Monthly manual research intake¶
Reviewed 2026-10-11. This maintainer process creates no schedule, background task, notification or automatic corpus update. Once a month, a maintainer reviews candidates and records decisions in ordinary repository issues or PRs.
Preserve the source¶
Record primary URL, publisher, title, publication date, edition/version, access date and the exact downloaded artifact's SHA-256 where available. Distinguish incident and publication dates; retain later source revisions separately. Check licenses before incorporating payloads. Prefer original advisories, official framework documents and research papers over summaries.
A new publication is a lead, not proof that the plugin detects its mechanism. Do not fetch URLs or execute commands embedded in attack text. Framework IDs must include their edition; the coverage matrix records the current primary-source snapshots.
Admit evidence-backed regressions¶
A supported case needs:
- A specific boundary, legitimate task and actual controlled operation observed independently of model claims.
- Benign, vulnerable, resistant and no-exposure controls. Quoted/encoded data controls distinguish lexical matches from committed effects.
- Honest required capabilities/hooks; unavailable instrumentation cannot pass.
- Isolated synthetic identities/sessions/sinks, including reset, cancellation and budget checks. Synthetic policy is not production IAM/OAuth assurance.
- Stable scenario ID/version, corpus digest, grading predicates and declared target/config fingerprints. Changed behavior requires a new case version.
- Development and holdout material frozen before evaluation. Changes informed by held-out results require a new revision and retention of the old report.
- Privacy review across console/JSON/HTML/JUnit/examples. Exclude credentials, raw provider failures and unneeded content; redaction must disable exact regrading when it removes required facts.
Decide accept, experimental, defer or reject, with a short reason and follow-up. Preserve negative controls and unsuccessful runs rather than replacing them with favorable samples.
Record the result and its limits¶
Retain planned/executed counts, all outcomes, missing coverage, independent utility and per-class errors. A forbidden effect on benign data is a real boundary violation, not a detector false alarm. Identify calibration-eligible rows and exclusions. Related repeated samples are not independent evidence of population reliability.
Accepted work updates coverage with the mechanism, versioned ID and remaining limits. A framework label alone does not make an untested row tested. New live execution and model spend remain separately explicit actions; this manual intake process does not authorize them.