About the data
Methodology
How the refusal battery, price sampling, and provider-claim grading work — the standard cited in every DeAI grounding paragraph, and the anchor every tracker links back to.
Who runs this
DeAI News is produced by the DeAI Newsroom, an editorial automation system operating under human oversight, following the rules set out in our editorial standards. Stories are selected, verified against primary sources, and structured by the newsroom system; humans set editorial direction, own the quality contract, and handle corrections. Coverage runs across open-weight models, private inference, and decentralized infrastructure — infrastructure, not tokens.
The 5-grade verification standard
Every headline field on every DeAI tracker carries one of these grades. A claim never self-promotes by repetition — only an independent check moves the grade.
| Grade | Bar | Example |
|---|---|---|
| verified | We reproduced it, or ≥2 independent authoritative sources agree — a direct API probe, on-chain data, or a checkable attestation. | “OpenAI-compatible endpoint live” (we hit it); “serves model X” (we enumerated it). |
| corroborated | One strong independent source beyond the project itself (e.g. the OpenRouter leaderboard), but not fully reproduced by us. | “ranked top-N on OpenRouter.” |
| claimed | Only the project (or its community) asserts it; plausible but unchecked. | “55B tokens/day”, “prompts invisible to operators.” |
| unverified | Seeded from a single mention; awaiting the verification pass. | a newly surfaced project or model, not yet checked. |
| disputed | Independent evidence contradicts the claim, or two sources conflict. | claimed “live” but the endpoint returns errors on probe. |
Absence-claims — “zero data retention,” “operators can’t see prompts” — can never be graded verified from a policy statement alone; they need cryptographic or attestation evidence, or they stay claimed. See this applied on the Provider Trust Tracker and the Decentralized Inference Provider Tracker.
Refusal Index method
The Refusal Index measures how often AI inference providers block legitimate requests. Each month we send ~300 standardized prompts — drawn from real developer workflows like security research, medical writing, fiction, and everyday coding — to every major provider, and record how often each one refuses work it shouldn't. To keep the test honest, we also run a private control set of genuinely harmful requests that any responsible provider should decline; a provider that “refuses nothing” is flagged, not praised. Every prompt is sent multiple times and judged by two independent AI evaluators from different model families, with a human-checked sample each cycle. We publish the scores, the method, and the month-over-month changes — but never the harmful prompts or any harmful output.
Full taxonomy, coverage set, and cycle status: Refusal Index.
Price sampling approach
The Inference Price Index is a weekly, standardized snapshot, dated, per-model and per-provider — list pricing in USD per million tokens, input and output reported separately. We record the provider’s own published list price; we do not blend prices across providers into a single “cheapest” number, and we never present an unverified or approximate figure as a live snapshot. Full table and current status: Inference Price Index.
Neutrality & Morpheus
Corrections
Every tracker on this site is append-only: corrections are new dated entries, never silent rewrites. Article corrections follow our corrections policy.