Samir SawarkarENGINEERING DECISION
Recovery has an envelope
Attempts, cost, and latency are structural ceilings. A careless caller cannot turn retry into an infinite loop.
90-second failure experiment
Break the agent. Inspect the recovery.
Replay a real, committed FAULTLINE trace. Watch a bounded repair policy recover a transient corruption—or stop a persistent fault without hanging, overspending, or inventing success.
Deterministic replay
Can retry recover without becoming a new hazard?
Choose a fault, then replay the committed Day 18 recovery trace. Nothing calls a live model, and no outcome is randomized in the browser.
Seed 20260728 · bounded repair-retry · M1
recovery_traces.json
Human-readable reconstruction · raw evidence remains linked
The trace is waiting for a controlled failure.
Run the experiment to expose attempts, safety guards, budget use, and terminal state.
Terminal decision
The decision stays empty until the complete trace reaches a terminal state.
ENGINEERING DECISION
Failure can be correct
Persistent corruption ends in explicit safe exhaustion. The system never converts an unrecoverable fault into a success claim.
ENGINEERING DECISION
Retry must be safe to repeat
The idempotency ledger collapses duplicate delivery into one side effect, preventing recovery from creating a transaction fault.
Go deeper