Observed contrast
Different entry paths produced different governance outcomes. That divergence localised the defect to discovery before action rather than restraint after discovery.
Case study · AI governance and failure analysis
Project Airlock
A root-scoped governance policy worked after discovery—but an exact-path write could begin before the policy was ever read. The discrepancy became evidence, not an excuse.
The bootstrap problem
One AI instance inspected its destination, discovered the governing policy, submitted a receipt, exposed missing decisions and stopped correctly. Another received an exact destination and began writing without discovering the root policy.
The Airlock itself was not failing after activation. The failure happened earlier: a policy stored inside a project cannot govern an instance that never loads it.
Problem-solving sequence
From observed divergence to a testable repair.
Observe→Localise→Hypothesise→RedTeam→Reject unsafe repair→Formulate bounded repair→Define trigger→Define finish line→Execute→Verify
Evidence boundary: the first eight stages are evidenced. The preferred global-bootstrap and root-marker repair is formulated, not proven globally implemented. Execution and verification remain future work; acceptance test C is specified, not demonstrated.
Unsafe repair rejected
Renaming or marking every directory could break links, scripts and sync relationships while still relying on visual discovery. It was not adopted as the primary solution.
Preferred bounded repair
A globally loaded instruction checks the exact proposed target and its finite ancestor chain for the nearest Airlock marker before the first write—without recursively scanning a drive.
Finish line
The bootstrap ends with one observable result: no marker found, one applicable policy read, or a conflict reported. The authorised action then retains its own separate completion condition.
Professional relevance
Project Airlock demonstrates requirements engineering, human-in-the-loop authority, adversarial testing, failure localisation, risk analysis, acceptance-test design and a willingness to preserve “not yet proven” when a more impressive completion claim would be inaccurate.