RAVEN. / research boundaryRAVEN. / research boundaryRAG retrieves; deep research loops; agents summarize. None of that verifies a token. An agent may retrieve, compare, and explain Raven receipts — it must not replace exact receipt verification with retrieved text, memory, summaries, screenshots, citations, or generated prose. Machine-readable: /agentic-research-boundary-policy.json.
Resolve chain, mint, requested action, evidence scope, and acceptable receipt age before expensive work. Never infer a missing mint or an ambiguous chain — ask or fail closed. If a user asks "is it safe", answer in Raven's bounded language; the word does not exist here. Prefer deterministic fields, enums, and schema inputs over model inference.
The loop is: parse structured request → retrieve exact receipt → validate schema → verify signature/key/staleness outside the LLM → evaluate completeness against the requested scope → retry within limits or stop with explicit gaps. Missing evidence becomes a coverage gap, never a model guess. Store large tool outputs in state with stable references — don't stuff raw evidence into active context. Context compression may shrink notes; it may never drop receipt invariants (verdict, findings, gaps, keyId, signature status, slot, engine version, mint, chain, staleness). Completion requires the completeness eval — not model confidence, and not context-window pressure. Bind only the tools the step needs; wrong-tool loops mean stop and rescope.
The writer gets structured verified fields plus coverage gaps plus allowed language. It may not invent evidence, omit blocking gaps, or strengthen "not verified". Draft reflection before display: no safety-affirming wording, no price/trading advice, gaps visible, verdict wording unchanged. A synthesis gap (draft misused available evidence) means rewrite; a research gap (evidence missing) means retrieve within limits or state the gap. And citations are references, not verification — a page citation never replaces ed25519 signature verification. Every material action still requires the exact receipt, verified, fresh, with gaps and policy applied. Drills 31–34 rehearse this: failure drills.