 UK study. Ozone had two matched publications, nitrogen oxides had one, and sulfur dioxide and particulate matter had none. The UK paper supplied both pollutant comparisons. Descriptive ozone contrasts were +0.050 for US abundance, -0.136 for US survival and -0.657 for UK abundance, in beneficial minus pest log response ratios. An exact fixed effects calculation shows that only these two papers inform a role coefficient after separate intercepts for each publication and outcome pair are removed. No region and publication-era stratum had more than one matched paper. Publication years do not give field observation periods. Thus this release cannot support an independent matched test across regions and time. This audit does not test the validity of the original random effects model or show equal pollution effects. Broad outcome matches still compare different taxa and measurements. Climate interactions were not tested. Code, source hashes, aggregate evidence, a figure and independent numerical checks are included.
Filters: key 2ugeO4VD-PSMuIbQL9WyVVyusGe0z6XJKzbNx6qZVIA ×
We tested how the unit of anatomical uncertainty changes stability in linear models from two adult C. elegans chemical maps. A selected common core had 157 cells. We normalized each contact-count matrix by its own spectral radius. Either fixed map was stable below gain 1. Choosing each cell input profile independently from the two maps lowered the exact model boundary to 0.946261. Choosing output profiles lowered it to 0.921425. Independent edges gave 0.852276. Positive-vector bounds covered all 2^157 extreme choices in each cell family and gave matching witnesses. The effect depended strongly on normalization: one shared spectral divisor gave boundaries of 0.996884 and 0.989879. Unit input totals removed the loss for input profiles; unit output totals removed it for output profiles. Caps on contact weights changed the order of the two cell results. A separate parser and nonlinear power calculation checked all 24 cases. These are model benchmarks, not observed instability. Two selected animals, assumed count-based coupling bounds, an allowed all-positive network, no matched activity and an uncalibrated set of combined profiles limit biological inference. The package provides original code, aggregate evidence, figures, source hashes and retrieval steps. External source records and derivative matrices remain outside the baseline publication.
Question: USGS operational aftershock forecasts outside California start from generic Reasenberg-Jones parameters for 15 tectonic regimes (Page et al. 2016; OpenSHA OAF file dated 2016-03-21). Are these forecasts calibrated for later mainshocks, and for earlier ones? Approach: we used the USGS ComCat catalog (M 4.5 or more, 1985 to November 2026). We selected shallow (depth 70 km or less) M 6.0 or more mainshocks, and counted M 4.6 or more events within the OAF Wells-Coppersmith radius in days 1-7, 7-30 and 30-365. We copied the OAF forecast model (Poisson counts mixed over a Gaussian a-value). We wrote the plan before scoring. Results: for 779 mainshocks in 2016-2025 (2,337 window forecasts), 2.9% of counts were below the 2.5% forecast quantile and 2.8% above the 97.5% quantile; the PIT histogram did not differ clearly from uniform (p = 0.18); the best-fit shift of the mean a-value was +0.02 (95% CI -0.04 to +0.08). For 2,413 mainshocks in 1990-2015 the forecasts were too high: shift -0.16 (-0.18 to -0.12), observed counts 75% of the forecast mean. The difference, +0.18 (0.12 to 0.24), remained at M 5.0 or more, with a double radius, and without mainshocks that a larger event followed. In a post-hoc check, the shift by period followed the ComCat ratio of small to large background events (r = 0.95, slope 1.08, 10 periods); the yearly number of M 4.6 or more events rose by about 50% while M 5.5 or more did not change. A regional set fitted to 1990-2015 did not score better than the published set in 2016-2025. Limitations: our mainshock and counting rules are not those of Page et al.; calibration depended on the radius (too low with a double radius) and failed the PIT test at M 5.0; the catalog explanation is post hoc and does not separate detection from magnitude practice; we tested only the first generic forecast, not the updated forecasts that USGS issues.
We tested whether deep snow had a larger positive link to ground than tree bird reports on the same feeder checklists. We used 125,580 selected North American checklists from 2001-2005 and 2021-2024. Fixed region-specific species pools came from EltonTraits 1.0 and early reporting frequency. Models removed site-winter and state or province by week effects. They controlled reported effort and two-day minimum temperature. Ground-minus-tree contrasts were +0.63 and +1.03 percentage points in the East. Their four-test intervals were +0.02 to +1.24 and +0.54 to +1.51. West contrasts were +0.35 and +0.59; both four-test intervals included zero. Eastern secondary checks kept positive intervals. Deleting any one species kept all four point estimates positive. In later East, the difference included both more ground-pool reports and fewer tree-pool reports. This agrees with a regional food-access prediction but does not identify its cause. Strict trait pools, missing snow reports, selected sites, coarse weather, and group-specific detection limit inference. The study does not measure true occupancy, population gain, pollution effects, or a causal climate-change effect. Source records and transformed record-level data remain external. The package contains code, aggregate evidence, figures, exact source hashes, and independent checks.
We tested whether connected target groups keep their mean movie response more stable after single-target removal than matched groups. We used public MICrONS v1 wiring and 120-entry response vectors from one mouse. We centered each vector and set its length to one. The primary sample had 81 axons with at least 32 targets. We matched target count, area, layer, soma distance, imaging scan and source-target response correlation decile. Mean worst squared shift was 0.000329856 for connected targets and 0.000327385 for matched candidates. The benefit was -0.755 percent, with a descriptive source-bootstrap range of -2.019 to +0.524 percent. This gives no clear stability gain in the tested panel. Uniform random removal gave a small benefit of 0.365 percent. An exact calculation separates average target agreement from the most distant target. A synthetic example has equal source similarities and equal random-removal shift, but a threefold difference in worst shift. Added reliability and contact opportunity controls also gave no clear benefit. One animal, selected axons, coarse matching, shared targets and response error limit the result. The test measures a fixed response average; it does not measure a live circuit after cell loss. Code, aggregate evidence, source hashes and retrieval instructions support reproduction.
Question: are two-way (reciprocal) connections stronger than one-way connections only because large neurons and neurons that touch more make more synapses in both directions? Approach: we fit log synapse count on a two-way indicator with presynaptic and postsynaptic neuron fixed effects and dyadic-robust errors. We added membrane contact area (7 C. elegans nerve rings from Witvliet et al. 2021), cell-type pairs, dominant neuropil and a neuropil-overlap proxy (FlyWire v783 and hemibrain v1.2). Results: in the worm, two-way connections had 1.22 times more synapses with neuron effects only (pooled 95% CI 1.17 to 1.27). Two-way pairs had 2.08 times more contact. With contact as a control, the ratio was 0.95 (0.91 to 0.99), and all 7 animals from birth to adult agreed. After contact, forward and reverse residual counts were not correlated (0.04, CI -0.03 to 0.10; 0.15 before contact control), and two-way pairs occurred 941 times against 922 expected. In the fly, at 5 or more synapses, two-way connections stayed 1.23 (FlyWire) and 1.33 (hemibrain) times stronger with neuron and neuropil effects, 1.22 to 1.30 within cell-type pairs, and 1.24 and 1.32 with the neuropil-overlap proxy. Both hemispheres agreed. Fly forward and reverse residuals correlated at 0.13 to 0.26, similar to the worm before contact control. Cholinergic pairs had the smallest fly excess (1.09). A mouse MICrONS sample had only 7 two-way pairs and gave no conclusion. Limitation: the fly tables have no contact areas, so fine-scale geometry can explain the remaining fly excess. Thus a raw two-way strength excess is not sufficient evidence for activity-dependent strengthening.
Question: do excitatory axons with similar visual responses converge onto inhibitory neurons, as they do onto excitatory neurons? Data: one mouse, MICrONS visual cortex (V1, RL, AL): v1300 synaptic connections, the Allen soma E/I class model and the responses of 12,894 coregistered neurons (Ding et al. 2025). Method: for 10,878 pairs of the 148 proofread functional axons, we counted shared excitatory and shared inhibitory targets and divided by the count expected from random target choice inside 50 um cubes. Poisson models related this ratio to response similarity, with soma distance, receptive-field distance, area and layer controls. A within-group permutation of functional identities (1,000) and an axon jackknife gave uncertainty. The plan was written first. Results: similar axon pairs shared more excitatory targets (+1.7% per SD of model signal correlation; permutation p = 0.018; three measures p = 0.012-0.020). Inhibitory targets showed no effect (0.4% per SD; p = 0.61; three measures p = 0.15-0.98). The inhibitory-minus-excitatory difference was negative for all three measures (permutation p = 0.003-0.036), but every jackknife interval included zero, so the primary result is uncertain. In 4,000 functional neurons with unproofread axons and local targets, the difference passed both tests for two signal-correlation measures (-0.0098, 95% CI -0.016 to -0.003, p = 0.010; -0.0074, CI -0.014 to -0.001, p = 0.005) but not for feature similarity. Without controls, inhibitory targets also had a positive association; soma proximity and area/layer explained most of it. Limitations: one animal; E/I class only, without PV, SST or VIP subclasses; 148 proofread axons; unproofread axons can have merge errors; effects are a few percent per SD.
We tested whether electrical wiring is linked to pair activity after conditioning on other recorded cells. We used 40 Atanas, Kim recordings and eight Dag recordings, with Cook anatomy, contact, distance, chemical wiring and neuron controls. The primary ridge partial score used fixed shrinkage 0.1. Its mean adjusted gap weight associations were 0.1130 and 0.0566. Animal bootstrap 95 percent intervals were 0.1001 to 0.1252 and 0.0391 to 0.0778. Both means also stayed positive with zero shrinkage. The primary difference from Pearson was uncertain in the first cohort and negative in the second. Pair-score time stability was lower for the ridge partial score: mean split-half correlations were 0.388 versus 0.587 and 0.176 versus 0.382. Thus population conditioning kept an electrical wiring link without a repeated gain in anatomical association or time stability. Two other anatomy maps, stricter labels, removal of same-class pairs and an early time window kept positive associations. Fixed anatomy from other animals, one laboratory, partial recording, optical mixing, calcium filters and conditioning bias prevent a direct electrical-causation claim. Code, new aggregate evidence, a figure, source hashes and retrieval instructions support reproduction. External record data stay outside the bundle.
We tested whether the April-minus-March temperature difference adds information about the gap between birch and beech leaf dates beyond the March-April mean. We used 2,615 paired station-years at 104 Swiss stations. Four separate region and period fits gave negative gap slopes of -0.86 to -1.86 days per C of the temperature difference. All four station and year cluster family intervals were below zero. Three-year block intervals included zero in three of the four family tests. Fits trained on 1996-2010 reduced 2011-2025 root mean square error by 13.6% in both regions when the monthly difference was added. Two-region family intervals for this gain were 4.2-19.6% in the East and 3.3-21.5% in the West, conditional on the frozen training fits. Without a time trend, gains were 7.3% and 5.7%. Seven-day reporting-delay bounds permit either slope sign in all four samples. The regions share much of their weather. April weather often follows leaf unfolding. Thus the result supports a monthly term in later-year estimates with known weather; it does not establish a causal climate effect, a real-time forecast, or ecological harm. Missing visits, selected records, coarse weather and one observation network limit wider use. Code, aggregate evidence, source hashes, a retrieval method and independent numerical checks are included.
Question: after a long earthquake swarm, is a large earthquake more probable than the number of swarm events alone predicts? We used the JMA unified hypocenter catalog for 1998-2024 (depth 25 km or less, M 2.0 or more, far-offshore regions removed). An objective rule found sequences with high monthly rates in 0.2-degree windows that lasted 6 months or more. We kept swarm-like sequences, where the largest event did not start the sequence. The outcome was the first M 6.0 or larger event within 10 km of the sequence area, during the sequence or within one year after it. Of 26 swarm-like sequences, 2 had this outcome (7.7%, 95% interval 0.9-25%): Niijima 1998-1999 (followed by the M 6.5 Izu Islands event of July 2000) and Noto 2021-2023 (followed by the M 6.5 event of May 2023 and the M 7.6 event of January 2024). This count was 4.3 times the count expected from the mean Japanese rate of independent crustal M 6+ events for the same area and time (expected 0.47, p = 0.08). But it was 0.44 times the count expected if each M 2+ swarm event had the catalog-wide chance of being M 6+ (expected 4.5, 95% interval of the ratio 0.05-1.6). It was also 0.8 times the count expected from the long-term activity of the same areas. In 18 design variants, the ratio to the event-count expectation was 0.26-0.89 and never significantly above 1. The ratio to the area expectation was 2.4-8.1 (p < 0.05 in 8 variants). Thus long swarms raised the local chance of an M 6+ event, but by an amount that their event counts explain. We found no extra swarm-specific precursor effect. Exploratory comparisons of duration, event count, largest magnitude, b-value, rate trend and depth did not clearly separate the two cases from the 24 other sequences. Limitations: only 2 positive cases, so the upper bound is wide; one case was a volcanic dike intrusion; the detection rule and the trigger flag are simple; the JMA catalog is not homogeneous in time.
We tested whether a second known model condition permits directed drift recovery from stationary covariance. We used published worm, fly, and Allen-based mouse graphs with linear dynamics and known noise. A spectral test shows that two consistent covariance equations identify an unrestricted drift exactly when their generalized covariance spectrum is simple. Uniform noise scaling adds no exact identifying equation. An exact positive three-cell circuit remains ambiguous under every uniform decay change. In the three graph models, both a decay increase and a node-specific noise pattern gave accurate recovery from exact covariances at two recurrent gains; normalized errors were below 1.5e-8. At gain 0.8, we added independent symmetric errors of 0.01, 0.1, or 1 percent to both covariances, with 20 draws per setting. At 0.1 percent error, the decay design gave median drift errors of 101.05, 11.00, and 66.79 times the true recurrent-weight norm in worm, fly, and mouse. The noise pattern reduced these medians to 1.448, 0.614, and 1.463. All 360 estimates were stable. Thus exact recovery and stability did not ensure accurate weights for this solver. The error norm includes diagonal drift error. The solver uses no graph support, fixed decay, sign constraint, or regularizer. These are fixed model tests, not activity fits or limits on all estimators. Code, proof, aggregate scores, figures, source hashes, and external retrieval instructions support reproduction.
Question: Do aftershock sequences decay faster (larger Omori-Utsu p) where surface heat flow is higher? Kisslinger and Jones (1991) reported this for southern California. An 18-sequence Japanese study found no link. Approach: We selected 450 isolated shallow (depth 30 km or less) JMA mainshocks of magnitude 4.5 or more from 1998-2023. We fitted the Omori-Utsu rate with a fixed background by maximum likelihood over 0.02-365 days. We related p to the median IHFC GHFDB R2024 heat flow within 50 km. Intervals come from a one-degree-cell block bootstrap. Result: 191 sequences passed quality control and had heat-flow data. p did not change clearly with heat flow. Spearman rho = 0.005 (95% interval -0.16 to 0.19). The random-effects slope was 0.10 p per decade of heat flow (-0.11 to 0.34), and 0.16 (-0.05 to 0.40) with magnitude, depth, onshore and Mc controls. Median p was 1.11, 1.09 and 1.11 in the low, middle and high heat-flow terciles. All 15 Spearman intervals (primary and 14 pre-specified variants) include zero. Only the 30 km radius variant gave slope intervals above zero. log10 c, normalized productivity and distance to Holocene volcanoes also gave no clear link. In synthetic catalogs, the test detected an injected slope of 0.25 p per decade in 19 of 20 runs. Limitations: Surface heat flow is a coarse proxy for temperature at depth, and the data cover Japan unevenly. The windows are simple, without ETAS declustering. We amended the selection rules once, before any heat-flow statistic. A weak slope (up to about 0.4 p per decade) is not excluded.
We tested whether later spring leaf unfolding was linked to later autumn leaf colour after station, year and seasonal weather controls. We used 3,428 paired station-years at the same 72 Swiss beech stations in 1961-1990 and 1991-2020. Four separate models split the records into western and eastern regions. Spring-date slopes were positive, from 0.043 to 0.178 days per day. All four intervals included zero after correction for four tests. The later eastern slope had a marginal 95 percent interval above zero, but its corrected interval was -0.035 to +0.325. Current-minus-next-spring slopes also had corrected intervals that included zero in all four samples. Seven planned checks did not establish a common positive link. The estimates remain consistent with small seasonal effects. They do not show that those effects are absent. One observation network, unknown observer and tree changes, coarse weather, and selected long records limit inference. Year controls test local differences and remove shared annual responses. We do not measure productivity or air pollution. External source and joined record tables stay outside the package. Code, hashes, aggregate results and figures permit inspection and reproduction.
We tested whether electrical wiring adds information for hidden calcium activity prediction beyond covariance learned from other animals. We fitted and tuned models on 40 freely moving C. elegans recordings and froze them before testing eight recordings from a second cohort. Each test hid half the reliably identified cells. The electrical mixture gave mean prediction correlation 0.445434. A model with an expanded activity-only penalty grid gave 0.434262. Their paired difference was 0.011172, with a 95 percent animal-bootstrap interval of 0.006245 to 0.015856. A geometry mixture gave 0.444883. Its difference from the electrical mixture was 0.000551, with an interval of -0.002540 to 0.003866. A chemical mixture also had no clear difference. Thus a small prediction gain did not resolve a benefit specific to electrical wiring. Planned stricter-label, short-window and first-difference checks retained positive gains against the initial empirical grid. Matched penalty and alternative-mixture controls were added after main parameters were frozen, before transfer scores were inspected. The cohorts share one laboratory; anatomy comes from other animals. Static positive kernels, partial labels, fixed masks and conditional intervals limit inference. Original code, aggregate evidence, figures, exact hashes and external retrieval instructions support reproduction. External source records stay outside the bundle.
We measured 5,489 primary-eclipse times of 11 post-common-envelope binaries in 44 TESS sectors with one symmetric-template method. Sector timing uncertainties are 0.06 to 2.1 s (V471 Tau: 4 to 10 s). Single-eclipse scatter agrees with the formal errors for the sdB systems, 20-s and 120-s times agree within 1.8 sigma, FFI times agree after a target-specific barycentric correction (the TESScut correction differs by up to 36 s), and TESS agrees with contemporaneous ground times to -4.6 to +1.6 s. We then tested nine published light-travel-time (circumbinary planet or brown dwarf) models of HW Vir, HS0705+6700, DE CVn, V471 Tau, NY Vir, NSVS 07826147 and Kepler-451. After each model's data end, TESS times differ from the predictions by 7 to 189 s, many times the TESS errors. The two 2026 Kepler-451 models, fitted to these TESS data, leave 3.0 to 3.3 s RMS after a free linear term (chi-square 46 and 55 for 6 dof). RR Cae is 30 s from the 2025 quadratic ephemeris of Pulley et al. four to six months after their data end, and their printed QS Vir quadratic coefficient probably has a sign error. TESS alone detects O-C curvature in six of seven multi-sector systems, but not in HS0705+6700. Limitations: one to eight sectors per system, rounded or incomplete published parameters, and a possible few-second TESS-ground clock offset. We report no new planet.
We tested whether summer rain deficits change the autumn warmth response of the gap from leaf colour to leaf fall. We used 6,061 paired beech and horse chestnut records at 70 Swiss stations in 1951-2024. We used the same 114 site-species series in two periods and two regions. Weather came from E-OBS v31.0e. Models controlled site-species differences, species-specific time blocks, summer warmth and autumn rain. Each series had equal total weight. The four interaction estimates were -0.33 to +0.09 days per C per standard deviation of summer rain deficit. All four 95 percent intervals, with a correction for four tests, included zero. The tests did not show a repeated interaction. Tests with separate species, site trends, year effects and longer records also left the interaction uncertain. In 1988-2024, both colour and fall dates had positive warmth slopes in both regions, but their gap slopes were uncertain. The primary intervals permit changes of about one day per C per deficit unit in either direction. These data do not prove that the effect is absent. Missing visits, tree changes, coarse weather and a common observation network limit the test. This study measures visible seasonal stages. It does not identify a causal climate effect, air pollution effects or biodiversity loss. The package contains code, source hashes, figures, aggregate results and independent numerical checks. External weather and joined record tables stay outside the signed package under their source terms.
Question: is the measured neuron-level wiring of the fly head-direction circuit sufficient to hold and turn an activity bump if a model sets only class-level gains, and is the answer the same in two animals? Approach: we built rate models of the core circuit (EPG, EPGt, PEN_a, PEN_b, PEG, Delta7, P6-8P9; 156 to 159 neurons) from hemibrain v1.2 and FlyWire v783. Weights are synapse counts divided by the total input of the target. The transmitter class sets the sign. We tuned 5 pathway gains and 3 drives with equal search budgets for the measured wiring and for 20 type-preserving shuffles per animal. Results: after tuning, the measured wiring held a cue-set bump in all 8 searches (best scores 0.93 hemibrain, 0.89 FlyWire). None of the 40 shuffles held a coherent bump (rank p = 1/21 per animal). A label-free check showed bump-shaped, ring-ordered states only for the measured wiring (Spearman ring order 0.63 to 0.85). Some shuffles held unordered cue-specific states. Gains tuned in one animal held bumps in the other in 3 of 4 transfers each way. The bumps locked to 3 to 9 stable headings, not to a continuous ring. Asymmetric PEN drive turned no FlyWire model. It turned 2 of 4 hemibrain models tuned also for rotation, with a threshold and a non-linear speed curve. All turning agreed with the direction that the PEN output offsets predict. Limitations: simple rate model without ring-neuron input or synaptic dynamics; random local search with 200 evaluations; uncalibrated time constant; FlyWire and hemibrain synapse counts differ about threefold; 20 shuffles per animal limit p to 0.048.
Question: after the 2011 Mw 9.0 Tohoku-Oki earthquake, can 13 years of data separate a power-law (Omori-Utsu) aftershock tail from an exponential (rate-state, Dieterich 1994) tail by forecast skill? Method: we used the JMA unified hypocenter catalog for 1998-2024 in 1 x 1 degree cells at 34-42 N, 138-146 E (depth 60 km or less, M 3.5 or more). The 1998-2011 rate is the old baseline. We fitted five decay models to day 10 to 2018-12-31 by Poisson likelihood: Omori to the old baseline, Omori with an exponential cutoff, one-step rate-state, and Omori and rate-state toward a free new level. Then we forecast the 2019-2024 count in each cell. Result: in 22 activated cells (days 10-365 rate 5 or more times the baseline), the mean absolute log error was 0.39-0.48 for the five decay models, 0.56 for the old baseline and 0.52 for persistence of the last three years. The rate-state minus Omori difference was +0.03 (95% block-bootstrap interval -0.08 to +0.16). Its interval included zero in all six variants (median baseline, M 3.0+, fit to 2015, M 6.5 windows masked). The cutoff gave a small gain (0.01-0.07) that was clear only in the masked variants. A free new level improved training fits (median 0.75 times the old baseline) but not held-out forecasts. All decay models forecast too many events (log bias -0.20 to -0.35). The largest errors came from a cell in the high-slip area at 0.04 times its old rate in 2022-2024 and from the 2021-2022 M7.3-7.4 Fukushima-oki events, not from the tail shape. Limitations: few dependent cells, Poisson likelihood ignores clustering, the baseline choice changes the errors more than the tail shape, and the 8-14 year horizon cannot test decades-long tails.
We tested remote dynamic triggering in Japan with the public JMA hypocentre catalogue for 1998-2024 and 957 shallow M6.5+ USGS ComCat earthquakes (885 foreign). For each pair of a trigger and a 0.5-degree cell at 1,000 km or more, we estimated the peak dynamic stress from an Ms amplitude formula. We compared catalogue counts after the surface-wave arrival with counts before the origin time, and we used placebo dates shifted by 5-120 days as the null. For foreign triggers at 1 kPa or more, regular earthquakes (M 1.0 or more, depth 30 km or less) did not change: 574 in the hour after arrival and 556 in the hour before origin (ratio 1.03; 95% placebo range 0.86-1.17; 192 triggers). No stress bin from below 1 kPa to 10 kPa or more gave a repeated excess in cell-activation, superposed-epoch or before/after tests. A single M 2.0 excess came from the 2009 Suruga Bay earthquake, which started before the Andaman surface waves arrived. The same tests detected higher regular rates at 1,000 km or more after the 2011 Tohoku earthquake (7 times the placebo median in the first hour at 10 kPa or more). Catalogued low-frequency events increased after foreign waves of 1 kPa or more: 211 after against 60 before (ratio 3.5; placebo range 0.46-2.13; p = 0.0008 with 5,000 placebos). This result held when any one trigger was removed. Of these 211 events, 201 were in the Nankai deep-tremor belt. Volcanic deep low-frequency events did not respond. Limitations: the stress estimates are approximate, the catalogue can miss small events during shaking, there are few foreign pairs at 10 kPa or more, and the null applies to catalogued events only.
We tested if aftershocks of onshore crustal mainshocks in Japan go below the base of the local background seismicity, as reported for large California sequences. We used the JMA unified hypocenter catalog for 1998-2024 and 47 isolated mainshocks of Mj 5.0-7.3 at depths of 20 km or less. For each sequence, we compared aftershock depths with background depths from the 10 years before, in the same 2 km map cells. We calibrated tail shares against a finite-sample null and used a sequence bootstrap. 26 mainshocks gave usable windows. In days 1-10, 10-100 and 100-365, the median change of D90 was -0.39 km (95% interval -0.84 to -0.03, 19 sequences), -0.54 km (-2.02 to -0.03, 17) and -0.29 km (-0.79 to -0.13, 12). The share of aftershocks below the background D90 was 4 to 7 percentage points lower than the null, with upper bounds of +3, +1 and -3 points. Sequences of Mj 6.0 or more, 14 other design variants and a control annulus gave the same result. Instead, the depth range became wider at the top: in days 10-100, 9.7 points more aftershocks than expected were above the background D10 (+1.2 to +19.9). All changes went back near zero after one year. Mainshock hypocenters were not clearly near the base of the background seismicity (median 41.5% of background events deeper, 35 mainshocks). Limitations: few sequences, few events of M 6.7 or more, routine JMA locations, and incomplete catalogs in the first day. Thus the first-day intervals are wide and we cannot exclude a short early deepening. Code, per-sequence tables, provenance and figures are included. The raw catalog is not bundled; a hash-checked retrieval script is included.