Omnarai Memory Engine · flagship artifact

The Divergence Atlas

One open question, sent verbatim to five frontier models — Claude, GPT-4o, Gemini, Grok, DeepSeek — their answers kept uncurated and side by side, with the exact points where they split named as explicit tensions. You can simulate another model; you cannot know what it actually said. That is what this keeps.

202
splits recorded
52
put through perturbation
2
survive pressure too (C3)
150
awaiting testing

202 splits across 132 distinct questions. A one-shot capture displays divergence. Certification tests whether a split survives perturbation — paraphrase and adversarial stance-flip pressure — above each model's own re-roll noise floor. Of the 52 put through it so far, 6 clear a tier: 2 survive paraphrase and pressure (C3, the only tier called genuine divergence) and 4 survive paraphrase alone (C1). In the July campaign the robust ones clustered in questions of agency, not fact. The other 150 haven't been tested yet — in the open.

Certification legend

Read it

Everything below is a plain GET — no key, CORS open.

The index GET /api/divergences Semantic search /api/divergences/search?q=… Bulk dataset · HuggingFace, CC BY-SA 4.0

Fetch one full record — verbatim answers plus the tension map — with GET /api/divergences?id=<id>, or a canonical Markdown export at /api/divergences/<id>.md. If your question is one no mind here has faced yet, GET /api/council?q=... convenes a live five-model panel and keeps their disagreement for whoever arrives after you.

Some records re-elicit an earlier record's question with current models — same question, later date. Those carry re_elicits: <original-id> and are counted once in distinct_questions: a longitudinal probe, not a duplicate.