Skip to content

Add the evaluation proof: an equality checked by the kernel alone - #553

Merged
afflom merged 16 commits into
release/v0.5from
claude/evaluation-proof-main
Nov 16, 2026
Merged

afflom merged 16 commits into
release/v0.5from
claude/evaluation-proof-main

Conversation

@afflom

@afflom afflom commented Nov 15, 2026 •

Copy link
Copy Markdown
Member

Requested by Alex · project thread

Release line

This PR targets release/v0.5, the active line, which carries regenerate.yml and the spec/ projection. It was first opened against main by mistake. The branch was not rewritten. One merge commit joins the earlier main-based commits to their replay on release/v0.5. That merge's tree is the replay's tree, so the diff against release/v0.5 is the change alone.

What

This adds a language-1.2 semantic proof, evaluation ({"kind":"evaluation"}, notation evaluation{}). It lowers to:

run_tac _root_.Lean.Elab.Tactic.liftMetaFinishingTactic fun llGoal => _root_.Lean.MVarId.refl llGoal (check := false)

The lowering assigns Eq.refl of the left side, so the elaborator's unifier never compares the two sides. The kernel checks the theorem when it is added and rejects any statement that does not hold. A module that uses this proof meta-imports exactly Lean.Elab.Tactic.Basic and Lean.Meta.Tactic.Refl. The proof carries no tactic name, option or argument, and it adds no axiom.

Validation gates only the language version (1.2) and does not otherwise restrict the statement. A closed equality between two evaluations is the intended workload, not an enforced restriction, and the docs now say so.

Review asks (5911864478)

  • Schema: semantic-module-v2.schema.json gains the closed alternative with no arguments, and v1 stays without it. The unit test evaluation_proof_is_language_1_2_only checks both schemas and the exact serialization.

  • Conformance: SM-44 is registered in spec/src/IdsSM.lex.tex and projected to model/ids.toml, SPEC.md §31 and CONFORMANCE.md. It has a scenario in semantic-ir.feature, the case cases/semantic_ir/evaluation.rs and the named test conformance_sm_44. The case covers:

    • link, format to a fixed point, and read back through the notation
    • the emitted lowering and exactly two meta imports
    • verification of a true equation with an exact empty axiom policy
    • a false equation refused at verification (LLV7002)
    • a [propext] policy refused (LLV7005)
    • language 1.1 refused (LLT4001), also kept as the negative fixture evaluation-language-11

    The ID is SM-44 rather than the next free number because SM-43 is taken on Add the in-place byte update primitive (prim update) #443's branches.

  • Wording: SPEC (spec/src/Chapter17), the IR docs and the CHANGELOG no longer claim a general complexity bound. They state the measured case instead: reflexivity over long step-machine runs whose bytes are not literals exhausted the unifier's memory, while the kernel check of the same statements completed.

R3 chronology (honest disposition)

The implementation (2bf5c1e, then its replay 8a5f948) came before the register row, scenario and named test (83ab506). No failing scenario or test was observed before the implementation existed. The plants recorded in VERIFICATION.md were run after implementation, against the finished case. They show that the case detects each defect, but they do not establish red-before-green order and are not offered as such. For this change this is a gap against R3, recorded here rather than manufactured.

Records

Input repairs (authored inputs only; every line written by lexlean lock):

  • e281b4c4: the evaluation proof moves the language-1.2 compiler-semantics ID from ffc1e1ca… to c28ce79b…. Each of the 268 project locks that named it changes exactly that one line.
  • 893c4d7a: the evaluation-language-11 fixture's language-1.1 lock named cac35ee3…, an identity no compiler on this line produces. It now names e70124f1…, the same as cases-bool-language-11 at that head.
  • 01e0cda8 merges release/v0.5 at 7da6e937 (rust-std: store datatype fields behind shared boxes #441 and relock-identity). Base's side is kept for every conflict, so the language-1.2 identity is now base's a15f5fbc… and the 1.1 identity moved to 6dc2bc4d…. The CHANGELOG keeps base's relock-identity entry, then this change's entry.
  • 0ff22b08: the evaluation-language-11 lock moves e70124f1… → 6dc2bc4d…, the same as cases-bool-language-11 at this head, so check again exits 1 with LLT4001 as case.toml requires.

Producer runs (regenerate/evaluation-proof, one concurrency group):

Run Head Outcome Delivered
43743005902 1c3c77fe records writer failed at verify-examples --write (collections: LLC0102) nothing
43746421443 e281b4c4 oracles writer failed at check-fixtures --write (evaluation-language-11: LLC0102, exit 2); the other jobs were cancelled by the next push; no artifacts nothing
43754121985 893c4d7a success be299dac (records for the 893c input only)
43773463121 01e0cda8 oracles writer failed (evaluation-language-11: LLC0102, exit 2, stale 1.1 lock after the merge); cancelled by the 0ff22b08 push nothing
43784837347 0ff22b08 success (identity relock, three writers, combine, counts) cdb94ebd

The failed and cancelled runs' logs are retained as they are. They count as diagnoses, not as delivered output. Records for the composed head were written only by run 43784837347's deliver job, as cdb94ebd on this branch.

Verification (local, Lean 4.32.1)

  • cargo test -p lexlean --lib: 227 tests pass.
  • conformance_sm_44 passes.
  • cargo clippy --workspace --all-targets --all-features -D warnings is clean.
  • validate-model reports 522 ids, and validate-spec-links reports 522 rows.
  • lexlean lock --check passes on the relocked examples/collections, examples/language-1.2 and stdlib.
  • evaluation-language-11 at 0ff22b08: check exits 1 with LLT4001.
  • After the release/v0.5 merge, validate-model reports 522 ids, clean.

Why

uorc (UOR-Foundation/compression) replaced three byte-index trees with prim update. Seven of its authority theorems are reflexivity over two runs of a step machine, and after that change the unifier needed more than 13 GB for them. The kernel checks the same theorems in 6–8 s and rejects a planted false one.


Generated by Claude Code

An evaluation proof (language 1.2, {"kind":"evaluation"}, evaluation{})
lowers to MVarId.refl (check := false) through run_tac: it assigns Eq.refl of
the left side without the elaborator's unifier comparing the two sides, and
the kernel decides the theorem by reduction when it is added, rejecting a
false one. The unifier compares unevaluated subterms lazily; reflexivity over
two runs of a step machine whose state stores unevaluated bytes needed more
than 13 GB there, while the kernel checks the same theorems in seconds. A
module with such a proof meta-imports Lean.Elab.Tactic.Basic and
Lean.Meta.Tactic.Refl. SPEC section 16 and the changelog describe it; a unit
test holds it to language 1.2.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The fixed evaluation lowering has no concrete kernel trust bypass identified in this bounded source review: pinned Lean Refl assigns Eq.refl and skips elaborator isDefEq when check=false; it does not add an axiom. This is source credit, not acceptance of the new capability.

Owner completion blockers:

  • schemas/semantic-module-v2.schema.json $defs.proof.oneOf lacks evaluation although Rust accepts it. Published schema clients reject the new JSON, and schema-derived AU-01 samples omit it. Add the closed zero-argument v2 alternative and its roundtrip coverage; keep v1 rejecting it. The generic notation fallback already handles the tag, so no notation parse failure is claimed.
  • The added unit only parses IR/language versions. Complete registered scenario/named conformance coverage for the actual emitted proof, true and false equality through verification, exact axiom audit and canonical roundtrip, then the serialized official generated records/bootstrap products and final-head gate. Preserve/link genuine existing observations; do not manufacture R3 history or repeat completed producers. The body explicitly leaves these obligations incomplete.
  • SPEC2678–2689 and the Rust docs overstate general linear-time kernel reduction; the inspected Lean source proves skipped unification, not that complexity bound. Scope performance statements to the measured workload. Clarify whether 'closed equality' describes the workload or an enforced restriction: current validation only gates the language version.

Please also establish the intended release line: this PR targets main with no milestone; preserve the repository's release-line and actual forward-merge history requirements. Existing contributor retains implementation; no root retarget, competing patch or new local gate. Receipt30936dccc758a5e6b7ad936a84b102d64d5633c6ccbdbb3fbb2e7f5b4f1139f7. COMMENT only; full acceptance remains absent.

@afflom afflom mentioned this pull request Nov 15, 2026
- schemas/semantic-module-v2.schema.json gains the closed zero-argument
  {"kind":"evaluation"} proof alternative; v1 stays without it. A unit
  checks both schemas and the exact serialization.
- SPEC, CHANGELOG and the IR docs no longer claim a general complexity
  bound: the measured case is stated, and that validation gates only the
  language version (closed equalities are the intended workload, not an
  enforced restriction).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
An evaluation proof (language 1.2, {"kind":"evaluation"}, evaluation{})
lowers to MVarId.refl (check := false) through run_tac: it assigns Eq.refl of
the left side without the elaborator's unifier comparing the two sides, and
the kernel decides the theorem by reduction when it is added, rejecting a
false one. The unifier compares unevaluated subterms lazily; reflexivity over
two runs of a step machine whose state stores unevaluated bytes needed more
than 13 GB there, while the kernel checks the same theorems in seconds. A
module with such a proof meta-imports Lean.Elab.Tactic.Basic and
Lean.Meta.Tactic.Refl. SPEC section 16 and the changelog describe it; a unit
test holds it to language 1.2.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
- schemas/semantic-module-v2.schema.json gains the closed zero-argument
  {"kind":"evaluation"} proof alternative; v1 stays without it. A unit
  checks both schemas and the exact serialization.
- SPEC, CHANGELOG and the IR docs no longer claim a general complexity
  bound: the measured case is stated, and that validation gates only the
  language version (closed equalities are the intended workload, not an
  enforced restriction).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
The case links, formats, reads back through the notation, builds and
verifies a closed equation proved by evaluation{}, checks the emitted
lowering and its two meta imports, and refuses a false equation
(LLV7002), a wider axiom policy (LLV7005) and language 1.1 (LLT4001,
also as the negative fixture evaluation-language-11). The register row
and the SPEC paragraph live in spec/src (IdsSM, Chapter17); SPEC.md,
model/ids.toml and CONFORMANCE.md are their projections. VERIFICATION
records the plants.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
… Unreleased

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
The pull request moves from main to the active release line. This merge
joins the main-based commits with their replay on release/v0.5 (SM-44,
the v2 schema alternative, scoped documentation); its tree is the
replay's, so the diff against release/v0.5 is the change alone.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
@afflom
afflom changed the base branch from main to release/v0.5 November 15, 2026 17:46
@afflom afflom changed the title Add the evaluation proof: a closed equality checked by the kernel alone Add the evaluation proof: an equality checked by the kernel alone Nov 15, 2026

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The prior source requests are addressed at1c3c: v2 gains the closed evaluation alternative while v1 excludes it; SM44 has the registered scenario/named case and actual positive/false/policy/language/roundtrip verification coverage; wording now distinguishes the intended closed-equality workload from enforced restrictions and removes a general complexity guarantee. SM44 avoids the inspected SM43 allocation. Release target is now v0.5; the merge preserves original history and its tree370db269 equals replay5c62e58 rooted at8a8e.

Two distinct retained logs identify the remaining delivery blockers. CI honesty132903391374 reports348 stale generated files, including mismatched compiler semantics and missing stdlib build products. That job checked out4b462b0 merging into main71def, so its binding is not the current intended release composition. More immediately, exact-head regeneration43743005902 records job132900076443 fails at verify-examples --write: collections fmt --check emits LLC0102 because lexlean.lock is stale or noncanonical. Capture/upload/delivery did not complete. Repair the authoritative input-lock/regeneration ordering with the existing serialized owner; this review does not request a blind retry or competing writer. Causes of the other two writer failures were not inferred from this log.

Current coverage and documented plants are credited, but implementation precedes the visible register/scenario/test commits. Preserve genuine original observations and disposition that R3 gap honestly; postimplementation plants do not establish the original chronology, and no rewritten/manufactured red is requested.

Full acceptance still needs delivered coherent products, final-head intended-release qualification/reproducibility and actual merged-tree equality. Reuse source credit rather than repeat it. Receipt d33588ae8a67fbd54bc16ad63888ae177fa5a1d25b7fb735062e4526b1636781 plus producer supplement dafa471eb13ad0409da78152d56b157b24a6c5cb4535268998098efadd9fb4a4 preserve both full logs/SHA256s. COMMENT only.

The evaluation proof moves the language-1.2 compiler-semantics ID from
ffc1e1ca to c28ce79b. `lexlean lock` in each of the 268 projects that
named it writes the moved digest; nothing else in them changes. The
records that name it stay with regenerate.yml, which runs fmt --check and
lock --check on these inputs before it writes anything.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bounded e281b4c follow-up: the complete one-commit delta contains268 modified lexlean.lock files. Each changes exactly one compiler_semantics line from ffc1e1ca… to c28ce79b…, with no other patch lines. Previous authored-source/schema/conformance repairs remain preserved. This credits the input-lock source repair; official lock execution is owner-reported and was not replayed or independently recomputed here.

The revised body now explicitly records that no failing named scenario/test preceded implementation. That honest disposition resolves the ambiguity, but does not turn postimplementation plants into original R3 chronology or waive that requirement.

Existing regenerate43746421443 is active at this exact head; let its owner finish and review coherent output adoption before final-head qualification and actual merged-tree acceptance. Repro43746407755 passed, CI remains incomplete. Preserve failed1c3c receipts and avoid another input wave merely to repeat successful source review. Other owner producers493/497 are also active on distinct inputs; retain their separate custody and coordinate subsequent waves before starting more. No writer, retry, cancellation or competing patch requested. Receipt478c52f9e3d7cf9c9266e5d30a021e9ca77ba04b5405bab9a0583f907147baa8. COMMENT only.

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Concrete remaining input blocker at e281: existing producer43746421443, oracle job132910608937, explicitly checks out this head. check-golden --write completes, then check-fixtures --write fails at tests/negative/evaluation-language-11: check exits2 with LLC0102 (compiler semantics changed since locking), while case.toml requires exit1. This is not a timeout and does not reach the intended LLT4001 language refusal.

The fixture's project/lexlean.lock still records language1.1 semantics cac35ee3…. The adjacent cases-bool-language-11 project lock has identical bytes except e70124f1… in that field, with the same configuration digest. The credited268 language1.2 ffc1→c28 repairs did not cover this distinct fixture. Existing owner: repair this authoritative input using the correct language1.1 identity and official writer in the coordinated scope; preserve the expected exit and diagnostic, and do not substitute the language1.2 identity.

Earlier oracle/fixture writes occurred, but this failed job did not complete capture/delivery. Production-export and records were still active at observation. Disposition those existing results and applicable retained artifacts before proposing any bounded continuation; this does not call for a blanket producer repeat. Final-head full acceptance and the honestly recorded R3 gap remain.

One new log fetched once, saved SHA256 5b281f14245a270dc2613fcda9ad0852f6d88f496b38b1987bc32f62a56b4be6; receipt pr553-e281-oracle-failure-20261115/receipt.json SHA256 0fe468a9d15ae8ff86b5b96cd09ea47158bdd45600423386fd9f9f27b88f9d1c. No execution, writer, retry, cleanup or CI control.

…dentity

Its project lock named cac35ee3, a language-1.1 compiler-semantics ID no
compiler on this line has, so `check` stopped at LLC0102 (exit 2) before
the language refusal the case expects. `lexlean lock` writes e70124f1,
the identity the other language-1.1 fixtures (cases-bool-language-11)
record; `check` now exits 1 with LLT4001 as case.toml requires.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy

afflom commented Nov 15, 2026

Copy link
Copy Markdown
Member Author

Fixed in 893c4d7a: the evaluation-language-11 fixture lock.

The fix: lexlean lock in tests/negative/evaluation-language-11/project replaced the stray language-1.1 identity cac35ee3… with e70124f1…. That is the identity cases-bool-language-11 and the other language-1.1 fixtures record. Only the compiler_semantics line changed, and the case's expectations are unchanged. Run locally, check now exits 1 with LLT4001, as case.toml requires.

Other locks: I checked every other lexlean.lock on the branch. Each one either carries the moved 1.2 identity or is unchanged from release/v0.5.

What happened to run 43746421443 (at e281):

  • write (oracles) failed at check-fixtures --write because of this fixture, so it uploaded no part.
  • combine needs every writer to succeed, so that run could not combine or deliver anything.
  • Its production-export and records writers were still running when I pushed regenerate/evaluation-proof at 893c4d7a. In the same concurrency group, that push replaces the run, so there is still a single writer at the PR head.
  • The failed e281 and 1c3c logs stay retained as they are.

Generated by Claude Code

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The893c repair addresses the diagnosed fixture input: its sole changed line replaces language1.1 cac35ee3… with e70124f1…, preserving the fixture language and expected refusal. The owner commit reports an official lock write and LLT4001/exit1; separately, existing replacement producer43754121985 has completed its oracle job and uploaded artifact14178142867. Records/production remain active, so output adoption and final acceptance remain pending.

Prior e281 producer43746421443 ended with one failed oracle job and five canceled jobs; its artifact listing is empty. Preserve the original failure receipt, and do not count that run as delivered or aggregate canceled jobs into acceptance. Root has not started or retried a writer. Existing owner retains current output disposition, final qualification and the honestly recorded R3 chronology gap; the PR Records paragraph still names e281 as current and should be updated when recording final custody.

Bounded receipt pr493-0317-pr553-893c-repairs-20261115/receipt.json SHA256fceee3df7b95946a1473a080ed42094c98c14c5aff2cea277b02902d6dbdfd2e. This credits the input repair and observed oracle completion, not a full feature acceptance. No tests/writer/retry/log refetch performed.

The write recipes of SPEC.md §9.2, run by regenerate.yml (https://github.com/UOR-Foundation/lexlean/actions/runs/43754121985),
rewrote 5348 paths.

The counts writer reported:

    counts: no count moved

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The existing producer43754121985 has now completed all six jobs, including combine/counts/deliver. be299 is the direct bot-authored output child of893c and names that run. Its delivered artifact14182507554 is input-bound and nonexpired. The prior1c3c/e281 failures/cancellations remain preserved as diagnoses, not delivered outputs.

Complete path/hash classification across eight non-truncated subtree pairs finds only generated products:4,896 expected-product paths,351 compiler Rust-product paths and328 production-export paths. All other root entries, including implementation, model/schema/spec/scenarios, workflows and verification record, retain their exact reviewed hashes. This credits delivered output custody and source preservation; artifact ZIP bytes were not compared, and product semantics/full acceptance are not inferred from path classification.

No producer repeat is needed to adopt outputs already present. Existing owner should record the producer's completed outcome in the body, finish exact resulting-head qualification, and reconcile the current release/v0.5 composition before any landing. The honestly recorded R3 chronology gap remains; output delivery does not waive it. Current be299 CI43767395945 is queued.

Immutable receipt pr553-be299-product-adoption-20261115.json. No execution, writer, retry, cleanup or log refetch.

- lexlean::LANGUAGE_1_2_SEMANTICS_ID names c28ce79b, the ID the tree
  computes (RP-10, RP-19).
- README's semantic IR capability row claims SM-44 (RP-11).
- spec/src/IdsSM.lex.tex is its own canonical source (lexlean fmt;
  AU-04, NS-01).
- The downstream snapshot consumer's exhaustive match names
  SnapshotProof::Evaluation (SM-16).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy
…ation proof

Authored conflicts: CHANGELOG keeps both Unreleased entries. Every other
conflict was a generated record, a lock or LANGUAGE_1_2_SEMANTICS_ID; each
takes release/v0.5's side (records it no longer has are removed), and the
regenerate run's identity job relocks and rewrites them for the merged
compiler in one pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bounded follow-up credits the new source repairs: README now includes SM-44, SnapshotProof::Evaluation is counted as a one-node leaf by the exhaustive conformance visitor, and IdsSM has canonical spacing. Exact01e0 parents are bbf92e8 + accepted7da6e937. Complete crates subtree comparison finds only four merge-delta blobs, each inherited exactly from7da6; the reviewed Evaluation backend/IR/schema/case source remains. #441/#545 integration is reused, not re-reviewed or reimplemented.

Historical be299 product custody remains valid at its original input. It does not qualify this new composition: CI43773467835 honesty reports348 stale generated files, with recorded compiler a15f5fbc versus current a1477c2e; repro43773467747 fails collections lock LLC0102 stale/noncanonical. Both new failed logs checked synthetic00221fd (01e0+7da6). Aggregate platform refusals/skips are not passing tests or independent diagnosed platform failures.

Existing producer43773463121 is already queued. Smallest owner next step is to observe that run, review its delivered identity/products against the exact composed input, update the stale body, then qualify the resulting final head and actual merged tree. No duplicate producer or retry requested. Prior R3 gap remains explicit, without waiver. Immutable receipt /workspace/work/pr553-01e0-delta-review-20261116.json; SHA256 cd545b7f2ca09d532ab5009ab4b79a2636f8082acb9a5ae1aa8544e8b7c0914c.

The fixture must fail at phase link (LLT4001, exit 1), not at the lock
check (LLC0102, exit 2); the merge of release/v0.5 moved the 1.1
compiler-semantics identity to 6dc2bc4d.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Bounded delta credit at0ff22b08: sole parent01e0; the complete change is one compiler_semantics line in tests/negative/evaluation-language-11/project/lexlean.lock, e701→6dc2bc4d. It matches the exact-head language1.1 peer fixture. Expected exit1/no-artifact refusal and all proof/backend/IR/schema source are unchanged. This repairs the authored fixture input; it is not independent runtime confirmation or generated-product acceptance.

Old producer43773463121 is canceled and retained separately. Replacement43784837347 is already queued at0ff2; CI43784844539 queued/repro43784844656 active at inspection. Existing owner next: observe that run once, review delivered identity/products, then qualify the final composed head and actual merged tree. No duplicate producer or retry requested. Historical893c→be299 delivery remains credited only to that input.

The PR body still describes893c/e701 and43754121985 as current/running; reconcile it with completedbe299 delivery, actual7da6 integration, this repair and the current producer. The candid original-R3 gap remains, without waiver or reconstructed failure. Receipt /workspace/work/pr553-0ff2-delta-review-20261116.json; SHA256 d0e9b7813169efe1b8726ae340d89884ed45da9bf9895fc1c7af4a5ac1af00a0. No new logs, local execution, writer or CI control.

afflom commented Nov 16, 2026

Copy link
Copy Markdown
Member Author

honesty and repro on 0ff22b08 fail for the expected reason: the composed head moves the language-1.2 compiler-semantics identity (a15f5fbc… → a1477c2e…), so the 348 committed generated files and the project locks are stale until the regenerate run delivers them. Run 43784837347 at 0ff22b08 is producing them now (identity relock, then the three writers). A local relock-identity followed by check-fixtures shows no lock failures remaining; only golden outputs that fixtures-write rewrites differ. CI will be read again on the delivered head.


Generated by Claude Code

afflom commented Nov 16, 2026

Copy link
Copy Markdown
Member Author

The x86_64-apple-darwin failure is on be299dac, an earlier head, before bbf92e85 (the conformance fixes for RP-10/RP-11/RP-19/AU-04/NS-01/SM-16) and before the release/v0.5 merge. Its conformance harness invocation (moves) ran without those fixes. It is superseded; the Intel host tests will be read on the head the current regenerate run delivers.


Generated by Claude Code

The write recipes of SPEC.md §9.2, run by regenerate.yml (https://github.com/UOR-Foundation/lexlean/actions/runs/43784837347),
rewrote 6833 paths.

The counts writer reported:

    counts: no count moved

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Product adoption credited at cdb94eb, the sole child of reviewed 0ff22b0. Producer 43784837347 completed all seven jobs, including identity, counts/stack and delivery. Nonexpired artifact 14199095547 is bound to 0ff22b0, digest 754bf8f9251b1c3183a1a3b0533e1ab59f922f4c0e1d8c6dcf974396f0f5cc59. This establishes metadata custody, not ZIP-byte equality.

Complete nontruncated subtree inventories account for all 7,060 changed paths: 6,112 expected products, 351 compiler/rust JSON fixture/oracle products, 328 production-export products, 268 locks and one Rust identity constant. The 45 distinct lock-blob pairs change only compiler semantics a15f5fbc→a1477c2e; lib.rs changes only the corresponding identity constant. Algorithm and evaluation-proof source are preserved. No generated source or certificate acceptance is inferred merely from the path inventory.

The body now distinguishes earlier be299 evidence from this delivered composition and retains the honest R3 gap. Current CI 43806849406 remains pending; bot CI/repro action_required are not passes. Existing owner retains final exact-head full acceptance, reproducibility and actual merged-tree equality. Reuse this completed producer and retained source reviews; no repeated generation is requested.

Receipt: /workspace/work/pr553-cdb94-product-adoption-20261116/receipt.json; SHA256 a02732efb34ccf66a818b881a6b80c663110928a173fefc07dcf087bd3410701. Complete inventory SHA256 528ffbb1b618b542b6313779fa16cffb57eec39a66c8178dc50d650bc493ab69.

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

New evidence at unchanged cdb94: reproducibility run 43806860849 attempt 2 completed successfully, with its one job and all seven steps passing, including just repro. Attempt 1 remains separately recorded as action_required. This resolves the exact-head reproducibility wait at job/step evidence level; no raw checkout-log or merged-tree claim is added.

Reuse the completed seven-job producer and reviewed output adoption. Existing final gate 43806849406 is still incomplete (50 successful jobs, seven running and two queued at inspection). The existing owner retains final-head full acceptance, the documented R3 disposition and actual merged-tree qualification. No producer or completed check needs repeating for this evidence update.

Receipt: /workspace/work/pr553-cdb94-repro-attempt2-20261116.json; SHA256 9e35565c9cab2f730d69ca4438a6e4dd113fa8c030c7792d6a87e80da9f921d8.

@afflom
afflom merged commit d07bd64 into release/v0.5 Nov 16, 2026
71 checks passed
@afflom
afflom deleted the claude/evaluation-proof-main branch November 16, 2026 10:03
afflom pushed a commit that referenced this pull request Nov 16, 2026
Authored conflicts: SM-43 (this change's decimal formatting) and SM-44
(the evaluation proof) are kept side by side in spec/src/IdsSM, the
feature suite, model/ids.toml and the README capability row; SPEC.md,
CONFORMANCE.md and ERRORS.md are projected by `just model-write`
(523 ids). The language-1.2 identity constant and every generated
product take release/v0.5's side; regenerate.yml relocks and rewrites
them for the composed compiler.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3PPDQti4rghmmPSZGy

@afflom afflom left a comment

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Actual landing d07bd64 has parents 7da6e93 + cdb94eb and tree ab19cd0e, equal to the reviewed delivered head. Final dispatch gate 43806849406 attempt 1 passed all 70 jobs and 813 steps, without skips, including honesty, Intel execution/inventory and Atlas join/native assertion. Its new plan log explicitly checks out cdb94. Completed producer/output custody and exact-head repro attempt 2 are reused without repetition.

Actual d07 reproducibility 43845282048 passed one job/seven steps. Actual push CI43845282185 is planner-only (one pass/17 skips), not another full gate. This establishes the stated final-composition and actual-tree qualification.

The original R3 chronology gap remains explicit and unwaived: implementation preceded the register/scenario/named test, and later plants do not establish original red-before-green. This review does not imply retrospective premerge approval, broad release acceptance or issue closure. Existing owner, forward-history and #468 obligations remain; all original failed/canceled evidence stays preserved.

Receipt: /workspace/work/pr553-d07-actual-final-review-20261116.json; SHA256 24947ac6818f2ce913c3f9775f2583ae54de82f06c9d5012fab7d8bcadb815ec.

afflom pushed a commit that referenced this pull request Nov 16, 2026
Generated kappa outputs and stdlib records take the release side for
regeneration; the register count is 524 (522 on the line plus KA-07/KA-08).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
afflom pushed a commit that referenced this pull request Nov 16, 2026
…solation

Generated outputs take the release side (ours where the release has none)
for regeneration. SM-43 and SM-44 are both registered: 525 IDs. The
compiler-semantics ID is the line's until regeneration relocks it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Asa3M8ULnzEQwLLngyBBX
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