mirror of
https://github.com/razzant/ouroboros.git
synced 2026-08-22 09:13:21 +00:00
Two GAIA v6.56.0 trace-verified losses were harness-shaped, not capability: 1. After an acceptance downgrade the agent shipped `FINAL ANSWER: blocked_with_evidence` verbatim. The snake_case outcome-tier ledger identifiers (best_effort / blocked_with_evidence) are now structurally rejected by extract_final_answer (they count as a missing answer, so the marker-nudge/salvage path recovers a real one). 'solved' stays extractable — an ordinary English word that can be a legitimate answer. The answer-protocol context rule names exactly the two snake_case identifiers (consistent with the extractor); the improvement-capsule closing line stays identifier-free to preserve the degraded-actor-no-tier-leak invariant. 2. A last-round refusal finalized with an EMPTY typed answer despite 24 tool calls of real research: the P2 final-marker nudge ALSO required a declared expected_output, but GAIA-shaped contracts carry the question in 'objective' with expected_output empty — the only salvage surface was silently suppressed. The protocol gate (answer_protocol_active) is sufficient on its own; the extra expected_output condition is removed. Bench canon untouched: both tasks stay scored as-is in the v6.56.0 table. Review: triad+scope converged over 4 rounds (r1 badge preflight, r2 self-consistency wording, r3/r4 clean). Full suite green, ruff F clean. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| ARCHITECTURE.md | ||
| CHECKLISTS.md | ||
| CREATING_SKILLS.md | ||
| DEPLOYMENT.md | ||
| DEVELOPMENT.md | ||