Skip to content

[LLM-Wiki-v2][F1] Build real and formal wiki verification skills; close safety and schema gaps #874

Closed

Description

@tommy-ca

Logical parent: #871. Depends on: #873 (plain issue links; no native GitHub subissue claim).

Scope and acceptance

Apply canonical skills/create-verification-skill/SKILL.md to generate a host-native project-local verify-llm-wiki CLI drive with Launch/Doctor/Drive/Evidence/Cleanup/Helpers and deterministic isolated scratch instances. Apply skills/create-formal-verification-skill/SKILL.md to model source immutability, page promotion/admission, and concurrent writer/linearizable update properties; pin code spans and sync SHA. Check a deliberate broken variant causes failure (negative control). Add Python tests for no overwrite, basename collision, symlink/path confinement, invalid page type/frontmatter, unindexed/deeply nested page, link resolution, conflicting concurrent ingests, torn log/index, stale digest and CLI exit codes. Repair defects only when independently verified. Formal TLC/SMT unavailable => BLOCKED, not proof. Keep functional vs formal verdicts distinct.

Evidence policy

Acceptance requires a reviewable PR and exact-head reproducible receipts. Planning is not implementation; no claimed Swarm/Arena/Interrogate model execution without actual independent work.

Partial implementation and exact-head evidence (2026-10-31)

  • Stacked draft PR #879 on docs(wiki): audit reference engines and propose governed migration (#871) #878 adds scripts/llm_wiki.py raw write-once/exclusive creation, link confinement and atomic index replacement, 11 new negative safety tests, 39-history bounded real-code checker + failing overwrite mutant, native Codex verify-llm-wiki and verify-llm-wiki-formal skills, optional NOT_EXECUTED TLA+ model, and offline CI wiring.
  • Earlier implementation head 674b2bfb354f71e8674c4cf5fa33e10dd9868262: Actions 41214985678 SUCCESS. Inspected job log: original wiki tests 9 passed, new safety negatives 11 passed, bounded model tests 3 passed; standalone model reports 39 bounded actual-code traces PASS and deliberate conflicting overwrite mutant detected at two-step trace A→B. Skill scaffold check PASS. This is offline functional evidence only.
  • Current successor head 0029b35fad97aa19fd67c42f35bd0ff3ec6f15ce includes schema/wiki documentation changes; Actions 41215096803 was IN_PROGRESS at this inspection. Do not treat prior-head CI as exact-head proof of later changes.
  • Remaining OPEN: actual TLC execution/negative TLA check, independent Swarm/Arena/Interrogate and Thermos reviews, crash-safe source+human-log transaction, adversarial directory-swap confinement (beyond simple preflight), independent source/rights review, multi-host observed-live coverage. Source+log writes are not atomic and not claimed as such. This child remains OPEN until acceptance gates are met.

Activity

  1. added 19 commits that reference this issue on Oct 31, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions