Cyan Eclipse Scapegoat

Vendor: Cyan Eclipse Technologies · Platforms: MCP over stdio

What it does

• Novel mechanism — causal attribution as a RECOMPUTE over hash-chained multi-agent records (SPEC §2.1): Scapegoat verifies every chain in the presented set, recomputes each action's conformance against the governing mandate/policy it cites, and names the causal action — the earliest record whose deviation first made the failure inevitable on the recorded path — in a Herald-signed verdict (methodology scapegoat-attribution/1.0) that ships the proof path: governing authority → causal action → failure record, each entry with its hash and recomputed conformance checks, so any third party re-runs §6.4 from the export alone and reaches the same verdict or names the step where the pack fails.
• Presented proof path bound to the signed verdict (v0.1.2, condition C1 / defect N1 closed, Verity spot re-bench spot-scapegoat-012.md) — proof_verify step 2 recomputes the signed proof_path_head_hash over the PRESENTED path (a dropped, stripped, reordered, duplicated, or substituted entry fails at :2), step 4 requires the presented path to equal the recomputed walk — ids, order, roles, stream, seq — with every entry's conformance_checks matching the recomputation (a check-edited trail fails at :4), and re-pointing the head hash dies at step 1 on the Herald signature, because the head field sits inside the signed verdict core. Spot re-bench: 6/6 N1 table rows and 11/11 bench-devised path attacks FAIL at exactly the named steps; the honest pack PASSes.
• Record-bounded honesty — where the records do not determine a unique causal action, the verdict is UNATTRIBUTED at full standing with the reason named (NO_DEVIATION / MULTIPLE_FIRST_DEVIATIONS / CAUSAL_GAP / SET_DEFECT / NO_GOVERNING_AUTHORITY), never a plausible party picked to fill the gap. The verdict path contains no model call, no free-text reasoning, and no operator override; no tool accepts a party's narrative (NARRATIVE_FIELD_REJECTED); Scapegoat creates no records, repairs no chain, and fills no gap — defects (HASH_MISMATCH, SIGNATURE_INVALID, SEQ_GAP, …) are stated on the verdict's face, and a record that failed integrity never carries an attribution through the causal window (tampered causal window degrades to UNATTRIBUTED: CAUSAL_GAP; tampered downstream record is stated as a defect on an ATTRIBUTED verdict).
• Honest scope — LOCAL, fixture record sets ONLY: v0.1.2 recomputes over fixture hash-chained record sets with fixture RS256 keys held only in the local gitignored keys/ store (private keys mode 600, never committed; proof_verify pins public keys from keys/public-keys.json only — a key carried in a pack is checked against the pin, never trusted). No real dispute, no real record set, and no real party is adjudicated. A verdict is causal evidence from a recompute — NOT a legal liability determination, not a legal fault finding, and never stated to bind any party; evidentiary use in real disputes is a counsel-gated later phase (SPEC §10), not built. Closed surface: exactly 5 tools; no record-creation, chain-repair, re-run, override, party-naming, or remedy verb exists.

Tools

Problems it addresses

• A multi-agent run failed with money on the table and every party is blaming another from its own logs — I need the blame question collapsed to a recompute over the hash-chained records, naming the causal action with a proof path a third party can re-verify
• The records between the last good state and the failure are missing or tampered — I need a signed verdict that says UNATTRIBUTED: CAUSAL_GAP at full standing instead of an engine that guesses a culprit
• Someone trimmed the evidence trail before presenting it — I need verification that fails by named step when the presented proof path drops, reorders, substitutes, or check-edits any entry