Skip to content

v0.10.1 — independent re-grade + verify-forgery bypasses fixed

Choose a tag to compare

View all tags
@momomuchu momomuchu released this 05 Jul 13:14
v0.10.1
b5abdfd

v0.10.1 — independent re-grade of the closures + two verify-forgery bypasses fixed

An independent adversarial verifier re-ran revert-and-test on all five v0.10.0 fixes (scratch copy → baseline PASS → revert the fix → assertion goes red → substance check). All five confirmed real and teeth-backed — none cosmetic. It also found two undisclosed bypasses that made v0.10.0's "a DONE typed on broken code is refused" claim overstated. Both are now addressed, honestly:

Fixed

  • MNM_DONE_DIR decoy redirect (silent — the dangerous one). The live re-execution accepted an agent-supplied directory, so pointing it at a clean decoy granted DONE on genuinely broken code. The override is removed: re-execution now runs in the project cwd, whose cwd-relative DoD/test anchors trip INTEGRITY-COMPROMISED on any redirect. Locked by evals/run.sh::verify-forgery-decoy-redirect.
  • MNM_SKIP_REVERIFY self-bypass (test hatch) — now loud. It disables the re-execution (needed to isolate the log layer in unit tests) and can no longer be silent: a real DONE with it set prints WARNING … NOT independently re-checked to stderr, and it's named in SKILL.md's residual.

Corrected claim (honesty over closure)

The re-execution closes the dangerous decoy/broken-code case. It does not close, in-band: (a) a verify-row skip of the adversarial panel when the code genuinely passes the gates, or (b) an agent setting MNM_SKIP_REVERIFY. Both are the same fundamental in-band limit — a guard that closed them would either re-run the panel (violating the author≠verifier independence it exists to enforce) or trust an env the agent owns. The real roots stay the ed25519-signed manifest, the independent mnm-verify sub-agent, and CI re-running the gates in a clean environment — none controlled by the agent under test.

Verification

  • 546 unit tests, 0 failed · 17 self-eval scenarios, 0 failed · both CI workflows green
  • All five closures independently revert-tested; the two bypasses independently reproduced and re-tested after the fix
Assets 2
Loading