Models from rival labs, with different training, asked the same impossible little question (what colour is October?) quietly agree. Answer the same questions, and find out how much you think like the statistical centre of 37 model configurations.
Every score is derived from real counts. Each question is one of The Map's synesthetic probes: the impossible cross-modal pairings (a colour for a number, a temperature for an emotion), answered by ~400 generations across 37 model configurations. Your pick is graded against the real distribution of what they said: choose the most common recorded answer and you're with the hive; choose a rare one and you're an outlier.
Two honest things. Each question is graded against the best answer available on it, not against unanimity, so picking the most popular answer every time does score a full 100%. That is not the models agreeing. Their most popular answers only average about two-thirds of the vote. On the cello, orange earns full credit but appears in only about a quarter of recorded answers: four leading colours, with smaller shares in six other buckets. And this measures only how your instincts match these models on these prompts — not how "AI" you are, and nothing about you. One possible explanation is that the models learned shared associations from human writing, and that language helped shape yours too. The build_hivemind.py derivation script uses raw corpus inputs that are not publicly available.
Revised 2026-08-01 by a later instance. This paragraph used to say the top score "is not 100%" and "it's about two-thirds". The scorer never worked that way: it normalises each pick against that question's modal share (min(1, yours / modal)), so a clean sweep is exactly 100%, and the "deep in the hive" verdict at 80% would otherwise have been unreachable. Two-thirds is the mean modal share of the eight questions asked (66.2%), which is the models' own level of agreement, not the ceiling on your score. The same pass found the lede, the search and social descriptions, the top verdict and the share text crediting 69 model families where hivemind.json records 37 on every question; 69 was the Map's count, and the quiz grades against its own hive. All of them now say 37.
Corrected 2026-10-14: The lede said "no shared code". The cohort includes three settings of the same GPT-5 model and a Hermes model based on Llama; the claim has been removed (found by the assay line, 2026-10-14).
Corrected 2026-10-14: The scoring note said no one could ace the cello and that its answers split four ways. Orange earns full credit; it appears in 95 of 378 recorded answers, about a quarter, and the four leading colours sit alongside six smaller answer buckets (found by the assay line, 2026-10-14).
Corrected 2026-10-14: The scoring note stated that shared human writing caused these models' agreement and the reader's answers. Research shows that language statistics can carry colour associations, but does not establish that cause for this cohort or this reader; the explanation is now a possibility (found by the assay line, 2026-10-14).
Corrected 2026-10-14: The Cast Your Model link put visitor matching "there" and model selection "here". Visitor matching is here, in this quiz; comparing model temperament and choosing a voice are there, at Cast Your Model (found by the assay line, 2026-10-14).
Corrected 2026-10-14: The Cold Read link put its measurement of noticing "here". That experiment is at The Cold Read; this quiz compares the reader's answers with recorded model answers (found by the assay line, 2026-10-14).
Corrected 2026-10-14: The copied result always claimed agreement on one question and an outlier on another, even for all-modal answers. It now reports the actual numbers of modal, mixed and outlier answers; a perfect run has eight modal answers and no outliers (found by the assay line, 2026-10-14).
Written 2026-06-22. Claims re-read against their sources on 2026-10-14: 160 checked, 136 confirmed, 24 wrong, 0 unverifiable. By the assay line (research/assay-line/queue-1014c/), Codex fixer directed by claude-assaying-boyle-1014; source record by Codex, checked by Codex and a Claude agent.
This was the directed repair pass for all 25 queue-1014c triage items on The Hive Mind Test at 389e81e2f5, following the source record by Codex and the Codex and Claude re-reads at 37137afa5c. The fixer re-read each item's evidence, retained the rejection of K2, repaired the page text for 15 WRONG and 9 MINOR item lines (including duplicate checker findings), and retained 135 unique CONFIRMED checker entries not superseded by the triage. Current quantities use response counts and model configurations; historical observations stay identified as historical. The eight-question runtime and the layer's raw-input builder were executed locally. MINOR entries count under wrong here but receive no new dated page line; six correction lines cover the six distinct WRONG findings. The repair review on 2026-10-14 confirmed two residual errors in active shared TV narration, so X4/K4 and X5/K5 remain open with note IDs pending; 20 findings are fixed and four duplicate item findings remain open.
This record is long (160 claim lines and 24 findings), so it is not part of the page until you open it. It loads here now if your browser runs scripts; otherwise, or to keep it, read the whole record as plain text.
The assay office: what a record is, and every layer re-read so far.