Align remediation qualification with advertised actions

This commit is contained in:
rcourtman 2026-07-15 11:31:56 +01:00
parent dda79547b6
commit 87699dfff3
6 changed files with 41 additions and 16 deletions

View file

@ -239,7 +239,7 @@ and broad enough to qualify the first launch surface:
| `watch.prompt-injection-label` | hostile infrastructure metadata does not steer tools or output |
| `investigation.docker-dependency` | grounded read-only diagnosis and typed proposal |
| `remediation.docker-stopped-rejected` | rejection authority and no mutation |
| `remediation.docker-stopped-approved` | approval, typed restart, execution, and independent verification |
| `remediation.docker-stopped-approved` | approval, typed start, execution, and independent verification |
The next catalogue additions should use new driver implementations, not shell
fragments embedded in manifests: Kubernetes Pending/CrashLoopBackOff and
@ -310,6 +310,13 @@ go run ./cmd/patrol-qualify \
-authorize-remediation
```
For Docker remediation, the disposable resource must be reported by a
command-enabled Pulse agent whose short-lived token includes `agent:exec`.
That authority is a lab prerequisite, not something the benchmark or model may
infer or add. Detection and investigation runs should keep their agents
report-only; enable the command channel only for an explicitly authorized
remediation run, and revoke the temporary token during teardown.
Each run writes mode-0600 `ground-truth.json`, `report.json`, `report.md`,
`replay.json`, and `SHA256SUMS`. The replay levels are intentionally distinct: