Review the evidence before an agent hands it on.

One complete local workflow

1. Read explicitly identified file revisions and compare their source hashes.
2. Compile a bounded packet and check that caller-labelled mandatory evidence survives. Hold on missing evidence or declared blockers.
3. Review the packet. Approval is bound to the tenant, agent, tool, destination, input, evidence and policy.
4. Save the dispatch intent in SQLite before the writer runs. Concurrent or restarted attempts cannot silently run it again.
5. Create an immutable artifact after checking the source revisions again. Save the observed receipt.
6. If acknowledgement is lost, inspect the destination and reconcile the observed artifact. A missing receipt never triggers an automatic second write.

Data stays in the local reference setup

The demonstration uses synthetic documents. The versioned file store holds source text and output artifacts; prepared plans contain selected text. A separate private gateway database stores approval and dispatch metadata. Both require owner-only filesystem access. The host is responsible for retention, backups and access control.

No cloud upload, payment, model inference or Fulcra account access is needed for this reference workflow. It runs outside the x402 endpoint and payment paths.

What still needs evaluation

This is a local implementation, not a live Fulcra deployment. The local file port checks revisions and writes atomically; the corresponding remote consistency and conditional-write semantics must be established before connecting an actual provider.

Reviewer identity is trusted host input. Hashes bind bytes; they do not establish identity, factual truth or complete evidence. Local crash and concurrency tests do not establish distributed exactly-once delivery, independent security certification or a production SLA.

Evaluate against the workflow you already use

Use the same clean, revised and incomplete packets with your existing controls and the local handoff. Compare missed changes, unnecessary holds, setup effort, reviewer time and accepted results. Continue only if the additional checks improve a recurring task.