Skip to content

release(v8.4.0): failure-class coverage by default (Phase A+B+C) - #89

Merged
aimsise merged 1 commit into
mainfrom
release/v8.4.0
Jun 9, 2026
Merged

release(v8.4.0): failure-class coverage by default (Phase A+B+C)#89
aimsise merged 1 commit into
mainfrom
release/v8.4.0

Conversation

@aimsise

@aimsise aimsise commented Jun 9, 2026

Copy link
Copy Markdown
Owner

TL;DR

v8.4.0 turns the harness's depth-gated verification machinery ON by default and makes the gates generate failure-class coverage rather than only grading volunteered ACs — bringing the default tier to parity-or-better with a max-effort build on a 2-subject dogfood A/B.

What ships (three bundled phases)

  • Phase Aevidence_floor = max(tier floor, AC-shape floor) (any behavioral AC floors at +1-independent, Size-independent); Gate 7's adversarial trigger broadens to every external-input boundary; tier-independent strongest-derivation oracle + per-AC [ORACLE-AUDIT].
  • Phase B — a default always-on failure-class eval panel (L-CORRECTNESS / L-ROBUSTNESS / L-CONTRACT-CONFORMANCE / L-UNIFORMITY / L-SIMPLICITY, [EVAL-PANEL]); Gate 9: Failure-Class Coverage (per-boundary R1–R4 matrix); plus a dogfood-driven adversarial-key / prototype-pollution vector in L-ROBUSTNESS + Gate 9 R2.
  • Phase CEC-SELFDOC channel + doc-verifier agent (classes A/E); Gate 10: Peer-Set Uniformity (class D); a refute-then-synthesize merge replacing the lone-non-critical-FAIL-demoting majority-merge; and a Gate 9 R1 round-trip-losslessness cue.

Migration

Non-breaking; every mechanism is kill-switched (constraints.eval_panel / failure_class_coverage / peer_uniformity / selfdoc_verification / refute_merge, plus the existing independent_evidence / oracle_verification), all default auto. The refute-then-synthesize merge flips a default on exhaustive tickets (a lone non-critical FAIL now survives unless refuted); constraints.refute_merge: off restores the prior majority-merge. See the [8.4.0] CHANGELOG entry for the full migration paragraph.

Verification

  • bash tests/test-skill-contracts.sh 821/821, bash tests/test-path-consistency.sh 144/144, full run-all.sh ALL SUITES PASSED.
  • plugin.json 8.4.0 == newest CHANGELOG [8.4.0] (CT-MODE-14), real ISO date (CT-MODE-13).
  • ac-evaluatorac-evaluator-hi body byte-identical (CT-EV-MODEL-1) across all paired edits.
  • Validated by a 2-subject dogfood A/B (a CSV↔JSON library + a stateful key-value CLI) against a spec-only max-effort build.

See CHANGELOG.md [8.4.0].

🤖 Generated with Claude Code

Turn the depth-gated verification machinery ON by default and make the gates
GENERATE failure-class coverage, not only grade volunteered ACs.

Phase A — split the evidence-independence floor off the Size×risk depth axis
(evidence_floor = max(tier floor, AC-shape floor)); broaden Gate 7's adversarial
trigger to every external-input boundary; tier-independent strongest-derivation
oracle + per-AC [ORACLE-AUDIT].

Phase B — a default, always-on failure-class eval panel (L-CORRECTNESS /
L-ROBUSTNESS / L-CONTRACT-CONFORMANCE / L-UNIFORMITY / L-SIMPLICITY, with a
per-ticket [EVAL-PANEL]); Gate 9 Failure-Class Coverage (a per-boundary R1-R4
matrix); plus a dogfood-driven adversarial-key / prototype-pollution vector in
L-ROBUSTNESS + Gate 9 R2.

Phase C — an EC-SELFDOC evidence channel + a read-only doc-verifier agent
(classes A/E); Gate 10 Peer-Set Uniformity (class D); a refute-then-synthesize
merge replacing the lone-non-critical-FAIL-demoting majority-merge; and a
Gate 9 R1 round-trip-losslessness cue.

Every mechanism is kill-switched and feature-named (no bare measure labels —
they collide with the prior plan's M-numbering). Validated by a 2-subject
dogfood A/B (a CSV<->JSON library and a stateful key-value CLI). 821/821
skill-contracts, 144/144 path-consistency, full sweep green; the
ac-evaluator <-> ac-evaluator-hi body stays byte-identical (CT-EV-MODEL-1).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@aimsise
aimsise merged commit 0efd83e into main Jun 9, 2026
2 checks passed
@aimsise
aimsise deleted the release/v8.4.0 branch June 9, 2026 18:39
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.

1 participant