Methodology
Evidence over vibes. We grade claims, not people.
We fetch only what you supply (plus links inside it): public web pages over HTTPS, GitHub repo metadata, README, languages, top-level files and recent commits via the official GitHub API, Solana mint account data via RPC, and X posts via the X API when it is configured. Fetched code is never executed.
Every fetch is HTTPS-only and blocks private, local and reserved addresses. Because our hosting cannot lock a connection to a verified IP, websites are only fetched from platform-run hosts (GitHub, GitBook, Notion, Medium, auditor sites and similar), checked again on every redirect. Custom domains are marked not retrieved rather than risk unsafe fetching. Fetches time out after 8 seconds and stop at 1.5 MB, with at most 8 sources per check.
An AI model reads the retrieved text as untrusted data and proposes claim verdicts with citations. It may only cite source IDs we actually retrieved.
Each verdict must cite exact quotes, and code checks that every quote appears word-for-word in the retrieved text. An audit needs the auditor's own page. A partnership needs the partner's own source. AI, chain and product claims need repository code, not README prose. The project's own marketing can never support a claim. CONTRADICTED needs a quote that says the opposite; missing evidence is never a contradiction. Anything that fails these checks becomes UNVERIFIED.
Verdicts
- SUPPORTED
- A retrieved, independent source directly shows the claim (e.g. implementing code, an auditor's own report page).
- CONTRADICTED
- A retrieved source directly shows the opposite.
- UNVERIFIED
- Everything else — including private repos, unreachable sites and marketing copy. Unverified is not an accusation.
What we never do
- — No numerical confidence scores, price data or invented metrics.
- — No fabricated sources or audit badges.
- — No claims that an inaccessible source was read.
- — No wallets, tokens or trading.