Measure recovery within a defined environment.

What it tests

150 deterministic failure-recovery scenarios across 50 fully observable weighted navigation layouts. This tests local policies after execution rejection, not LLMs or customer agents.

Failure taxonomy

Collision

A move into a blocked cell is rejected. The state stays unchanged.

Invalid path

A nonadjacent move is rejected. This category shares the collision failure-state prefix.

Premature completion

A completion request before reaching the target is rejected.

Method

Reset, execute the reference prefix, inject a rejected action, run a bounded local policy, verify transitions and final state, then replay. Recovery requires reaching the target at the original exact optimal total cost.

Baselines & results

WFRB v0 local baseline experiment.

Recovery + exact-cost oracle passed, per 300 experimental cells for each policy in one invocation. The second invocation repeats these cells; it does not double the independent sample.

The planner has full grid and cost access and recomputes a shortest path. Random walk proposes uniform cardinal moves. Blind repetition repeats the rejected action. These are controls, not competitive commercial agent baselines.

Random walk reaches the target in 27/300 cells, but only 4/300 satisfy exact-cost recovery. Task completion and verified recovery are different measures.

Reproducibility

Two fresh local invocations of 900 trials each: 1,800 executed trials. Their result streams are byte-identical; timestamps and runtime vary.

Results SHA-256

Limitations

Fully visible, solvable grids and rejected actions that leave state unchanged. No hidden map changes, LLM, customer agent, external service or paid API was tested. Correlated categories and deterministic seeds are not independent observations.

The perfect planner result validates this narrow harness. It is not evidence of general recovery or commercial performance.

Try Explorer

The Explorer uses a separate public navigation sample. Its paths are not WFRB agent execution logs.

Private evaluation

Customer-agent results require separate authorized runs, a scoped environment and agreed checks.