What we say,
and how.
The rules every finding, case study, report and post is held to. They apply to the scanner, the public record, the weekly and quarterly reports, our channels and anything written with AI.
Sources come first.
Every finding rests on public Robinhood Chain data: transfers, swaps, balances, funding transactions, contract code and events. Each one links to the transactions, blocks or wallets behind it, so a reader can check it without asking us.
Labels on addresses (exchanges, bridges, known deployers) carry their source. Community labels are proposed with evidence and voted on; they are marked as community labels and never shown as team findings.
We do not use leaked data, private messages, doxxing or guesses about who is behind a wallet. An address is not a person.
Three evidence states, always visible.
Observed: read directly from chain records. Inferred: a hypothesis drawn from observations, stated with its reasoning and the other explanations it leaves open. Unavailable: a source failed or does not cover the case. An unavailable check is never shown as passed, as zero or as a clean result.
Every flag carries a strength (strong, moderate or weak). Every scan shows when it was read and what it could not see. See the documentation and the methodology.
A scan is a reading, not a rating.
We never call a token safe, rug-free or guaranteed, and a low score is never presented as clearance. A scanned token is described as scanned. This is checked automatically on every change to the site and the API (scripts/ci/wording-lint.mjs), in every language we publish.
Nothing we publish is investment advice, a price prediction or a recommendation to buy or sell.
Facts about creators, never verdicts.
What we say about a creator, developer, deployer, team or insider is limited to measured, checkable quantities: “the creator wallet sold 82% of its tokens”, “3 of this wallet’s 5 earlier launches were recorded as collapsed”. We do not assert intent, motive or character, and we do not use words that do.
Every generated flag that mentions a creator must carry a number or a template value; copy that asserts intent is rejected automatically (scripts/ci/creator-phrasing.mjs). Outcomes such as “collapsed” describe what happened to a price, by fixed rules, not why.
Linked wallets are candidate relationships. Shared funding does not prove common control, and links through exchanges, bridges and routers are never counted.
Case studies.
Cases are chosen by a stated rule from the record (for the case library: the most-traded launches recorded as collapsed or with the creator selling out), not by who is involved. Every sentence in a case restates a recorded field: block, swaps, volume, peak, last trade, creator selling.
We publish no case built on sample or fictional data. If the index is unreachable, the page says so and shows nothing in its place.
Corrections.
Anyone, including a creator named by a finding, can ask for a correction, a takedown or a privacy change, with no sign-in. Requests are acknowledged and decided on a published clock and logged in public without naming the requester: corrections and takedowns.
A disagreement with a reading of the data (for example, that two wallets are not related) is an appeal, decided by reviewers in Cyclops Eyes. When a method changes, the change is listed in the score changelog and every scan keeps the method version that produced it.
Conflicts we disclose.
Cyclops has a token. Torch trading fees on it fund operating costs only (data, hosting, AI, team); they never fund rewards or buybacks. Contributor rewards come only from the disclosed Rewards wallet, whose address and balance are published on the reserve page.
The scanner treats our own token like any other launch. Any team or developer buy of it is disclosed with the wallet and amount. We do not accept payment to score, list, delist, feature or omit a token, and paid access changes limits and how soon some feeds arrive, never what a scan finds.
AI, with citations.
AI writes optional summaries of a scan and answers questions about it. Every sentence of a summary cites the observation ids it rests on; a sentence without a citation is not shown. AI never sets a score, a flag, an outcome or a label, and the rule-based reading is always shown beside it.
Machine-translated pages are marked as such until a native speaker has reviewed them.
Records you can check.
Every stored scan has a signed receipt; you can recompute its hash and check its signature on the verify page, and recompute its chain facts at the block it was read. The outcome dataset and the method are published for citation on the research page.