SRT hivemind artifacts

Every number in Where the Hivemind Comes From has a file here. This repository is the evidence, not a model: sampled continuations, hidden states, fitted-map scores, execution outcomes and run logs.

Paper: paper_hivemind.md. Code: github.com/space-bacon/SRT.

Reproduction principle. If a claim is in the paper and you cannot rebuild it from the files here plus the scripts in the GitHub repo, that is a bug and we want to hear about it. Four claims were corrected during the work by exactly that process, and each correction is recorded below rather than quietly overwritten.

What the paper argues. Jiang et al. (arXiv:2510.22954) measured 70+ closed models writing alike. Open weights let us take that apart: the convergence is not inherited from pretrained geometry, is not mostly bought by instruction tuning, and largely arrives with the deployment format. It is also suppressible, by a lever nobody pulls.

---

Layout

atlas/ — representation transport across 12 models from 8 labs

Held-out retrieval through a fitted map reaches 0.9181 across lab boundaries against a shuffled floor of 0.00101 and a self-map ceiling of 0.999. Transport cost is 0.0787. Shared corporate lineage is worth only 0.0357 of that.

coder_matrix1024/ — the scale test, Qwen2.5-Coder 0.5B to 32B

36 arms, 164 HumanEval prompts, K=8, 1024 new tokens. scaling_curve.json holds the per-rung table.

The format effect grows with scale, from 0.0878 at 0.5B to 0.1589 at 32B, a slope of +0.0433 per decade over a 66x range. The tuning term moves the other way at −0.0402.

coder_ladder/ is the same matrix at 192 new tokens, where 43% to 80% of candidates were cut off mid-function. Its similarity slope (+0.0441) matches the clean run; its pass rates do not, and every pass-rate figure below reads from the 1024-token pools.

ministral_ladder/ — independent replication, Mistral 3B/8B/14B

18 arms of prose stems. Role structure alone is worth +0.3809 to +0.4354, persona alone +0.0612 to +0.0821, and instruction tuning alone is slightly negative. The format effect rises at +0.0348 per decade, reproducing the Qwen sign independently.

ministral_suppression/, crosslab_suppression/ — can it be undone

Eight genuinely distinct personas take inter-model similarity from 0.6464 to 0.3817. Eight production-style personas, each a rewording of "helpful assistant", leave it at 0.6441. Across three labs at 14B-31B the deployed arm is worse than inert: inter-model similarity rises, an inter_drop of −0.0111.

mbpp_ladder/, ensemble/, verifier_1024/, code_select/ — what the pool is worth

Candidate pools and execution outcomes. verifier_1024/ holds the HumanEval read comparison on the 1024-token pools and copies of the MBPP reads; verifier/ is the superseded 192-token HumanEval run; ensemble/ holds six frontier models from five labs.

holonomy/ — a retracted selector, kept

The transport-defect selector and the sweep that killed it. Retained because the retraction is the useful part.

---

Using the selector rather than reading about it

The winning read is packaged as srt_select in the source repository. It picks one of K replies using only the replies: no reference solution, no test suite, no scoring model, no training.

from srt_select import select
best = select(user_message, replies)

Handed the benchmark's entry point and signature it scores 0.5854 on HumanEval and 0.8094 on MBPP, against floors of 0.4679 and 0.7185. Recovering the entry point and the argument shapes from the chat turn alone is the deployable case, and it costs coverage rather than accuracy: 62.3% of HumanEval problems resolve and 98.6% of MBPP ones do, giving 0.5476 and 0.7991 when unresolved problems are scored as an arbitrary pick. verifier_1024/chat_consensus.json and verifier/chat_consensus_mbpp.json are those two runs.

Selecting means executing every candidate, so it is a remote code execution primitive if pointed at the wrong host. srt_select.sandbox states what the confinement does and does not cover. Run it somewhere you are willing to lose.

---

Reading these files correctly

ensemble/union_ceiling.json is superseded and carries a SUPERSEDED key. Use union_ceiling_fixed.json. In the first run GLM-4.7-Flash was rendered with its <think> block open, so it spent its whole budget on analysis and never emitted code. Its pass@1 there (0.0526) is a render bug, not a capability. Corrected it is 0.7904, the union moves to 0.9878 against a best single model of 0.9634, and GLM turns out to contribute the only 2 problems no other model solves. The first run reported zero unique solves for every model, which was an artifact of that one broken member.

The union ceiling is an oracle, and a selector does not reach it. pooled_select.json runs real selection over all 48 pooled candidates from the six frontier members. It scores 0.9390, solving the same 154 of 164 problems as the best member's own eight candidates, for a gain of +0.0000. Recovering the target from the replies alone finds one more problem, which is noise at this size. Read 0.9878 as a bound on what better selection could win, never as an ensembling result: pooling models buys nothing that one good model plus a selector does not already give.

Example-filtering leads on both benchmarks, and agreement is the coverage play. Filtering by the prompt's stated examples recovers 68.1% of the gap on HumanEval and 76.8% on MBPP; agreement recovers 44.1% and 53.4%. MBPP states an example for 425 of 425 problems; HumanEval states one for 128 of 164, and agreement covers the other 36. On the superseded 192-token HumanEval pools agreement appeared to lead at 82.9%. That ordering was a truncation artifact: agreement was discarding code that could not run, against a floor that included it.

consensus*_covered_only_superseded.json carry a SUPERSEDED key. Those runs averaged the agreement read over the problems it could run while the floor and oracle were averaged over all of them, so 12 of 36 HumanEval arms appeared to beat their own oracle on the 192-token pools and 7 of 36 on the 1024-token pools (1 of 10 on MBPP). The files without the key score every problem: consensus falls back to the first reply where nothing runs, as srt_select.select() does, and consensus_strict charges those problems as failures. On the 192-token pools the correction is 0.4426 to 0.3762; on the 1024-token pools 0.6301 to 0.5854; on MBPP 0.8174 to 0.8094. Found by Dipankar Sarkar from the published artifact. The bound is now enforced twice: per arm inside consensus_select.py and chat_consensus.py, and as one pass over every results file by scripts/check_oracle_bound.py, which holds all-rows selector columns to their oracle and reports declared subset rates (consensus_on_covered, on_resolved_only) without judging them.

The learned verifier does not transfer. It captures 15.9% of the gap on HumanEval and 6.3% on MBPP, and on MBPP appending it to example-filtering makes that baseline slightly worse. Its AUROC is 0.778 on both budgets; a length-only control that scored 0.581 on the 192-token pools scores 0.526 on the clean ones. It is reported as a negative result.

Selection value decays with scale, at −0.0704 per decade on MBPP, from +0.1538 at 3B to +0.0097 at 32B, and at −0.0676 per decade on the 1024-token HumanEval matrix. It is a weak-model amplifier and does not push a strong model past its own ceiling. The MBPP slope is the chat-native read; agreement gives −0.0830 and execution-guided filtering −0.1192, so the direction is read-independent and the quoted figure is the most conservative of the three. See hivemind_census.json.

Ladder magnitudes are not comparable across domains. Code prompts constrain output far more than open-ended stems, so instruct-raw already sits at 0.74-0.79 on HumanEval against 0.4435 on prose. Compare slopes across domains, not levels.

The Ministral chat arm was regenerated. Its first version is kept under corrupt_chat_pre_fix/. apply_chat_template(tokenize=False) does not round-trip special tokens through Mistral's tokenizer backend, so the model read its own control tokens as prose. Anything in that directory is a record of the bug, not data.

---

Scale

110,704 generations across 97 arms, 12 models with saved states, 6 frontier models in the ensemble, two code benchmarks with execution outcomes from a harness validated at 164/164 and 425/425 on canonical solutions.

Citation

@misc{srt_hivemind_2026,
  title  = {Where the Hivemind Comes From: Geometry, Tuning and Format,
            Separated on Open Weights},
  author = {SRT},
  year   = {2026},
  url    = {https://github.com/space-bacon/SRT}
}