Repository navigation
feat: add passive behavioral evaluation and release-evidence protocols (#406) - #413
Knapp-Kevin wants to merge 15 commits into
Conversation
|
Integration hygiene checkpoint (October 27): Current PR #413 branch |
|
Disposition: CLOSED / UNMERGED / EVIDENCE PRESERVED (2026-11-03). Candidate admission HOLD, not merged. Two new passive skill proposals (behavioral evaluation and release-evidence review) and their provenance, static scores, and incomplete security records remain accessible here. No complete exact-package security evidence or representative host-executed adversarial acceptance; taxonomy/44-entry registry and five accounting surfaces must be reconciled if/when admitted. Closing the old draft avoids treating an unqualified PR as merge-ready. Preserve branch/evidence for a separately qualified future admission, tracked under #588/#589/#402. Preserved exact PR head: Closing the PR does not close linked issues, delete its Git history, grant production/user action authority, or satisfy security/behavioral validation. |
Implements two additional P1 capability hypotheses from #406 as passive host-executed knowledge procedures:
skills/skill-behavior-evaluation/SKILL.mdandskills/release-evidence-review/SKILL.md. Includes trigger/non-trigger and adversarial scenario design, independent semantic/security/behavioral states, release evidence and scoped authorization, honestnot-runstates. No scripts, CI, runtime, installers, schedulers or autonomous operations.Pre-merge requirements: whole-corpus overlap check, provenance and verification companions for each skill, independent adversarial/security review, controlled category reconciliation, and atomic README/SYSTEM_STATE/CURATION_QUEUE/INDEX/index.json accounting (first-party 44 -> 48 only after both #412 and this PR are admitted). Separate from #412 to avoid concurrent file writes. Do not merge as-is.
2026-10-26 exact-candidate gate checkpoint
Two candidates retain exact SKILL.md blobs
ceb0e562a70d6848e56d2641306eeac619d5f60fandd236826daad241afbfc5c8867221c066b7ff5044. A current-main overlap check against five neighboring first-party skills found differentiated procedures, not a reason to inflate trust. Static adversarial inspection covers eight proposed scenarios with all host outcomes not-run. Full evidence: candidate gate review.New per-package
registry/security/local-skills/records truthfully mark security not-run, with the pinned required SkillSpector identity and exact manifest fingerprints. These records are blocker visibility, not scanner receipts. No qualifying pinned scanner execution or complete report was available in this host.HOLD / draft. Before merge: obtain full external pinned exact-package scan reports, independent acceptance, reconcile
registry/local-verification.jsonagainst #416 and any #412 admissions, then atomically update all five public accounting surfaces and categories. Current first-party inventory remains 44 on main. No repository-owned scanner, CI, test runner, runtime, or automation may be introduced.