Team pilot runtime

1. Record the starting Git revision and initialize ignored local run logs.
2. Launch a Codex coding worker with the requested job and stream its output.
3. Build bounded, sanitized observations concurrently with coding. Include the diff against the starting revision, staged changes, new-file previews, worker output and acceptance results.
4. Ask Jev nine assessment questions through the configured provider. Save the observation and scores. Apply deterministic lifecycle, budget and completion rules.
5. When work is ready, run the programmer's fixed --verify command using /bin/sh in the target repository. This verifier is a local command process, not another Codex interpretation.
6. Accept verification only when that command exits zero and source fingerprints before/after match. Missing or failed verification escalates. Optimistic Jev scores cannot override it.
7. Refresh the fingerprint after the final model call and immediately before applying FINISH. Any source change invalidates the result. A subsequent coding worker also resets verification.
8. Return FINISHED only when fresh command evidence and semantic thresholds both pass. Otherwise continue within limits or escalate for review. Ctrl-C cancels and terminates active processes.

The worker_factory injection point only replaces coding workers. Acceptance always uses the configured local command. The offline demo explicitly uses a trivial simulation command.

Logs contain state, output, sanitized observations, probabilities, interventions and acceptance command/exit/fingerprint. Previous observations are excluded from subsequent observation inputs so the audit log does not recursively inflate model requests.

The programmer guide defines the source scope, limits and remaining weaknesses. The older theory documents describe the upstream experiment; this page describes the pilot.