release(v8.7.0): forward-direction lossless verification (MR-ROUNDTRIP) - #95
Merged
Conversation
…Gate 4 representation-foreclosure P0/P1 from the dogfood56/57 A/B: a writer that rounds an exactly-representable value to a lossy canonical string (format(x) lossy at rc=0) slipped through every parse-side check. Adds forward-direction coverage as a property-based mechanism across the executed sweep, AC-authoring, producer lock-in, and the determinism hook. P0-1 (forward MR-ROUNDTRIP, the central fix): - accept-set-conformance-harness.md: MR-ROUNDTRIP relation + worked shape (run_writer_roundtrip / make_value_corpus; grammar-derived inter-anchor band; independent parse oracle) as the W-axis forward counterpart to MR-CANONICAL. - ac-evaluator.md (+ -hi twin, byte-identical): L-ROBUSTNESS MR-ROUNDTRIP duty; persisted sweep fields roundtrip=/intermediate-sampled=; intermediate-sampled=n self-incrimination on a triggered W boundary. - accept-set-verify.sh: P5 forward-depth gate (W boundary roundtrip=y intermediate-sampled=n -> BLOCK), parallel to P2 astral, same kill-switch/fail-open. P0-2 (authoring): Gate 9 R1 round-trip clause requires a forward generator sampling the inter-anchor intermediate band; an anchors-only grid no longer satisfies R1. P0-3 (durable lock-in): producer (test-writer/implementer/test-authoring-guidance) commits a GENERATIVE property-test on TRIGGER, unconditional on a found leak, not a leak-conditional literal -- closes the generation gap both A/B builds reproduced. P1 (confound): Gate 4 representation-foreclosure rule (a foreclosed numeric type / precision model under an advertised lossless/exact END is a FAIL) + planner self-audit wiring + brief Out-of-Scope ENDS-not-MEANS guidance. Property-based + product-agnostic (CT-DECONTAM-1 = 0 hits); meet-or-beat (additive; parse-side MRs untouched). Verification: test-skill-contracts 862/862 (+CT-AASC-15..20), test-accept-set-verify 32/32, test-path-consistency 144/144, ShellCheck clean, ac-evaluator twins byte-identical (CT-EV-MODEL-1). P2 (plateau early-exit cost optimization) deferred to a separate increment. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GKqmv9NH3naVcoWAB3fr22
…ed round-trip sweep (dogfood58) dogfood58 (efficacy test of b07871e on a data-size format writer) found a SPLIT: the product was correct (format(1023)="1023 B", trap fixed) and authoring propagated (Gate 9 R1 intermediate-band reached the round-trip AC), but the new W-axis RUNTIME gate never engaged on the writer -- the round-trip was recorded under an off-grammar label (boundary=numeric-external roundtrip=n/a) instead of canonical boundary=W, so hook P5 was structurally un-triggerable. Root cause: the orchestrator's accept-set trigger had a K-axis recognition clause but NO W-axis sibling for the writer-pairing, and the evaluator was free to improvise a boundary label -- recognition-variance (the exact failure the AASC determinism hook exists to remove) re-entered at the label layer. Defense-in-depth, all property-based (no unit/format vocabulary): - #1a orchestrator (SKILL Step 3a): a W-axis trigger clause parallel to the K-axis one -- a ticket pairing a reader (parse/decode) with a canonical writer (format/serialize/encode) advertising round-trip/lossless/exact/canonical MUST include that AC in triggered-on= for the W axis, and the evaluator MUST encode boundary=W roundtrip=y. - #1b ac-evaluator (+ -hi twin, byte-identical): a deterministic label rule on MR-ROUNDTRIP -- a round-trip/invertibility sweep IS the W axis; never an improvised non-canonical label. - #2 hook accept-set-verify.sh: new P6 -- a round-trip-bearing sweep (roundtrip=y or intermediate-sampled=y) under any non-W boundary is non-conformant by construction -> BLOCK, closing the mislabel bypass of P5. Canonical A/U/K lines carry roundtrip=n/a intermediate-sampled=n/a so they never trip. - #3 harness: the MR-ROUNDTRIP corpus MUST deterministically include the maximal-mantissa magnitudes below promotion (not random samples that cap short) -- closes the dogfood58 [MEDIUM] just-below-next-anchor corpus gap. Verification: test-skill-contracts 865/865 (+CT-AASC-21/22/23), ShellCheck clean, ac-evaluator twins byte-identical (CT-EV-MODEL-1), CT-DECONTAM-1 0 hits. Functional hook proof (SW_ACCEPT_SET_CONFORMANCE_MODE=on): the dogfood58 line boundary=numeric-external ... intermediate-sampled=y now BLOCKS (P6); canonical boundary=W roundtrip=y intermediate-sampled=y passes; shallow boundary=W intermediate-sampled=n still blocks (P5); canonical boundary=A passes. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GKqmv9NH3naVcoWAB3fr22
Bump plugin.json to 8.7.0 and add the CHANGELOG [8.7.0] entry for the forward-direction lossless verification feature (commits b07871e + acd735e): the W-axis MR-ROUNDTRIP relation in the executed accept-set sweep, the deterministic W-axis writer-pairing trigger (/impl Step 3a), the hook P5/P6 forward-depth backstops, Gate 4 representation-foreclosure + Gate 9 R1, and the producer generative property-test lock-in. Pre-flight: test-skill-contracts 865/865 (+ CT-AASC-15..23, CT-MODE-13/14), test-accept-set-verify 32/32, test-path-consistency 144/144, ShellCheck clean, ac-evaluator / ac-evaluator-hi byte-identical sans name:/model:. Validated live by dogfood59 (paired parse+format subject): boundary=W roundtrip=y intermediate-sampled=y fired end-to-end across three tickets (corpus 9190 IEC + 9004 SI, 0 divergences), the shipped writer was correct + doc-truthful, and the committed suite (208/208) locks the round-trip + Unicode-digit complement RED. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JZoJM8ViKq39XLyYrWRVjx
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
v8.7.0 — forward-direction lossless verification (MR-ROUNDTRIP)
TL;DR. Adds the write-side counterpart to the parse-side MR-CANONICAL in the executed accept-set sweep. When a ticket pairs a reader (
parse/ decode) with a canonical writer (format/ serialize / encode) that advertises a round-trip / lossless / exact / canonical guarantee, theac-evaluatordrives exactly-representable values through the real writer and verifiesparse(format(x)) === xover a grammar-derived inter-anchor intermediate band — catching a writer that silently rounds an exactly-representable value to a lossy canonical string atrc=0, a class invisible to every parse-side relation.What's in it (commits
b07871e+acd735e+ release9fa5022)parse∘formatoracle. New persisted fieldsroundtrip=/intermediate-sampled=. Twins mirrored byte-identically. CT-AASC-15..22./implStep 3a) — a paired reader+writer advertising round-trip is itself the trigger even when the only lexical cue sits on the reader (parallel to the K-axis clause); the evaluator MUST encodeboundary=W roundtrip=y, never an improvised label.P5(shallow forward:roundtrip=y intermediate-sampled=n) andP6(round-trip-bearing sweep under a non-Wlabel) block recognition-independently. Fail-OPEN. CT-AASC-23.Verification
test-skill-contracts865/865 (+ CT-AASC-15..23, CT-MODE-13/14, twin byte-identity CT-EV-MODEL-1, CT-DECONTAM-1 0 hits)test-accept-set-verify32/32 ·test-path-consistency144/144 ·shellcheckclean · twins byte-identical ·plugin.json8.7.0 == newest CHANGELOG[8.7.0]P5/P6covered by CT-AASC-23 + functional probes; dedicated cases intest-accept-set-verify.share a tracked follow-up.Dogfood evidence
P6.parse+format,chain=on/uc=off) — confirmed end-to-end: orchestrator recognized the paired parse/format as the W axis, evaluator encoded canonicalboundary=W roundtrip=y intermediate-sampled=yacross three tickets, a 9190 IEC + 9004 SI MR-ROUNDTRIP ran with 0 divergences, the shippedformatwas correct + doc-truthful (README "Round-trip fidelity (caveat)"), committed suite 208/208 locks the round-trip + Unicode-digit complement RED. Zero product defects shipped.Additive + kill-switched
The W-axis relation only ADDS coverage; absent a paired canonical-writer boundary it never triggers; parse-side relations + A/U/K axes are byte-unchanged; the hook predicates are fail-OPEN.
SW_ACCEPT_SET_CONFORMANCE_MODE=off/constraints.accept_set_conformance: offreverts byte-for-byte to v8.6.0.See CHANGELOG.md
## [8.7.0].🤖 Generated with Claude Code