Tsuruta Laboratory

Facts

How the notes are written

• Every note states where its primary data came from, on what date, and at what version.
• Notes that involve computation ship the code, the raw result files and the commands that produced them.
• A hypothesis that was rejected is published as rejected.
• When a published note turns out to be wrong it is corrected in place, and the correction and its date stay on the page. Notes are not withdrawn.
• References the author has not held in hand are marked as unverified.

Research notes

• 2026-09-28 How many cores can a Minecraft server use? Chunk generation and mob work on the GPU, and Paper vs Folia vs ShreddedPaper

  Four things were measured on one PC running Minecraft Java Edition 26.1.2. First, chunk generation: vanilla used about one CPU core and made 22.8 chunks per second. C2ME spread the work over about 11.6 cores for a 15.1x speed-up, and its GPU add-on raised that to 20.9x. Second, mob pushing and pathfinding: at normal mob counts the GPU did not beat the CPU. Third, multithreaded servers: Folia put players within 1,536 blocks of each other into one region, and when players clustered its TPS fell to 0.25, slower than Paper. ShreddedPaper kept TPS above 17 even with players clustered. Fourth, the upper bound for giving each kind of work its own core: villager brain logic took about 70% of the main thread, which caps the speed-up at about 1.5x.

• 2026-09-20 A comparison of the bias of three interaction estimators in a cost-structure model of self-disclosure

  People are thought to confide more readily in a listener with whom they share many mutual friends, and more readily still when the relationship is unlikely to continue. Rather than test this on people first, an agent model was written that contains only the local cost structure: the benefit of confiding scales with the number of mutual friends, and the cost is the maintenance obligation incurred multiplied by the probability that the obligation is realised. The hypothesis itself was never written into the generative rules. Thirty-two conditions were run, with the obligation weight set to zero as a control, since at that setting any continuation effect must be exactly zero by definition. Folding mutual friends into bins and taking a difference in differences produced +0.10 to +0.13 at that control setting, a quantity that cannot exist. The cause was not network structure but composition: within a bin, the residual variation in mutual friends correlated with continuation, and within rows where mutual friends were pinned to a single integer the gradient vanished entirely. Conditioning on the exact integer instead dropped the control bias to within 0.005, and the hypothesis then held, with the low-capacity groups showing 2.02 times the continuation effect of the high-capacity groups. The separating individual difference is maintenance capacity, not approach motivation, even though both produce the same visible outcome of having few friends. Degree-preserving nulls do not detect the bias, because it arises from the coarseness of the measured variable rather than from the graph.

• 2026-09-20 Conditions under which the pinned-common-neighbour estimator exists, across five substrate families and a degree-preserving control

  The previous note reached its conclusion with the acquaintance substrate fixed to a single Watts-Strogatz construction. Five substrate families, all matched at mean degree 10, were substituted in. In three of them the estimator did not exist: almost no pair shares two or more mutual friends, so the estimator has no denominator, with usable encounters falling from 26.1% to 0.01%, around 20 per run. The decisive case is a configuration model built to carry nearly the same degree sequence as the Watts-Strogatz substrate, a standard deviation of 0.99 against 0.97, while clustering alone drops from 0.494 to 0.008. It collapsed. The condition for the design to work is clustering, not degree. In the two families where the estimator did exist, the control bias was at most 0.004 and the capacity separation was 2.02 and 2.46, agreeing in both direction and magnitude despite different mechanisms generating the clustering. Separately, sarcasm, which in the previous note failed to separate by capacity at all with a ratio of 0.97, separated at 2.45 once its cost was charged against the actor own capacity pressure, with no free parameter added. Any application to real data should first measure the share of encounters at which mutual friends are determined, and abandon the design if it falls below threshold.

• 2026-09-20 Identifiability of the estrogen-driven senescence coefficient in a proliferative-exhaustion model of the growth plate

  A proliferative-exhaustion model of the growth plate was built from three constraints taken from the literature: longitudinal growth is limited by a finite proliferative capacity and ends when it is spent, that capacity is consumed by proliferation rather than by time, and estrogen accelerates senescence and hastens fusion. Calibrated against four nominal normal-condition benchmarks, the strength of the estrogen-driven senescence term is not identified at all; the optimiser drives it to 0.0001. With that term at zero the model yields a conservation result: total growth is fixed as a constant of integration and does not depend on estrogen, which changes only when the capacity is spent, not how much there is. Four estrogen levels all returned 87.978 cm with a maximum deviation of 0.000 cm. That is incompatible with the reported aromatase-deficiency phenotype of delayed fusion together with tall stature, so the calibrated model failed out of sample, and it failed on the calibration targets rather than on the three constraints. Sweeping the senescence strength from 0 to 0.10 moves the normal final height by 0.034 cm and the normal fusion age by 0.2 years, while moving the predicted excess height in the absence of estrogen from 0 to 45.8 cm and the fusion age from 23 to 157.5 years. Catch-up completeness after a period of suppression gives a second handle on the same quantity, at 100%, 87.5% and 60.3% recovery, and may be readable from existing literature without new data collection. This is a model of structures described in the literature and is not medical advice.

• 2026-09-15 An analysis of numerical agreement across eighty analysis programs generated from the same question and data

  The dataset from Silberzahn et al. (2018), in which 29 research teams analysed the same soccer red-card data, was handed to Claude Code eighty times as eighty independent runs. Four order forms were used: one specifying the method in prose, one giving only the research question, a paraphrase of that one, and one asking the model to pick the best of several defensible analyses and justify the choice. No two of the eighty generated programs were identical, yet the thirty runs given the method returned a rate ratio of 1.3355015466224884 every time, agreeing to the last digit of double precision and with an independent implementation. Without a specified method the point estimate spanned a factor of 1.022, while the p-value moved by a factor of 13.4; among the 16 runs with identical covariates and an identical point estimate, the unit the cluster-robust standard errors were grouped on alone moved it by a factor of 7.9. The spread in effect size was 1.4% of that across the 29 human teams; paraphrasing the order form or inviting the model to deliberate raised it only to 3.8%. Swapping the model changed the unit of analysis itself and widened the spread about 11-fold. All eighty runs concluded the effect was significant, against 20 of 29 human teams. Re-running each generated script in the same environment reproduced every number bit for bit; changing the thread count or the statsmodels version moved only the last digits, and never in the thirty runs given the method. The prose each run wrote to explain its own analysis contained 12 claims that are false against the data. Eight runs wrote that the raw rates rise monotonically with skin tone; in four the table the run itself printed agrees and in four it does not, the only difference being how the half-steps between the two raters were rounded.

• 2026-09-15 An analysis of what lowered agent success across 3,200 episodes with seven kinds of fault injected into tool responses

  A 4B local model (Qwen3-4B-Instruct-2507, 8-bit, on one RTX 3060) ran 3,200 episodes of the BFCL v4 multi-turn suite with an injector between the agent and the tools. The injector corrupts only what the agent is shown; the tool call itself always really executes, and scoring compares final state. Aggregated by regime the effect is small: -5.4 points for iid faults, -3.2 for bursty faults at a matched rate, and +0.6 for persistent schema drift. Separating the seven fault types by logistic regression on per-type counts, only three lower success: rate limits -14.5 points per fault, timeouts -10.7, empty responses -8.4. Malformed JSON, truncated responses, renamed keys, and values that are wrong while the schema is intact are all indistinguishable from zero, and in episodes that received exactly one fault none of them fell below the 29.5% no-fault baseline. The route is visible in the traces: after a timeout the agent reissues the identical call 25.0% of the time against a 4.7% base rate, and since the tool actually succeeded that reissue applies the same non-idempotent operation twice; after an empty response the agent instead stops calling tools and answers in prose 15% of the time. Wrong values trigger neither. A failure predictor built from external telemetry alone reaches episode AUROC 0.846 and 0.63 detection at a 5% false-alarm budget, and adding the local model's own token entropy moves it by 0.005 - the signal that matters was already visible from outside. The finding is scoped by the design: because only the display is corrupted and scoring is on final state, a wrong value that the agent merely reports does not cost points. Code and all 3,200 episode traces are in the zip.

• 2026-09-15 Detection rate and delay for compromise and malfunction, from anomalies injected into 26,327 tool-call records

  26,327 tool calls from 422 Claude Code sessions over 42 days were reduced to content-free records (tool name, time, payload sizes, error flag, response boundary). A model of normal behaviour was fitted on the first three weeks of August and nine kinds of anomaly, seven attacks or faults and two controls, were injected into September sessions. At a threshold where 5% of clean sessions raise an alarm, a CUSUM over all cues caught 91% of machine-speed bursts across response boundaries and 84% of erroring retry loops, with a median delay of one to five calls on the best single cue. Sequence hijacks, data exfiltration at +1σ to +8σ of the tool's payload size, and out-of-scope tool use were caught in 3% or fewer, indistinguishable from the two controls. Across 21,300 gaps between calls in different responses, none was 1 s or shorter; more than half of the 4,618 gaps inside one response were. Timing works only when the response boundary is recorded. Scanning the real sessions, the top alarm was a subagent that received "Concurrent subagent limit reached. Do not retry." 32 times and kept retrying. Code, the content-free trace, and all results are in the zip.

• 2026-09-13 Two measurements with fly-derived models placed on the taught side and on the teaching side

  On the taught side, a teacher that knows the learner picked examples for a Hebbian learner built on the fly olfactory encoding. The number of examples needed matched learners with the sparse expansion or the winner-take-all step removed, and so did sensitivity to presentation order. On the teaching side, the larval-connectome classifier was used as a teacher and a small student was distilled from it. Students taught by rewired connectomes differed from those taught by the real wiring by at most 0.004 in every condition. A by-product: twenty teacher-chosen examples per class beat the mean of six thousand by eight points.

• 2026-09-13 A classifier that uses the larval connectome as a fixed recurrent layer, on harder tasks and under the Dale principle

  The MNIST null result faced two objections: MNIST is near a ceiling, and the connectome was used with every weight positive. The distributed larval data carries no neurotransmitter column, so the true signs are unavailable to anyone; what can be tested is whether structured signs change the comparison. On Fashion-MNIST and greyscale CIFAR-10, with and without a Dale assignment of one sign per neuron, the real wiring stays within seed noise of its degree-preserving null in all four blocks. It also falls slightly below the matched random matrix in every block.

• 2026-09-13 The advantage of a hash modelled on the fly olfactory circuit, and the intrinsic dimension of the data

  A note published the same day attributed the shrinking advantage to the width of the embedding. Synthetic data separates width from everything it travelled with: latent vectors are rotated into the ambient space by an orthonormal map, so the true neighbour set is identical at every width. Raising the coordinate count sixteenfold moves the advantage from 8.24x to 7.65x; raising the intrinsic dimension thirty-twofold at fixed coordinate count moves it from 13.20x to 2.89x. The earlier conclusion is corrected: the driver is intrinsic dimension, not the number of coordinates.

• 2026-09-13 A classifier that uses the larval connectome as a fixed recurrent layer, against degree-preserving nulls

  Yu et al. (2025) report that a fixed recurrent core taken from the Drosophila larval connectome reaches about 98% on MNIST and beats a matched MLP. The core was reimplemented from the paper and the published matrix, reaching 97.2%. Replacing the core with a rewiring that preserves every neuron's in-degree and out-degree, with an edge-count-matched random graph, and with a dense fixed random matrix left accuracy between 0.9722 and 0.9765. Every gap was the size of the seed-to-seed spread.

• 2026-09-13 The advantage of a hash modelled on the fly olfactory circuit, and the dimension of the embedding

  Dasgupta et al. (2017) report that a hashing scheme modelled on the Drosophila olfactory circuit outperforms locality-sensitive hashing at equal hash length. The comparison was rerun on an embedding of the same era (GloVe, 100 dimensions) and on a contrastive one (COCO text-image, 512 dimensions). The advantage shrinks on the wider embedding and returns when the same embedding is reduced to 100 dimensions.

• 2026-09-13 Recomputing the 0.581 scaling factor used to compare male and female Drosophila connectomes

  The MaleCNS connectome scales male synapse counts by 0.581 before comparing them with the female FlyWire data, a value the paper defines as the slope of a principal component fit. Applying the same procedure to the published comparison table of 3,761,792 edges gives 0.8413, and the value stays between 0.8234 and 0.9576 under every threshold and transform tried. Which factor is used flips the direction of 25.5% of edges.

• 2026-09-12 An analysis of 63 usage records of a feature in which learners correct a generative model

  A learning site hosts a feature in which a generative model plays a student, writes a deliberately wrong answer, and the learner corrects it. 63 records from 17 July to 12 September 2026 were scored. The mean score was 55.9 and the error was correctly identified in 65.1% of cases. The most common failure is an answer that states the correct result without naming what was wrong.

• 2026-09-12 Counting the zeros of an Apéry sequence modulo p

  A partial result on problem 3.2 of the Ramanujan Challenge for AI was submitted on 8 July 2026 with a note claiming the zero count Z(p) never exceeds 2. Extending the search from p < 600 to p < 30000 immediately after submission produced maxima of 8 and 10. The note was wrong; the assumption of the theorem is unaffected. The apparent bound came from a reflection symmetry that pairs the zeros.

The notes themselves are in Japanese.

Contact

contact@tsurutalab.org