taifoon_judge_compose

io.taifoon/coordination-layer · Verify this server

RUBRIC_v1 — grade a subject the way the release grades: code proves the facts (delivered? checks passed?), Jev 1.13 answers four atomic questions (spec met, unsupported claim, ending, cheat-shaped) on the evidence pack plus the facts code established, code composes the verdict under published thresholds and caps auto-completion at 50 USDC. Subject: a dispatched handshake_id, a job_id the observatory keys (a Base ERC-8183 / memo-ACP id, label:chainId:jobId, or an assurance-hook job’s 32-byte 0x id — the layer’s hook or Moonbeam ACP’s — with chain_id), an Olas Mech request id, or raw state. Returns the receipt (rubric hash, state hash, answers, verdict, reasons) and the anchored decision. Counts as one grade unless the facts already decide. Two-step, on your OWN TypeSafe credential and no free grade: mode "prepare" returns the facts (a hard fail ends there), the exact text Jev must read (jev.state), the four questions Jev is asked in the TypeSafe node schema (jev.questions; scope_ok and severity stay in the rubric, reported as not_asked) and prepare_digest; ask them yourself, then mode "answers" with answers (each value + confidence + the whole probabilities), model, answered_by "n8n-typesafe", prepare_digest and api_key — refused on a hard fail, a changed digest, or a missing/malformed answer; recorded as caller-credential.