RAVENThe anti-slop layer
Agents generate convincing text, code, summaries, and UI. None of it is launch evidence unless Raven checked the supported evidence and signed a receipt. Raven's job: a crisp boundary between "the agent said" and "the verifier checked". Machine-readable: /anti-slop-policy.json.
LLMs may explain — never alter
An LLM may explain a receipt, summarize findings, and help gather context to ASK Raven about. It may not alter a receipt, upgrade a verdict, remove coverage gaps, convert pass into "safe", treat missing evidence as a pass, or override staleness. Derived interpretations may only hold or DOWNGRADE confidence — never upgrade it.
Summary contract
A current receipt-v1 summary must preserve: subject, rulesVersion, findings, coverage gaps, slot, payloadHash, receiptId, signer identity/trust status, integrity, freshness, and the consumer-policy action. It must not invent a verdict field. Legacy v2 summaries separately preserve verdict, keyId, engineVersion, observed slot, replay/attestation hashes, signature status, and their version label. Forbidden: omitting gaps, vague paraphrases of findings, "safe/approved/good to buy", or hiding failures. With gaps remaining the summary says: not enough evidence for a full pass. The summary is never authority; the receipt is.
Generated integration code must prove itself
AI-generated integration code is untrusted until tests pass: current receipt-v1 payloadHash/receiptId/signature verification, tampered/missing-signature rejection, untrusted-signer rejection, expected-subject checks, fail-closed verifier errors, stale handling, gap preservation, no invented verdict, no "safe", no secret logging, no browser keys, no wallet signer permission. /pubkey is a cross-check only. Legacy v2 keyId/verdict handling stays separately versioned. No publish without human approval.