Build a Production Permit for a Free-Model Write Job

Description and use

Create a JSON permit and a checker script to ensure an LLM job runs only when all evidence and rules are satisfied. This provides fail‑closed behavior and rollback capability.

Problem

Free-model LLM jobs can falsely report success, risking unintended writes to production data. There is no built‑in fail‑closed mechanism to ensure that a job only runs when all safety checks pass.

Outcome

The permit system prevents accidental writes by requiring a signed JSON contract and evidence files. It provides clear rollback paths and enforces strict output validation against golden fixtures.

Workflow

• Create a folder structure with permit.json, check_permit.py, job.sh, fixtures, sandbox, out, and evidence directories.
• Write permit.json with fields: version, job, owner, lane, created_at, blast_radius (root, glob, max_files, max_bytes_out, deny_globs), evidence paths, and fail_closed rules.
• Add a golden fixture (good_note.md) and a thin fixture (thin_note.md) to test success and failure modes.
• Implement check_permit.py to read permit.json, verify evidence files exist, enforce lane and glob rules, and refuse if any check fails.
• Run check_permit.py before executing job.sh. If it exits with code 0, the job proceeds; otherwise it aborts.
• Use human_double_ack files to confirm real writes, enabling rollback by deleting permit.json or restoring previous output hash.

Evidence

The permit is one JSON file plus a checker. The model does not start until every gate has evidence.

Source

Rollback is one permit file on disk. Delete permit.json and the job must refuse.

Source