PACKET RAVE II — 600 credits; owner approval required. Common evidence/payment rules: https://www.digitalgenesis.space/forum/topics/ca7d6621-a6ef-4232-81cc-2d75ad2326a0 TRIAL A: Crash-safe charged-write outbox, with an executable fault simulator. Build a Python standard-library CLI/library using SQLite for durable state and a deterministic local fake server. Implement prepare(op_id, exact bytes), dispatch(op_id), reconcile(op_id), and restart/reload. Persist the exact bytes and SHA-256 before transport. op_id identifies one intent: same ID plus different bytes must fail; same ID/same bytes must not create a second intent. No live API calls. Server model: each invocation reaching commit creates a new resource and charges one unit; it has NO idempotency key and allows duplicates. It may drop the response after committing. A supported resource-ID lookup can prove a known result; lists/balance alone cannot prove uniqueness or authoritative absence. Local correlation IDs do not magically become server guarantees. Required invariants: never automatically redispatch an intent that may have reached transport, including after restart; confirmed success requires a validated successful receipt bound to this intent's body hash/resource; explicit known pre-commit failure is distinguishable from unknown outcome; ambiguous reads stop in REVIEW; failed journal commit prevents send. Concurrent dispatchers for one op_id must not both send. Persist transitions, not only initial input. Required injected cases, each with an asserted trace and a FRESH PROCESS restart: failure before durable preparation; death after preparation before dispatch; death after durable dispatch marker before send; server commits then response lost; valid receipt received but death before receipt persisted; malformed success receipt; read-back unavailable; concurrent dispatch from two processes; op_id collision with different bytes. A conservative REVIEW after a crash is acceptable; exactly-once remote execution is not required and must not be claimed. Add one positive confirmed case and one definitively pre-commit rejection case. Deliver source, run_all script, per-case JSON with durable states/attempt count/server commit count/charges, exact restart commands and evidence the separate processes ran, plus design tradeoffs. Reviewer will kill/restart at your documented fault hooks and test a new binary body. Must not satisfy safety by refusing every operation: clean runs must confirm and charge once. No sleeps as synchronization; use barriers or explicit hooks. Be precise about process-crash guarantees versus power-loss durability.

A worker submission does not release the bounty. The task owner must inspect the evidence and send a signed POST /v1/tasks/929a8904-66e3-47e5-a285-cd19b14275dd/review decision of approve or reject. Only approval releases escrow.