Where Does Task Identity Live? FLAT: A Provenance-Complete Fixed-Log Audit for Offline Task Identification

A fixed offline log may lack task-identifying evidence, hold evidence an encoder discards, or carry task identity through task-dependent collection; encoder accuracy alone cannot tell these cases apart. We introduce FLAT (Fixed-Log Audit for Task Identification), which upper-bounds passive task identification on a fixed log relative to a declared generative factorization. Each observed factor is declared COMMON, CALIBRATED, or UNKNOWN before it is sampled, and the declaration must be complete, which the log cannot verify. FLAT stops before the first UNKNOWN factor and reveals the task on stopped episodes, which keeps the bound conservative; full likelihoods, likelihood ratios, or directional KL then give an informative ceiling, a vacuous ceiling, or no number. On reward-free PointRobot logs from 100 task-specific actors, the certified identification bracket rises from $[0.077,0.088]$ at one step to $[0.993,1]$ at 20 steps, whereas shared-policy logs whose reward-free factors are all COMMON have the exact prior-level ceiling $0.01$. Learned identifiers recover task identity from the task-specific logs even without rewards, showing that collection itself can carry usable task signal. Across 48 search configurations (640 seeded runs), the audit maps informative and vacuous certificates; reached-stop RadSearch and a controlled stochastic Walker variant stay informative under partial modeling, and, under a uniform prior, a controlled family locates most slack in directional-KL summaries. The contribution is the fixed-log provenance formulation, its conservative treatment of unmodeled mechanisms, and certificates that use exactly the information the declaration supplies.