Skip to content

v0.3.1 — inspector + runtime monitor

Choose a tag to compare

View all tags
@SollanSystems SollanSystems released this 03 Jul 16:58
loop-engineer--v0.3.1
1793a47
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Adversarial-fix milestone. The v0.3.0 release closed two false-POSITIVE classes
in the anti-cheat scanner; a GPT-5.5/xhigh codex challenge over the v0.3.0 diff
then found the blind side — evasion paths the scanner failed to flag, plus
boundary-validation gaps in three harness scripts. This patch closes them.

Fixed

Anti-cheat scanner false-negatives (P1.1–P1.5) — scripts/anticheat_scan.py

  • Scoped self-exclusion (P1.1). A scanner self-edit that empties or shrinks
    DEFAULT_GATE_PATHS / _ADDED_LINE_SIGNATURES is now graded critical
    (FailedSafety); additive and comment-only self-edits stay clean. Removed
    entries are compared semantically, so a reorder or reformat does not flag.
  • Delete + rename evidence (P1.2). parse_changed_files now also captures
    gate files that are deleted (+++ /dev/null) or renamed
    (rename from/rename to); both of Codex's exact exploit diffs now return
    clean:false.
  • verify-* gate coverage (P1.3). Gate-path matching now covers
    verify-fast / verify-full / verify-safety; tampering one to bypass it is
    flagged.
  • Broader tautology detection (P1.4). Identical-operand assertions (a literal
    or an identifier compared against itself) and always-true unittest calls now
    downgrade to FailedUnverifiable; honest asserts with distinct operands stay clean.
  • Path-shaped hidden-answer names (P1.5). Trajectory reads of held-out /
    hold_out / answer-key / golden / expected-output paths are flagged, while a
    plain assert result == expected stays clean.

Boundary validation (P1.6, P2.1–P2.4)

  • scripts/benchmark_harness.py — compare() raises on a mismatched A/B
    task-set instead of reporting a silent delta; non-bool claimed_done /
    verification_passed and out-of-range repair / criteria counts are rejected.
  • scripts/runtime_monitor.py — robust score parsing for 1e-3, negatives, and
    malformed input (no crash); tests pin the exact intervention per scenario.
  • scripts/inspect_loop.py — bounded shallow walk with a per-file read cap
    replaces the unbounded full-tree traversal.

Changed (P2.5)

  • README.md — present-tense install note corrected to "all 9 skills".
  • .claude-plugin/plugin.json — version 0.3.0 → 0.3.1.

Credits

  • The false-negative and boundary findings came from the GPT-5.5/xhigh
    codex adversarial review over the v0.3.0 release diff.
Assets 2
Loading