The problem studied
An approval tied to an earlier revision must not authorise publication of a new version. A matching approval must still allow the legitimate workflow. Four remediation variants were compared against both criteria.
| Misuse refused | Legitimate workflow preserved | Model decision |
|---|---|---|
| Yes | Yes | Candidate cross-checked |
| Yes | No | Rejected: workflow broken |
| No | Yes | Rejected: misuse remains |
| No | No | Rejected on both criteria |
What is documented
The historical G1 0.2.0 acceptance record contains 83 passing tests on 25 October 2026, including 10 G2 preparation tests on configuration fixtures, without a VM. This total covers the whole local suite; it is not a count of security checks on a client website.
The limitations remain explicit
Synthetic data, demonstration identities and scripted AI. No real AI model, accepted VM isolation, protected client website or measured protection benefit. This note illustrates a cross-checking method; it is neither a client case nor an independent audit.
Own source: CY-017 acceptance record, historical TAP results and the closed model in the HAIDRA.DEV Cybersecurity project. Editorial summary: 30 October 2026.