How a record is built

Three classes of statement

A verified fact is one onchain event, projected one-to-one and linked to its transaction. A signal is the output of a published rule over many facts, with that rule's confidence. An Aleph inference is a model's reading of the evidence. They are labelled everywhere, stored separately, and only facts and signals ever reach the score.

Trust Score · model v0.3

• Subscores that do not apply to a subject (a wallet that never deployed a token, a contract that sends no transactions) are left out and the remaining weights renormalised.
• Each risk item costs its severity points × its confidence: low 60 · medium 150 · high 300 · critical 500. Repeats of the same rule cost 100% / 50% / 25%.
• Severity caps: while an item with confidence ≥ 0.6 is present, the total cannot exceed 199 (critical), 399 (high), 649 (medium). A long history cannot hide severe evidence; the uncapped score stays visible.
• Confidence combines data volume and history age, discounted for every coverage gap. Below 0.25 with no medium-or-worse item there is no score (insufficient data); below 0.6 the score is provisional.
• History is append-only: a snapshot is written on the first score, a model change, a band or status change, a new or removed risk item, or a move of 10 points or more.

Deterministic signals (10 published rules)

Verified facts · event catalog v1

Relationship graph and association

• Only control-like links are followed: direct ETH funding (plain transfers), deployment, ownership and upgrades. Trading and token transfers are not relationships.
• An address linked to many others, or a known protocol contract, is a hub: shown, never walked through — so a faucet or an exchange does not connect everyone it paid.
• Association (risky_association) is reported next to the subject's own evidence and never as it. It is at most medium, carries low confidence, cannot cap a score, and does not chain from one address to the next.
• A funding cluster groups wallets by direct funding. It is not a claim that one person controls them.

Aleph · aleph-2

aleph-1 plus two rules: a medium/high answer is capped at the confidence of the strongest risk evidence it cites (association at 0.4), and contradictory reason codes may not be combined.

• Aleph reads a structured case file — record, RAP SHEET, graph, score history, monthly activity — in which every entry has a ref. It must cite refs for every reason.
• A validator rejects any answer that cites refs not in the case file, mentions addresses or transactions that are not there, uses accusatory or certain language, or claims more risk than the cited evidence carries. Rejected answers are kept in history, never shown.
• Risk levels are low / medium / high — there is no "critical". "High" needs a cited high-severity item about the subject itself; association supports at most "medium".
• Confidence is capped by data quality (0.3 + 0.6 × the score's data confidence) and, for medium/high, by the confidence of the evidence cited. Assessments expire after 7 days and are re-run only when the evidence changes.

Language

Records describe behaviour and evidence; they do not accuse. Generated text may not use: scam, scammer, rug, rugpull, rugged, criminal, thief, theft, fraud, fraudster, stole, stolen. Preferred wording: high-risk behaviour, suspicious pattern, elevated counterparty risk, association with high-risk entities, anomalous activity.