Skip to content

experiment(exp63): prove rule-overlay coverage loss and real Aegis fix - #52

Open
heybeaux wants to merge 7 commits into
mainfrom
nightly/2026-10-13-exp63
Open

heybeaux wants to merge 7 commits into
mainfrom
nightly/2026-10-13-exp63

Conversation

@heybeaux

@heybeaux heybeaux commented Oct 14, 2026 •

Copy link
Copy Markdown
Owner

New nightly development: exp63 / Spec69 / RT53

Paired Aegis PR follows immediately. Unique question: can opaque matcher/scope cascade replacements silently remove enabled protection while severity stays strict? Deduplicated all specs/README/claims and prior unmerged exp60/61/62; old24/25 already built, not repeated.

Pre-registration c8bc120; BEFORE observations correction060e6b8 pins high/critical deny on every arm. Two fetched-main real-Aegis baselines4be84ed: unsafe allow10/10, accuracy14/25. Runtimebc879d6: SAME25 scenarios/seed/harness, unsafe allow0/10, rejection10/10, accuracy25/25. Controls10/10 base +15/15 candidate behavior. No tuned thresholds, holdout unused.

Committed trace+source/dist SHA receipts, full replay37 events ×3; CLAIMS/SYNTHESIS/JOURNAL/EVIDENCE updated. Full local build/typecheck green; verify:evidence 49 claims105 traces983 assertions. New executable CI builds core+exp63 using isolated locked toolchain (historical absolute file deps remain untouched), audits entire evidence corpus/receipts, checks out exact Aegis runtime commit, freshly reruns and checks metrics/green, not just exit0. Identical clean-worktree CI rehearsal passed.

HOLD / uncertainty

Runtime policy tightening needs owner migration review (opaque regex broadening also needs explicit consent). Local evidence is not remote CI/merge confirmation. Prior unmerged60/61/62 not combined. Aegis integration is real public mergeLayers->loadPack->evaluate, commands are NEVER executed. No regex inclusion, post-merge alias, predictor calibration or production-incident claim. No .acr modifications or Territory refresh. Both primary dirty checkouts and existing traces preserved.

Paired PR: heybeaux/aegis#69

@coderabbitai

coderabbitai Bot commented Oct 14, 2026 •

Copy link
Copy Markdown

Important

  • 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration
  • Configuration used: defaults
  • Review profile: CHILL
  • Plan: Advanced
  • Run ID: 3484d427-4128-498f-abeb-f99230b3fe86
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@heybeaux

Copy link
Copy Markdown
Owner Author

Final headd0ca6d9 executable exp63-evidence SUCCESS run38524070942: core/new-experiment build+typecheck, entire corpus49claims105traces983assertions, exact before/after audit and fresh frozen exp63 against pinned real Aegisbc879d6 GREEN. Paired Aegis69 head88548b2 release-check SUCCESS run38524236053;1146 tests, stored50/50. CLEAN/MERGEABLE, but paired changes held for owner migration review of new matcher-retune consent; not merged. CodeRabbit success explicitly review-skipped. New experiment evidence—not an old15/16/17 replay.

@heybeaux

Copy link
Copy Markdown
Owner Author

Observable-interface audit reproduced verifier-only false acceptance: --latest reused the old committed green, missing secondary metrics passed, and altered row fields could hide behind self-check booleans. Fixed in a066950: complete fixed metrics/row recomputation; immutable frozen plan; exact newly-created run ID + prior IDs/start-time check; live source/dist hash validation.15 verifier tests pass including artifact mismatch. Executable CI run38525410093 SUCCESS on a066950 freshly reruns the SAME unchanged experiment green. Original red/red/green traces/thresholds/harness untouched; audit is NOT a second novelty claim. Aegis runtime remainsbc879d6;1146 release tests/stored50/50 and105840 boundary combinations independently passed. Migration review hold persists; not merged.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants