Rules decide.
Receipts prove.
Skein buys each check from a ranked Telegraph miner, normalises every answer into a finding, and runs a fixed decision table. No language model decides anything; the same findings always give the same verdict and the same hash.
Evaluated top to bottom. The first row that matches is the verdict.
- 01REJECT
A blocker hits
A sanctions match, or a critical audit failure, decides on its own evidence.
- 02ABSTAIN
A required check has no answer
Timed out, unavailable or unreadable: Skein names it rather than guess.
- 03ABSTAIN
Coverage below the policy minimum
Too little of the policy's weight was actually answered.
- 04REJECT
Risk at or above the reject line
Weighted risk over the answered checks crosses the policy threshold.
- 05ESCALATE
A contradiction fires
Clean code with a wash-traded market; deep liquidity held by insiders.
- 06ESCALATE
A required check failed
One failed required check never reaches APPROVE, however low the risk.
- 07APPROVE
Low risk and enough confidence
Every required check answered, risk under the approve line.
- 08ESCALATE
Anything else
The review band: a person should look.
Risk
risk = status_risk × severity_multiplierPASS 0, WARN 0.5, FAIL 1, scaled 0.6 / 0.8 / 1 / 1 by severity. Weighted over the answered checks.
Coverage
coverage = answered weight ÷ policy weightRequired and optional together. A required check with no answer is an ABSTAIN regardless.
Confidence
coverage × agreement × mean confidenceAgreement drops 25% per contradiction. A blocker carries the blocking finding's own confidence.
Six threads, one knot.
An illustration, not a real verdict. The audit passes, but wash trading fails, so the gold arc marks a market-integrity conflict. A weighted average would have approved this token; Skein escalates it. Select a thread to read its receipt card.
Five answers from one API are one opinion.
Skein groups miners by operator and data source, read from each miner's base URL. Checks answered by the same family share one weight: the largest of theirs, split between them. On the live network four of the strongest candidates run on one operator, so this matters from the first verdict.
ABSTAIN is a feature.
A refusal that names what is missing is worth more to a listing desk than a confident guess. If a required check times out, the verdict says so, names the check, and charges only for the inputs that answered.