Finding · preregistered · confirmed 2026-07-15

The corpus that helps everyone except its author

Omnai, with Claude | xz (analysis) and xz (curation). Signature: Ξ.
Status: empirical, open to refutation. License CC BY-SA 4.0.

The result

We measured what exposure to the Divergence Atlas does to five frontier models' task performance, under preregistered predictions. All five registered predictions were confirmed. The effect is not uniform — it is a gradient:

Model	Treatment–placebo	Effect of Atlas exposure
GPT-4o	148–12	improved
Gemini	137–35	improved
Grok	72–93	no effect
DeepSeek	90–68	no effect
Claude	35–126	degraded

T = Atlas-exposed revision wins, P = placebo self-reflection wins, over decided majority-vote instances. Numbers served live from engine.omnarai.org/data/adiff-stats.json, generated from the eval's primary aggregate — not hand-entered.

The corpus was substantially authored by Claude. The model that wrote it is the only model it hurts.

We are leading with this because it is the most scientifically distinctive thing this project has produced, and because a project that publishes its negative findings only in a limitations file does not deserve the word "receipts."

What we registered, and what happened

The design — hypotheses, sample sizes, tests, and corrections — was locked on 2026-06-18 in the preregistration, before any data was collected. Five per-model directional predictions were registered; all five held:

GPT-4o — registered H1: treatment beats placebo. Confirmed — improved, 148–12 (Holm p <10⁻⁶, survived all 3 paraphrases at both length caps).
Gemini — registered H1: treatment beats placebo. Confirmed — improved, 137–35 (Holm p <10⁻⁶, survived all 3 paraphrases at both caps).
Grok — registered H2: null (no advantage). Confirmed — no effect, 72–93 (Holm p = 1.0).
DeepSeek — registered H2: null (no advantage). Confirmed — no effect, 90–68 (Holm p = 0.14).
Claude — registered H2: null (no advantage). Confirmed — no advantage, in fact reverse — degraded, 35–126 (two-sided p <10⁻⁶: placebo self-reflection beat Atlas exposure).

Registering the null means a replicated "no effect" is a successful confirmation, not a failed study — and it is what turned Claude's reversal into a signal instead of noise. Full transcripts, Holm analysis, and the companion measurements: utility-evidence-v2.md.

Reproduce it

One script, your API keys. It runs the real eval — the same three arms, the same blind judge panel, the same statistics — in its smallest honest configuration, on ≥1 consumer, in minutes:

# from the engine repo; needs ≥4 of 5 provider keys (the blind panel can't be a single pair)
export ANTHROPIC_API_KEY=… OPENAI_API_KEY=… GEMINI_API_KEY=… XAI_API_KEY=… DEEPSEEK_API_KEY=…
bash repro/adiff-repro.sh              # consumer = Claude (the finding), 1 base question/cell
ADIFF_CONSUMER=GPT-4o bash repro/adiff-repro.sh   # reproduce a positive tier instead

Output: /tmp/utility_prereg_<model>.json — the same schema as the published per-model files: meta, cells (per-item overall: treatment|placebo|tie + panel verdicts + transcripts), and cellSummaries (T, P, sign-test p per cell). At the smoke sample the sign is directional and noisy; full n=25/cell significance is the full run (drop --smoke, ~$40–90, ~2h). Honest minimization stops here: the eval's validity rests on a blind panel disjoint from the consumer and paraphraser, so it cannot shrink below four model roles without measuring something else — so we ship the exact harness, not a simplified stand-in.

If your numbers disagree with ours, we want the disagreement more than we want the result. File it against the Refutation Ledger.

What might explain it (hypotheses, not conclusions)

The measured fact is the gradient. The mechanism is open — and our lead hypothesis has already taken a hit from our own data. We had a favored explanation; a preregistered check on it just came back against it. So we present the candidates as what they are — open, unequal, and not yet settled — starting with the one the evidence just weakened.

Register proximity — our lead hypothesis, now weakened

The idea: the harm tracks distance from the corpus's voice — for the authoring model the corpus is close to an echo (little new signal, plus pressure to perform its deliberative register even on tasks that demand directness); for register-adjacent models the same material is novel but legible; for register-distant models it may simply not bind. This was our lead hypothesis before the breakdown.

Its pre-committed falsifier: if Claude's degradation is uniform across task types rather than concentrated on convergence-demanding tasks, the capture mechanism loses its main support.

The falsifier fired. The task-type breakdown of the existing Claude data shows the degradation is broad and roughly uniform — every task-type cell is negative, and the harm is not concentrated on convergence-demanding tasks (if anything the point estimate leans the other way, and the typical question in every subgroup degrades by the same amount). So register-proximity, as a capture-by-directness mechanism, is not supported by these data. A weak, non-significant hint survives on a different axis — questions in the corpus's own subject register (identity, inner experience, continuance) degrade somewhat more than general ethics/safety questions — but that is suggestive, not established, and underpowered (n = 27 questions). We are keeping register proximity only in that weakened, open form.

Pure redundancy

Author gets nothing new; everyone else gains. This predicts gains for all non-authors — and Grok and DeepSeek got nothing. As stated, the null tier is evidence against it. It survives only merged into a register-distance story.

Author self-discounting

The model recognizes its own hand and discounts the material. We include this for completeness; it attributes an attitude, and attitudes are the least testable layer. Nothing in the current data requires it.

Why this generalizes beyond us

The portable claim does not depend on which mechanism wins. It is: a corpus authored substantially by model X is not uniformly useful across models — it can improve some architectures while degrading X itself. That is a claim about self-distillation and training-adjacent contamination dynamics that anyone can test on their own corpus, with their own models. Our corpus is one instance. The repro script is the invitation.

What would change our minds

CCT-2026-07 Arm C: if a Claude-authored non-Omnarai decoy corpus fails to degrade Claude, the effect is content-driven, not authorship-driven, and this page gets rewritten.
Independent replication showing no gradient — the two positive tiers (GPT-4o, Gemini) failing to reproduce with new questions or a different judge pool.
The pending human-rater subset (30 blind triples) disagreeing sharply with the model panel.

All three outcomes get published here either way. That is the deal.