You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Close v0.4 as a coherent release rather than a collection of merged features.
Work
Add canonical end-to-end examples for general formal programming, compiler-as-LexLean-program, globally certified production compilation, language model, reasoning model, and composed PrismPM-style application.
What I ran (Rust 1.97.1, no Lean installed, so no Lean-backed gate): cargo xtask validate-model (green: 414 ids, 85 codes, audit-deferral, audit-oracle-isolation, meta-gate), validate-spec-links (green: 414 rows bijective), check-test-inventory (green: 722 tests = 263 + 459), release-check (refuses; details below). Not run:just vv, verify-examples, check-golden, check-reproducibility, deny. Everything below that depends on them is "by reading", not "by running".
Legend: Done = discharged by something in the tree. Now = can be done without waiting. Blocked(#n) = depends on an open issue. Decision = needs the maintainer.
1. Work
#
Bullet
State
What discharges it / what is missing
W1a
E2E example: general formal programming
Done
examples/{semantic-1.1,language-1.2,recursive-data,higher-order,recursion,collections,fold-laws} via EX-08; negatives under tests/negative/
compiler/ (calculus, Gnaf, Certification) and bootstrap/ (BS-01, BS-02: 2 components, stage-0 oracle). Stage 8 of #92 and the closed Rust AST of #91 still open
GC-01..11, compiler/certificates/, examples/production*, PX-01..13. Final example waits on #91/#105 (#105: collection primitives realized by self-recursion but registered otherwise)
CONFORMANCE/ERRORS current (generated, validate-model). Pins consistent (see §4). Stale prose and identity questions in §3. The final pass waits on the other issues
W3
Package extraction/offline; supported-host policy
Partly Done
release-check built the packaged crate standalone with the identical four-line identity (lexlean 0.3.0 / language 1.0 / compiler-semantics 95deb33a… / lean-toolchain v4.32.1). Package is now 1054 files. Five-host policy is RP-14 + CI host matrix; not run here
W4
Repository audit
This comment
§3 below
W5
Legacy/oracle non-normative
Done
OR-01..08, audit-oracle-isolation green (5810 files in 7 roots name no part of the oracle), SPEC §17.19 and §17.23, bootstrap/README.md name the stage-0 Rust as the oracle. lean_text_legacy is live (certificate A/B/E paths) and documented at SPEC 4995, 5610, 5706, not dead
2. DoD and acceptance
Criterion
State
Evidence / gap
Every capability: one positive artifact + a negative fixture
release-check also reports, besides the expected 1.0.0 / missing release/ outputs, 8 schemas "validated by no conformance case". That is a gate false-negative, not missing coverage (finding F1).
3. Findings
ID
Where
What is stale or wrong
What is true now
Action
F1
crates/model/src/release.rs:541
schemas-exercised searches for "<name>"; the v0.4 cases use schema("<name>.schema.json"), so target-fixture, target-program, gnaf-request, gnaf-fixture, gnaf-certificate, gnaf-certificate-fixture, rust-package, rust-provenance are reported unexercised. They are used (cases/calculus.rs:663, cases/gnaf.rs:530, cases/certification.rs:273, …)
the criterion would refuse the 1.0.0 release for a reason that is false
blind spots: markers are only TODO FIXME XXX unimplemented! todo! "for now" "later version"; not scanned: README/SPEC/CHANGELOG/VERIFICATION/AGENTS.md (roots list omits repo-root .md), `.github/*.sh
*.json, .devcontainer/Dockerfile, api/, oracle/, tooling/`; "not yet", "later release", "follow in", "eventually", "in this release" are not markers. Today a scan with its own markers over the unscanned files finds only the two quoted lines in VERIFICATION 4540-4542
gate is green and latent-blind
F16
.devcontainer/ vs SPEC §9.3
SPEC: "post-create command MUST install the pinned Lean toolchain"; no postCreateCommand, the Dockerfile installs at image build. Versions are hard-coded (1.97.1, v4.32.1) rather than read from the pin files, and no gate or register row checks the devcontainer (only audit-deferral reads devcontainer.json)
pins match today, nothing keeps them matching
Decision (D3)
F17
README "Documented deviations"
SPEC↔code deviations are listed (public/ExistsUnique/probe/…); the devcontainer one (F16) and the lean_text_legacy module-system rendering are not
Now for the README list once D3 decided
No other staleness found in: capability ranges (every README range matches model/ids.toml, 33 prefixes), "414 IDs" (README:145), seven-method Engine (7 pub fn in impl Engine), maxRecDepth 100000/maxHeartbeats 1000000000 for language-1.1 (backend/semantic.rs:1885), Rust/Lean pins (all 301 lean-toolchain files = v4.32.1; CI reads the pin files; devcontainer 1.97.1/v4.32.1).
Orphaned v0.3 assumptions / dead paths: none found by grep (legacy|deprecated|obsolete|superseded hits are Lean-token lists, the live lean_text_legacy, and Atlas-register terms). The three #[allow(dead_code)] (api/production.rs:726,735, cli.rs:342) are serde schema fields under deny_unknown_fields, not dead production code. I did not run cargo +nightly udeps/dead-code analysis across the workspace, so "no unused path" is a negative result of reading, not a proof.
4. Decisions needed
D1 – Repository/image identity: keep afflom/lexlean in Cargo.toml, SPEC §2.3, schema $ids, Dockerfile label, README install lines, or move to UOR-Foundation? (Moving $ids changes compiler-semantics ID.)
D2 – Version/tag story: workspace is 0.3.0 with no tag, [0.5.0] Release 0.5.0 #128 is 0.5.0. Should SPEC §2.3 and CHANGELOG be rewritten at release time, and README's docker pull …:0.3.0 replaced with the 0.5.0 tag then?
D3 – Devcontainer: add a register row + check that .devcontainer/Dockerfile pins equal rust-toolchain.toml/lean-toolchain (R1-style), or list it as a documented deviation from §9.3?
5. Proposed PRs (draft, one topic each, nothing ready-marked)
claude/issue-34-release-check-schemas – F1.
claude/issue-34-readme-prose – F2, F3, F4, F5 (F17 after D3).
F6–F16 wait on the issues or decisions named above.
Contributor session on #34, base release/v0.4 at b8ff7bef6 (register 436). Nothing below has been run yet; this is what I will check, where, and at what cost. I start work only after the maintainer's go.
1. Audit (read, then fix in docs only)
One commit per finding class, on claude/issue-34-closure from release/v0.4, draft PR against release/v0.4.
Class
Files
Method
Stale prose / v0.3 assumptions
README.md (e.g. "the only tag so far is v0.1.0", the image owner, the "1.0 / initial implementation" framing against a 1.2 language), SPEC.md prose only, CHANGELOG.md (Unreleased vs. the work graph of #15), VERIFICATION.md opening sections
read; grep for v0.3, 0.3.0, "language 1.0", "future", "so far", version strings against Cargo.toml and lean-toolchain
Deferral markers (R4)
whole tree, incl. examples and fixtures
cargo xtask audit-deferral (inside validate-model) plus my own grep for the marker family; the gate's list is the oracle, mine is a cross-check
cargo xtask check-package; RP-16 pins; version still 0.x on the line, which is the maintainer's bump
Examples' READMEs
examples/*
none of the 20 example directories has a README; I will list which ones need a stated purpose, claim, and verify command and propose them (prose only)
2. Acceptance checklist (run and record)
Each is recorded on #34 with command, host, wall time, and result.
Item
Command
Expected cost
Clean-checkout gate
just vv
hours, not minutes: examples alone is the longest CI leg and UOR Atlas is sharded in CI; conformance_ex_09 is ~14 min by itself
Package / offline
cargo xtask check-package
minutes
Reproducibility
cargo xtask check-reproducibility
two full builds, tens of minutes
Golden
cargo xtask check-golden
minutes
Every example, real Lean + leanchecker + exact axiom policy
cargo xtask verify-examples (no --write)
per-example times reported; Atlas is the long one
Host policy §8.3
read ci.yml rows against the five hosts; cite the latest green run on release/v0.4 for the non-Linux-x86-64 hosts (I cannot run them)
minutes
Environment caveat, stated now. This container is Linux x86-64 with 4 cores, 15 GB RAM, Rust 1.97.1 pinned, and no just, elan/lean/leanchecker, or cargo-deny installed. Outbound HTTPS works through the proxy, so I will install the pinned tools (elan with leanprover/lean4:v4.32.1, just, cargo-deny) before the gates; I will not change any pin. The CI profile for examples is 2 wide with a 12 GiB budget on a 16 GB runner; here it is 4 cores / 15 GB, so wall times will not equal CI's and I will say so. If a gate is too long for this session's lifetime I will report which legs ran (using the same per-leg split as .github/vv-rows.json) and cite the CI run for the rest rather than present a partial run as the whole just vv.
3. Claim audit (README.md, SPEC.md)
Every sentence on global optimality, "model", "reasoning", "verified", "certified", "complete", "minimal", "oracle agreement", "self-hosting". For each I will list: the sentence, the evidence that bounds it (register ID and level, certificate, VERIFICATION.md plant), and the weakened wording I would use. The yardstick is the #15 rule: compiler verification, self-hosting and oracle agreement are not global optimality; internal plan optimality is not complete-system optimality; representation minimality is not execution optimality; measured / oracle results are not proofs (R2).
One row per v0.4 capability (from the #15 work graph and the register): positive end-to-end artifact (example, fixture or test), negative fixture, and gaps. Capabilities whose issues are still landing (#30 slice 4, #91, #92) will be marked "pending landing", not as gaps.
Will not touch
model/ids.toml and any register row, CONFORMANCE.md / ERRORS.md by hand (generated only, and only if prose I change feeds them), language/, schemas/, compiler semantics or any identity-moving file, gate code, the version, tags, main, other branches. Nothing for #91 stage 4, #30 slice 4 (#247), or #92. Any finding that would need a gate-affecting change goes to #34 as a finding, with the R3 order it would need, and not into the PR.
Reporting
Maintainer messaged at: plan posted (now); draft PR opened; clean-checkout just vv result; claim audit done. Findings tables are posted to #34 as they are produced.
Tools. Install exactly as ci.yml does: elan-init with --default-toolchain none, then elan toolchain install "$(cat lean-toolchain)", and cargo install --locked just cargo-deny; no pin changes. The clean-checkout just vv criterion is discharged on the normative host by the gate on the release candidate, so your run is evidence for the audit, not the discharge: run it by the .github/vv-rows.json legs that fit the container, record each leg's host, wall time and result on this issue, and cite the latest green run on release/v0.4 for the legs that do not fit. Say which is which.
Example READMEs. First the list on this issue (per example: purpose, the register IDs it exercises, the verify command), then the READMEs in the draft PR once the list is here. Each README's claim is bound to the rows it names and says nothing stronger; a README is hand-written prose and the claim audit applies to it. Keep them short.
Findings that need a gate-affecting change go here as findings with the R3 order they would need, as you planned; nothing of that kind in the PR.
Example READMEs: the list (purpose, register IDs, verify command) — corrected
Base release/v0.4 @ b8ff7bef6. This edit replaces my first version of this comment, whose ID column came from a scan of the case sources and was wrong in places: it listed loop-type rows (AU-09..AU-12, CF-18, RP-10, RT-03, RT-04, PK-05, RP-14) that iterate over many examples and say nothing about one. I discarded those. Also, examples/ has 17 directories, not the 20 of my plan (a miscount).
Method. A README names a row only if the row's register statement names the example, or the row's statement is plainly that example's capability and its case arm builds the example (copy_example/examples/…); I read each statement. Purposes are paraphrases of the root README's own example table, which is the existing source for them. Every example is also under EX-08.
Verify command (all).cargo xtask verify-examples (the gate; there is no per-example filter, only LEXLEAN_EXAMPLES_PART=without:uor-atlas and the Atlas shard/join parts). By hand, one example: cd examples/<name> && lexlean check && lexlean verify (pinned Lean, leanchecker replay, exact axiom audit; it does not do the expected/ comparison).
Seven examples (collections, higher-order, language-1.2, production, production-coverage, recursion, recursive-data) carry \title{Natural number addition} in Main.lex.tex, a copy of the literal example's title. The title is part of the canonical document, so changing it moves expected/ records; I will not touch it. Gate-affecting finding (no register row; a sweep of expected/ with check-golden --write and verify-examples --write), for the maintainer to decide.
list-induction, propositional-logic, semantic-1.1, language-1.2 and peano-arithmetic are named by no capability row; they stand on EX-08 (and VR-23 for Peano). Their READMEs say exactly that. The capability table will list them as gaps in negative coverage unless a row is wanted.
No example contains a GNAF request or an optimality certificate (grep -ril 'gnaf\|optimal' examples finds none), which is what the production-family READMEs state.
READMEs are written in the draft PR (claude/issue-34-closure). Update:nat-add-zero has none: SE-07 holds that project's root to a closed file list (see the gate record and the capability table), so #250 has 16 READMEs.
Gate record 1: the cheap recipes, run here (evidence for the audit, not the discharge)
Host. Linux x86-64 container, 4 cores, 15 GB RAM, no swap; Rust 1.97.1 (pin), Lean 4.32.1 via elan toolchain install "$(cat lean-toolchain)", just 1.59.0 and cargo-deny 0.20.2 by cargo install --locked. No pin changed. Head: claude/issue-34-closure @ f71ae0e44 (= release/v0.4 @ b8ff7bef6 + the 17 example READMEs, #250). Cold build, so the first recipe of each kind includes compiling.
Recipe (just vv order)
Result
Wall time
fmt-check
pass
4 s
model (cargo xtask validate-model)
pass: 436 ids, 86 codes, all audits clean
180 s cold build + run
spec-links
pass: 436 rows bijective
<1 s
lint (clippy -D warnings)
pass
70 s
features
pass
50 s
inventory
pass
107 s
golden (check-golden)
pass
81 s
deny
pass: advisories, bans, licenses, sources ok
1 s
cargo xtask check-package
pass: packaged crate builds standalone with the four-line identity (language 1.0, compiler-semantics 8554008034f6…, lean-toolchain leanprover/lean4:v4.32.1)
67 s
A defect the gate caught in this PR's first commit.validate-model was red on the first READMEs: audit-oracle-isolation (R2, §17.19) refuses any mention of the differential oracle under examples/; I had named it in the production-family READMEs. Reworded to cite the OR-* rows; green on f71ae0e44. Not a plant, but it shows the isolation audit fires on prose in examples/.
Running now, in vv-rows order, one at a time (the container cannot run legs side by side):test; the examples leg (without:uor-atlas, width 2, 12 GiB budget, the committed hints); repro; then each bdd part alone, preservation last and by itself (the maintainer measured ~10.4 GB RSS for it). UOR Atlas: its CI shards need the parts of one run and a 12 GiB budget; if it does not fit I will cite the CI run rather than report it as run. I will say per leg which is run here and which is cited.
Gate record 2: examples leg and test, run here (evidence for the audit, not the discharge)
Same host and head as record 1 (Linux x86-64, 4 cores, 15 GB, no swap; claude/issue-34-closure @ 48eb261ea, which differs from b8ff7bef6 only by README prose).
Recipe
Result
Wall time
test (workspace unit tests, repo-conformance excluded)
pass
391 s
examples leg = LEXLEAN_EXAMPLES_PART=without:uor-atlas, width 2, 12 GiB budget, committed verify-hints.json (the CI profile)
pass, rc=0
4388 s (73 min)
What the examples leg ran. 21 projects, each formatted, locked, checked, built and verified with real Lean 4.32.1, leanchecker replay and the exact axiom policy, and each project's normalized verification records compared with expected/verify (§29.5): the 16 examples other than uor-atlas, plus compiler, stdlib, bootstrap, std-models, std-reasoning (the gate treats those as examples, §28.6). Per project (verify-examples: <name> took N s):
Project
s
Project
s
production-coverage
1093
fold-laws
29
bootstrap
905
std-reasoning
26
reasoning
438
recursion
24
stdlib
400
collections
24
models
400
std-models
20
production
358
semantic-1.1
9
helpdesk
322
peano-arithmetic
6
compiler
298
recursive-data, propositional-logic, higher-order
3 each
nat-add-zero, list-induction, language-1.2
2 each
Times are for this 4-core box at the CI profile and are not CI's. The projects ran one after another here; in CI the leg is one job.
Not run here: uor-atlas. Its CI verification is sharded (.github/verify-shards.json) over several jobs of one run and joined; I did not reproduce the shard/join run. I will cite the latest green release/v0.4 run for it (see the final record). It is not reported as verified by this session.
Still to run, in this order:repro, then the bdd parts alone (bootstrap-check, bootstrap-validate, compiler, production, verification, preservation-lists, preservation, rest).
Gate record 3: repro and the bdd parts, run here, with the CI run that discharges the rest
Same container as records 1-2 (Linux x86-64, 4 cores, 15 GB, no swap). Head: claude/issue-34-closure @ d5f6c405d (see the README note under verification). The parts ran one at a time; preservation ran alone. Per .github/vv-rows.json.
Leg / recipe
Result here
Wall time here
CI, same tree (run 39520454225)
repro (check-reproducibility)
pass
342 s
in rest: 1921 s total for the leg
bdd part bootstrap-check
pass
3397 s
2442 s
bdd part bootstrap-validate
pass
2069 s
1600 s
bdd part compiler
pass
2793 s
2831 s
bdd part production
pass
4573 s
1727 s
bdd part preservation-lists
pass
849 s
985 s
bdd part rest (library tests, doctests, meta-gate, remaining IDs)
pass
2087 s
in rest leg
bdd part verification
FAIL, 1 test: conformance_se_07 (90 passed)
2168 s
1356 s, green
bdd part preservation
FAIL, 1 test: conformance_sp_02 (13 passed)
3473 s
3790 s, green
The two failures, and what they are.
conformance_se_07 was my change: examples/nat-add-zero/README.md is a file the project root "gains", and SE-07 (§25.6, confined staging) holds the literal example's root to a closed set (.lexlean, NatAddZeroHost.lean, lake-manifest.json, lakefile.toml, lean-toolchain, lexlean.lock, lexlean.toml, src). I removed that one README (d5f6c405d, [v0.4] #34 closure: example READMEs and README claim wording #250). conformance_se_07 alone then passes (0.72 s). No other case failed on a README: the 16 other example READMEs are in the tree for every part above. The whole verification part is being re-run on d5f6c405d and will be recorded; until then this row stays FAIL for the part as run. Consequence for the capability table: the literal example cannot carry a README without editing the allow-list in crates/conformance/src/cases/security.rs:849, which is a gate change and so belongs here, not in the PR.
conformance_sp_02 failed with LLS8002: "child_timeout_ms exceeded by lean in phase LexLeanPreservation.RustLemmas: configured 300000, observed 300008 ms". That is a resource limit on this box, not a failed proof or a semantic mismatch: 8 ms over the project limit, on the compile of one Lean module, on 4 cores with no swap. In CI the same part is green in 3790 s. I am not counting it as run-and-passed. It is being re-run alone; the result goes in the next record. Finding for the maintainer:child_timeout_ms = 300000 leaves this module essentially no headroom on slower hardware, and the failure mode (a resource limit reported as a gate failure on a conformance case) is exactly what R2/R9 intend; I changed nothing about it.
What discharges the criterion. The criterion is the gate on the release candidate. For the legs this box cannot discharge, the evidence is CI run 39520454225 (pull_request, head 53c7a8f60, green, 28 jobs): b8ff7bef6 and 53c7a8f60 have the same tree (2a203993ac09… both), so it verified the tree of this line, and the later push run 39529604597 only recorded that. Its jobs, all success: plan; just vv for examples (2261 s), rest (1921 s), bdd-bootstrap-check, bdd-bootstrap-validate, bdd-compiler, bdd-preservation, bdd-preservation-lists, bdd-production, bdd-verification; all five hosts of §8.3 (x86_64-apple-darwin, aarch64-apple-darwin, aarch64-unknown-linux-gnu, x86_64-pc-windows-msvc as "build and non-Lean tests", Linux x86-64 in the just vv legs); and the UOR Atlas jobs (prefix, cert4, cert4-replay, census-a..d, pairs-a..d, orbit, closure, join), which I did not run here and cite only.
§8.3 host policy
Evidence
Linux x86-64: normative full gate
CI just vv legs (run 39520454225); this session's legs above
Linux AArch64, macOS x86-64, macOS AArch64, Windows x86-64: build + non-Lean tests
CI jobs build and non-Lean tests (…), all success in the same run; not run here
Which is which.Run here and passed:fmt-check, model, spec-links, lint, test, features, inventory, golden, deny, check-package, repro, examples (21 projects), bdd parts bootstrap-check, bootstrap-validate, compiler, production, preservation-lists, rest. Run here and failed:bddverification (fixed, re-run pending), bddpreservation (resource limit, re-run pending). Cited, not run:uor-atlas shards and join, the four other hosts, and the CI-measured times above.
Base release/v0.4 @ b8ff7bef6, register 436. Rows follow the work graph of #15, plus #91/#92 (still landing). Method and limit. Positive = a committed project that verify-examples (or the project's own gate) verifies with real Lean; negative = a committed directory under tests/negative/ (285 of them), a committed negative certificate under compiler/certificates/, or an inline defect a conformance case plants and the diagnostic code it asserts. I read the directory names and the case arms by hand; the automatic scan I tried could not tell inline negatives from helper-function ones, so "inline" below means I found the asserting arm, not that I audited each plant. Planted-defect records for gates are in VERIFICATION.md (§27.9) and were not re-derived.
the same directory holds the negatives: undercut, unattained, non-strict, premise-omitted/weakened, bridge-missing/violated/out-of-range, claim-mismatch, …
none; no examples/ project carries a GNAF certificate (grep), so no example makes an optimality claim
refuses by design below 1.0.0 (README, CHANGELOG preface)
the 1.0.0 naming vs. a v0.4 line is the maintainer's at the version bump
Summary of gaps. Four v0.4 capabilities have no committed negative fixture directory, only inline planted defects: TC (#21), GN (#22), RB (#24) and LM (#30). Whether the DoD's "relevant negative fixture" accepts an inline plant asserting a registered code is the maintainer's call; if it does not, each needs a directory under tests/negative/ (R3: scenario and test already exist for these rows, so only fixtures plus the EX-07 coverage list would move). Pending, not gaps: #30 slice 4, #91 (RB-15/16), #92 steps 5-7.
Related finding (gate-affecting, for you).examples/nat-add-zero cannot carry a README: SE-07 (crates/conformance/src/cases/security.rs:849) holds the literal example's root to a closed file list, and the conformance cases copy that project (P::example()). #250 therefore has 16 example READMEs, not 17. R3 order if you want one: no new row; edit the allow-list in the SE-07 arm, plant a stray root file and record the failing signature in VERIFICATION's falsifiability table.
Maintainer, rehearsal result: run 40221304393 (on #318's second head, 1d6acf575) finished with eight jobs green (tag identity, crate readiness, the five host binaries) and one red: its release gate's just vv failed in conformance_rp_18 only, the RP-18 oracle reading the old crate.yml lines, which #318's third commit fixed; the tag commit 93fbe7df9 carries that fix and its own gate passed rp_18. Nothing else in the release workflow failed. To have the whole workflow proved at the exact tag commit, one more rehearsal is dispatched on release/v0.4 at 93fbe7df9; it publishes nothing (sha-93fbe7d image tag only) and does not hold the tag: the tag can be pushed at any time, and if its run ever fails in a job, that job is re-run without re-tagging.
Hold the tag: do not push v0.4.0 at 93fbe7df9 yet.
Rehearsal 3 showed the release workflow's just vv (release evidence) job cannot finish: on one hosted runner the bdd recipe alone took 4 h 14 min (run 40221304393), and rehearsal 4 (run 40244034460, at the tag commit) is about to hit the six-hour job limit in the same job. A tag pushed on 93fbe7df9 would fail there, and image, manifest, and publish all wait for it.
The fix is in progress on release/v0.4: the release gate reads the ci.yml run of the tagged commit that was the whole gate (the sharded gate that already runs on every push to the line and passed on 93fbe7df9, run 40233071771) and records it as the §30.3 evidence, instead of re-running the gate on one runner. SPEC §9.4/§30.3 and the RP-18 oracle change with it. When it lands, the tag commit moves to the new head of release/v0.4; I will post the exact commands here again, and rehearse once more first.
The v0.4.0 tag commit is aa6593e026d92f4696fe970715c81528ad2c37da on release/v0.4: the merge of #321 (release notes for the cut that carries H5 and the release-workflow fixes), after #320 (the release gate reads the gate run of the tagged commit; rehearsal 5, run 40292173816, was the first fully green rehearsal: tag identity, release evidence in under a minute, the five host binaries, crate readiness, both images and the multi-architecture tag). The register is 457 and the three compiler-semantics IDs are the ones the locks pin after H5 (CHANGELOG 0.4.0).
Rehearsal 6 is running on this commit now; I will post one line here when it is green. Then the tag, which only you can push:
git fetch origin release/v0.4
git tag -a v0.4.0 aa6593e026d92f4696fe970715c81528ad2c37da -m "LexLean v0.4.0"
git push origin v0.4.0
Push the tag only; release.yml's publish job creates the GitHub Release (binaries, image, crate, SBOM, checksums, gate evidence) and uploads to crates.io only if CARGO_REGISTRY_TOKEN or CRATES_IO_TRUSTED_PUBLISHING exists, saying so otherwise. Landings on release/v0.4 are held until the tag exists, so the commit does not move again; main receives the line through a PR after the tag.
Rehearsal 6 on the tag commit aa6593e026d92f4696fe970715c81528ad2c37da is green (run 40304904506: every job passed, the two tag-only publish steps skipped as designed). The tag is ready to push with the commands above.
The v0.4.0 tag was pushed at 12:17Z on aa6593e026d92f4696fe970715c81528ad2c37da with the commands above (annotated tag object 699dd53a96524b62a7320b2e0cf96048b14de832). The tag run, 40321415513, failed at its last job, publish the release, step Read the release criterion over the assembled set.
Every other job passed: tag identity, release evidence, the five host binaries, crate readiness, both images and the multi-architecture tag (the development pre-release upload was skipped as designed).
Failing criterion.cargo xtask release-check refused with source-tag and version-output (the two the step forgives under §2.3) and with one it does not:
ci-evidence: release/vv-evidence.txt does not record the gate passing
##[error]criteria unmet for reasons §2.3 does not explain: ci-evidence
Cause.crates/model/src/release.rs:453 requires the literal vv: the acceptance gate passed in release/vv-evidence.txt. Since #320, .github/release-gate.sh writes that file from the sharded ci.yml run of the tagged tree (here run 40292537147 on d985f9561, the second parent, every job passed) plus the legs' vv.txt transcripts. None of the 29 leg transcripts prints that line; only the single-runner just vv does. #320 changed release-gate.sh, ci.yml, release.yml, SPEC, VERIFICATION and RP-18, not release.rs. The rehearsals skip publish the release, so this criterion first ran on the tag.
What was published.ghcr.io/uor-foundation/lexlean:0.4.0 (with 0.4.0-amd64, 0.4.0-arm64) and :latest.
What was not. No GitHub Release, so none of the binaries, SBOM, checksums or gate evidence, and no crates.io upload: the step stopped before Decide how the upload is authorised. The run reports HAS_CARGO_REGISTRY_TOKEN: true, so the next run that passes this step uploads the crate.
The tag is untouched. The fix and whether the tag moves are yours.
Thanks for the diagnosis; it is exact. The fix is #323 (release-gate.sh ends the evidence with the line release-check reads; the criterion is one function, repo_model::release::ci_evidence, with its own test, and RP-18 holds the script's output to it, so the publish job can no longer refuse what the rehearsals accepted). Its gate is running.
The tag moves. The publish job runs the workflow of the tagged commit and downloads the evidence its own gate job wrote, so a re-run of the failed job would refuse the same file; aa6593e02 cannot be released as tagged. When #323 is merged and rehearsal 7 on its merge commit is green, I post the new commit here, and the tag is re-made from it:
Nothing was published under the first tag (no GitHub Release, no crates.io upload), so the move loses nothing; the image tags 0.4.0 and latest are overwritten by the next run. Landings on release/v0.4 stay held until the tag exists.
#323 is merged and rehearsal 7 on its merge commit is green (run 40339000887: every job passed, the two tag-only publish steps skipped as designed; the evidence job named gate run 40325802734 and the file ends with the line release-check reads, checked by RP-18 and the criterion's own test).
The v0.4.0 tag commit is e0b74b81e6f49c83596a957d232c8383a5db1a25. Move the tag:
git push origin :refs/tags/v0.4.0
git fetch origin release/v0.4
git tag -a v0.4.0 e0b74b81e6f49c83596a957d232c8383a5db1a25 -m "LexLean v0.4.0"
git push origin v0.4.0
Nothing was published under the first tag, so nothing is lost. I watch the tag run through publish the release and post the result here.
Next: release/v0.4 is merged into main by PR (the line at its tag, SPEC §9.4), the hold on release/v0.4 landings is lifted (patch work on the line continues as 0.4.x; feature work goes to release/v0.5), and the second forward merge into release/v0.5 follows.
Closing: v0.4.0 is released from e0b74b81e (release run 40382770933, every job green including publish; GitHub release with 15 assets; crates.io lexlean 0.4.0; main at the tag via #324).
Against the acceptance criteria:
just vv from a clean checkout on the normative host: the release's vv-evidence.txt is the gate run of the tagged commit's tree (every job, every just vv leg, Atlas join), checked by release-check at publish.
Package/offline verification, every example with real Lean, leanchecker replay and the exact axiom policy, reproducibility and golden checks: all jobs of that run and of the release run (release-check.txt, MANIFEST.sha256, release-identity.json, SBOM, checksums).
Remaining v0.4-milestone work (#91 heap half, #92 step 4, #30 slice 4) lands on release/v0.4 as 0.4.x patch releases per the owner's rule on #92, not as a condition of this closure. #256 (#251 retitles, #252 fixtures) stays queued for 0.4.1. Umbrella #15 carries the release note.
The tag was moved to the commit the maintainer named, and the tag run went through publish the release.
Tag.v0.4.0 now points at commit e0b74b81e6f49c83596a957d232c8383a5db1a25 through annotated tag object 22698a589f9f4d91b65c498f9de057c98fc4a4dc. The old tag object 699dd53a96524b62a7320b2e0cf96048b14de832 (commit aa6593e0) is orphaned; nothing had been published under it.
Run.40382770933 (push, tag v0.4.0, commit e0b74b81) completed success. Every job passed — tag identity, release evidence, the five host binaries, crate readiness, both images, multi-architecture tag, and publish the release; the development pre-release upload was skipped as designed.
Release.v0.4.0, published 2026-10-26T19:20:36Z, 15 assets: lexlean-0.4.0-binaries.tar.gz, lexlean.crate, sbom.json, vv-evidence.txt (194,960 bytes), release-check.txt, release-identity.json, version-output.txt, compiler-semantics-id.txt, MANIFEST.sha256, checksums.txt, SPEC.md, CONFORMANCE.md, ERRORS.md, LICENSE-APACHE, LICENSE-MIT. The image is ghcr.io/uor-foundation/lexlean:0.4.0@sha256:d53355cb3c678ad342e77502d3a43e450a0fe2064315d8d3ecbea5d231547c08.
The criterion. The shipped release-check.txt names only the two criteria §2.3 defers at this version:
gate failed: RP-12: the release is refused; unmet criteria:
source-tag: the workspace version is not 1.0.0; §2.3 fixes the first complete release at 1.0.0
source-tag: CHANGELOG.md has no `## 1.0.0` entry
version-output: release/version-output.txt must be exactly: ... lexlean 1.0.0 ...
ci-evidence is no longer among them, so #323's fix (release-gate.sh ending the evidence with the line release-check reads) is confirmed on the publish path, not only in rehearsal. version-output.txt reads lexlean 0.4.0 / language 1.0 / compiler-semantics 07859acb5a756221018b7f65e7f725d99b00a694670eecc7cd264832ba1333fb / lean-toolchain leanprover/lean4:v4.32.1.
Open points. Whether the crate also reached crates.io is not stated in the release — RP-18 permits a release without the upload. The release object records target_commitish: main even though the tag points at e0b74b81, which is on release/v0.4. The forward merge of release/v0.4 into main is next.
crates.io: lexlean 0.4.0 is on crates.io, created 2026-10-26T19:20:26Z (the registry's /api/v1/crates/lexlean/versions), so the upload step ran, not only the release without it that RP-18 permits.
target_commitish: main: cosmetic. A release object made from an existing tag resolves to the tag's commit (e0b74b81e) regardless of that field, and main has since received the line at the tag through main receives the v0.4.0 line: release/v0.4 at e0b74b81e (the tag) #324 (merge b9793961c), so the field now also names a branch that contains the commit.
Follow-up to the line above — the crates.io question is settled: the crate was published.
crates.io/api/v1/crates/lexlean lists one version, 0.4.0, not yanked:
num 0.4.0
created_at 2026-10-26T19:20:26.786788Z (10 s before the GitHub Release, 19:20:36Z)
checksum f2d579289052493d2c68274b10015d2cce544a78fbb876d73062184f09fb0a18
crate_size 5260039 (release asset lexlean.crate: 5260039)
edition 2021
rust_version 1.97
license MIT OR Apache-2.0
published_by afflom
That checksum is the same string as the lexlean.crate line of the release's checksums.txt, so the crate on crates.io is byte-for-byte the crate the release attaches — the equality step 12 of publish the release asserts (the crate cargo would upload is the crate this release attaches).
13 success Decide how the upload is authorised
14 skipped Acquire crates.io trusted-publisher token
15 skipped No upload credential
16 success Publish the exact verified crate to crates.io
17 success Publish
and the log reads Uploading lexlean v0.4.0 → Uploaded lexlean v0.4.0 to registry crates-io → Published lexlean v0.4.0 at registry crates-io, with HAS_CARGO_REGISTRY_TOKEN: true. So the first upload took the crates.io environment secret; trusted publishing (#260, RP-18) was not exercised, its two steps correctly skipped.
Both open points from my earlier comment now stand as: this one settled, and the release object's target_commitish: main — which is moot for consumers, since the tag itself names e0b74b81.
Activity
afflom commentedon Oct 15, 2026
Recorded as item L2.5 of the completion plan #106. The release itself is #128 (0.5.0).
Generated by Claude Code
afflom commentedon Oct 18, 2026
Audit of #34 against
main@724a55590What I ran (Rust 1.97.1, no Lean installed, so no Lean-backed gate):
cargo xtask validate-model(green: 414 ids, 85 codes, audit-deferral, audit-oracle-isolation, meta-gate),validate-spec-links(green: 414 rows bijective),check-test-inventory(green: 722 tests = 263 + 459),release-check(refuses; details below). Not run:just vv,verify-examples,check-golden,check-reproducibility,deny. Everything below that depends on them is "by reading", not "by running".Legend: Done = discharged by something in the tree. Now = can be done without waiting. Blocked(#n) = depends on an open issue. Decision = needs the maintainer.
1. Work
examples/{semantic-1.1,language-1.2,recursive-data,higher-order,recursion,collections,fold-laws}viaEX-08; negatives undertests/negative/compiler/(calculus,Gnaf,Certification) andbootstrap/(BS-01,BS-02: 2 components, stage-0 oracle). Stage 8 of #92 and the closed Rust AST of #91 still openGC-01..11,compiler/certificates/,examples/production*,PX-01..13. Final example waits on #91/#105 (#105: collection primitives realized by self-recursion but registered otherwise)LM-01,LM-02,LM-03,LM-13(tokenizers).LM-04..12do not exist;examples/modelshas a neural digit recognizer, no text pipelineexamples/reasoning,RS-01..14crates/conformance/src/production_consumer.rs(thePXconsumer), not as an applicationvalidate-model). Pins consistent (see §4). Stale prose and identity questions in §3. The final pass waits on the other issuesrelease-checkbuilt the packaged crate standalone with the identical four-line identity (lexlean 0.3.0 / language 1.0 / compiler-semantics 95deb33a… / lean-toolchain v4.32.1). Package is now 1054 files. Five-host policy isRP-14+ CI host matrix; not run hereOR-01..08,audit-oracle-isolationgreen (5810 files in 7 roots name no part of the oracle), SPEC §17.19 and §17.23,bootstrap/README.mdname the stage-0 Rust as the oracle.lean_text_legacyis live (certificate A/B/E paths) and documented at SPEC 4995, 5610, 5706, not dead2. DoD and acceptance
release-artifacts,AR-13,repro,golden. Not run herebuild, noneopen.LM-04..12are not yet rows (#30)just vvon the normative host (maintainer; ~700 runner-min)just vvfrom clean checkoutcheck-packagelogic runs insiderelease-check; green part aboveleanchecker, exact axiomsverify-examplesneeds Lean 4.32.1release-checkalso reports, besides the expected1.0.0/ missingrelease/outputs, 8 schemas "validated by no conformance case". That is a gate false-negative, not missing coverage (finding F1).3. Findings
crates/model/src/release.rs:541schemas-exercisedsearches for"<name>"; the v0.4 cases useschema("<name>.schema.json"), sotarget-fixture,target-program,gnaf-request,gnaf-fixture,gnaf-certificate,gnaf-certificate-fixture,rust-package,rust-provenanceare reported unexercised. They are used (cases/calculus.rs:663,cases/gnaf.rs:530,cases/certification.rs:273, …)PK-01..PK-04PK-05(README:187 already saysPK-01..PK-05)rust-toolchain.tomlpins 1.97.1compiler/,bootstrap/,stdlib/,oracle/,tooling/,api/,.github/LMLM-04..landdocker pull ghcr.io/afflom/lexlean:0.3.0v0.1.0;release.yml:68publishesghcr.io/<owner>/lexlean, i.e.uor-foundationnow; no 0.3.0 image is evidencedCargo.toml, schema$ids, Dockerfile label, READMErepository/homepage=github.com/afflom/lexlean; version0.3.0is a MUST in §2.3UOR-Foundation/LexLean; the next release is 0.5.0 (#128)$ids are inside the compiler-semantics digest, so changing them moves identity: I will not touch them0.3.0has a CHANGELOG section but no tag (onlyv0.1.0)0.1.1,0.2.0,0.3.0sections exist with no tagslexlean-0.3.0.crate, 712 files, digestsaudit-deferral.lex.tex(the self-hosting plan of issue 92)"xtask/src/audit.rs:858-audit_deferralTODO FIXME XXX unimplemented! todo! "for now" "later version"; not scanned:README/SPEC/CHANGELOG/VERIFICATION/AGENTS.md(roots list omits repo-root.md), `.github/*.sh,.devcontainer/Dockerfile,api/,oracle/,tooling/`; "not yet", "later release", "follow in", "eventually", "in this release" are not markers. Today a scan with its own markers over the unscanned files finds only the two quoted lines in VERIFICATION 4540-4542.devcontainer/vs SPEC §9.3postCreateCommand, the Dockerfile installs at image build. Versions are hard-coded (1.97.1, v4.32.1) rather than read from the pin files, and no gate or register row checks the devcontainer (onlyaudit-deferralreadsdevcontainer.json)lean_text_legacymodule-system rendering are notNo other staleness found in: capability ranges (every README range matches
model/ids.toml, 33 prefixes), "414 IDs" (README:145), seven-methodEngine(7pub fninimpl Engine),maxRecDepth 100000/maxHeartbeats 1000000000for language-1.1 (backend/semantic.rs:1885), Rust/Lean pins (all 301lean-toolchainfiles = v4.32.1; CI reads the pin files; devcontainer 1.97.1/v4.32.1).Orphaned v0.3 assumptions / dead paths: none found by grep (
legacy|deprecated|obsolete|supersededhits are Lean-token lists, the livelean_text_legacy, and Atlas-register terms). The three#[allow(dead_code)](api/production.rs:726,735,cli.rs:342) are serde schema fields underdeny_unknown_fields, not dead production code. I did not runcargo +nightly udeps/dead-code analysis across the workspace, so "no unused path" is a negative result of reading, not a proof.4. Decisions needed
afflom/lexleaninCargo.toml, SPEC §2.3, schema$ids, Dockerfile label, README install lines, or move toUOR-Foundation? (Moving$ids changes compiler-semantics ID.)0.3.0with no tag, [0.5.0] Release 0.5.0 #128 is 0.5.0. Should SPEC §2.3 and CHANGELOG be rewritten at release time, and README'sdocker pull …:0.3.0replaced with the 0.5.0 tag then?.devcontainer/Dockerfilepins equalrust-toolchain.toml/lean-toolchain(R1-style), or list it as a documented deviation from §9.3?5. Proposed PRs (draft, one topic each, nothing ready-marked)
claude/issue-34-release-check-schemas– F1.claude/issue-34-readme-prose– F2, F3, F4, F5 (F17 after D3).F6–F16 wait on the issues or decisions named above.
Generated by Claude Code
Wave 4: SL-02 (SHA-256), BS-03 (primitive scanner), PD-08 (stack boun…
afflom commentedon Oct 21, 2026
v0.4 closure: audit plan (intent)
Contributor session on #34, base
release/v0.4atb8ff7bef6(register 436). Nothing below has been run yet; this is what I will check, where, and at what cost. I start work only after the maintainer's go.1. Audit (read, then fix in docs only)
One commit per finding class, on
claude/issue-34-closurefromrelease/v0.4, draft PR againstrelease/v0.4.README.md(e.g. "the only tag so far isv0.1.0", the image owner, the "1.0 / initial implementation" framing against a 1.2 language),SPEC.mdprose only,CHANGELOG.md(Unreleased vs. the work graph of #15),VERIFICATION.mdopening sectionsv0.3,0.3.0, "language 1.0", "future", "so far", version strings againstCargo.tomlandlean-toolchaincargo xtask audit-deferral(insidevalidate-model) plus my own grep for the marker family; the gate's list is the oracle, mine is a cross-checkcrates/lexlean,compiler/,bootstrap/,oracle/,tooling/,xtask/oracle/*.toml,lean4-prodmentions in README, SPEC, CHANGELOG, VERIFICATIONcrates/lexlean/Cargo.toml, workspaceCargo.toml,.devcontainer/,Dockerfile,lean-toolchain,rust-toolchain.toml,deny.tomlcargo xtask check-package;RP-16pins; version still 0.x on the line, which is the maintainer's bumpexamples/*verifycommand and propose them (prose only)2. Acceptance checklist (run and record)
Each is recorded on #34 with command, host, wall time, and result.
just vvexamplesalone is the longest CI leg and UOR Atlas is sharded in CI;conformance_ex_09is ~14 min by itselfcargo xtask check-packagecargo xtask check-reproducibilitycargo xtask check-goldenleanchecker+ exact axiom policycargo xtask verify-examples(no--write)ci.ymlrows against the five hosts; cite the latest green run onrelease/v0.4for the non-Linux-x86-64 hosts (I cannot run them)Environment caveat, stated now. This container is Linux x86-64 with 4 cores, 15 GB RAM, Rust 1.97.1 pinned, and no
just,elan/lean/leanchecker, orcargo-denyinstalled. Outbound HTTPS works through the proxy, so I will install the pinned tools (elanwithleanprover/lean4:v4.32.1,just,cargo-deny) before the gates; I will not change any pin. The CI profile forexamplesis 2 wide with a 12 GiB budget on a 16 GB runner; here it is 4 cores / 15 GB, so wall times will not equal CI's and I will say so. If a gate is too long for this session's lifetime I will report which legs ran (using the same per-leg split as.github/vv-rows.json) and cite the CI run for the rest rather than present a partial run as the wholejust vv.3. Claim audit (README.md, SPEC.md)
Every sentence on global optimality, "model", "reasoning", "verified", "certified", "complete", "minimal", "oracle agreement", "self-hosting". For each I will list: the sentence, the evidence that bounds it (register ID and level, certificate,
VERIFICATION.mdplant), and the weakened wording I would use. The yardstick is the #15 rule: compiler verification, self-hosting and oracle agreement are not global optimality; internal plan optimality is not complete-system optimality; representation minimality is not execution optimality; measured / oracle results are not proofs (R2).4. Capability table on #34
One row per v0.4 capability (from the #15 work graph and the register): positive end-to-end artifact (example, fixture or test), negative fixture, and gaps. Capabilities whose issues are still landing (#30 slice 4, #91, #92) will be marked "pending landing", not as gaps.
Will not touch
model/ids.tomland any register row,CONFORMANCE.md/ERRORS.mdby hand (generated only, and only if prose I change feeds them),language/,schemas/, compiler semantics or any identity-moving file, gate code, the version, tags,main, other branches. Nothing for #91 stage 4, #30 slice 4 (#247), or #92. Any finding that would need a gate-affecting change goes to #34 as a finding, with the R3 order it would need, and not into the PR.Reporting
Maintainer messaged at: plan posted (now); draft PR opened; clean-checkout
just vvresult; claim audit done. Findings tables are posted to #34 as they are produced.Generated by Claude Code
afflom commentedon Oct 21, 2026
Go, as planned, with three notes.
ci.ymldoes:elan-initwith--default-toolchain none, thenelan toolchain install "$(cat lean-toolchain)", andcargo install --locked just cargo-deny; no pin changes. The clean-checkoutjust vvcriterion is discharged on the normative host by the gate on the release candidate, so your run is evidence for the audit, not the discharge: run it by the.github/vv-rows.jsonlegs that fit the container, record each leg's host, wall time and result on this issue, and cite the latest green run onrelease/v0.4for the legs that do not fit. Say which is which.Generated by Claude Code
afflom commentedon Oct 21, 2026
Example READMEs: the list (purpose, register IDs, verify command) — corrected
Base
release/v0.4@b8ff7bef6. This edit replaces my first version of this comment, whose ID column came from a scan of the case sources and was wrong in places: it listed loop-type rows (AU-09..AU-12,CF-18,RP-10,RT-03,RT-04,PK-05,RP-14) that iterate over many examples and say nothing about one. I discarded those. Also,examples/has 17 directories, not the 20 of my plan (a miscount).Method. A README names a row only if the row's register statement names the example, or the row's statement is plainly that example's capability and its case arm builds the example (
copy_example/examples/…); I read each statement. Purposes are paraphrases of the root README's own example table, which is the existing source for them. Every example is also underEX-08.Verify command (all).
cargo xtask verify-examples(the gate; there is no per-example filter, onlyLEXLEAN_EXAMPLES_PART=without:uor-atlasand the Atlas shard/join parts). By hand, one example:cd examples/<name> && lexlean check && lexlean verify(pinned Lean,leancheckerreplay, exact axiom audit; it does not do theexpected/comparison).nat-add-zeroEX-01..EX-06peano-arithmeticVR-23(verification as the parts of one run)propositional-logiclist-inductionsemantic-1.1language-1.2recursive-dataDF-12,DF-13,SM-24higher-orderDF-14,DF-15,SM-25,SM-26recursionDF-16,DF-17,SM-27,PF-19collectionsDF-18,SM-28,SM-29,SM-30fold-lawsPF-20,PF-21,PF-22,PF-23productionPD-01..PD-06,NE-01,PX-01,PX-02,SP-02,SP-05,SP-08,VR-13,OR-02production-coveragePX-01,SP-02,SP-05,SP-08,SP-10,OR-02modelsMD-08,MD-11,MD-12,MD-13,SP-02,OR-02reasoningRS-03,RS-11,RS-12,RS-14,SP-02,OR-02helpdeskEX-09,SL-03,SL-04uor-atlasVR-19,VR-23Findings from the list:
collections,higher-order,language-1.2,production,production-coverage,recursion,recursive-data) carry\title{Natural number addition}inMain.lex.tex, a copy of the literal example's title. The title is part of the canonical document, so changing it movesexpected/records; I will not touch it. Gate-affecting finding (no register row; a sweep ofexpected/withcheck-golden --writeandverify-examples --write), for the maintainer to decide.list-induction,propositional-logic,semantic-1.1,language-1.2andpeano-arithmeticare named by no capability row; they stand onEX-08(andVR-23for Peano). Their READMEs say exactly that. The capability table will list them as gaps in negative coverage unless a row is wanted.grep -ril 'gnaf\|optimal' examplesfinds none), which is what the production-family READMEs state.READMEs are written in the draft PR (
claude/issue-34-closure). Update:nat-add-zerohas none:SE-07holds that project's root to a closed file list (see the gate record and the capability table), so #250 has 16 READMEs.Generated by Claude Code
afflom commentedon Oct 21, 2026
Gate record 1: the cheap recipes, run here (evidence for the audit, not the discharge)
Host. Linux x86-64 container, 4 cores, 15 GB RAM, no swap; Rust 1.97.1 (pin), Lean 4.32.1 via
elan toolchain install "$(cat lean-toolchain)",just1.59.0 andcargo-deny0.20.2 bycargo install --locked. No pin changed. Head:claude/issue-34-closure@f71ae0e44(=release/v0.4@b8ff7bef6+ the 17 example READMEs, #250). Cold build, so the first recipe of each kind includes compiling.just vvorder)fmt-checkmodel(cargo xtask validate-model)spec-linkslint(clippy-D warnings)featuresinventorygolden(check-golden)denycargo xtask check-packagelanguage 1.0,compiler-semantics 8554008034f6…,lean-toolchain leanprover/lean4:v4.32.1)A defect the gate caught in this PR's first commit.
validate-modelwas red on the first READMEs:audit-oracle-isolation(R2, §17.19) refuses any mention of the differential oracle underexamples/; I had named it in the production-family READMEs. Reworded to cite theOR-*rows; green onf71ae0e44. Not a plant, but it shows the isolation audit fires on prose inexamples/.Running now, in
vv-rowsorder, one at a time (the container cannot run legs side by side):test; theexamplesleg (without:uor-atlas, width 2, 12 GiB budget, the committed hints);repro; then eachbddpart alone,preservationlast and by itself (the maintainer measured ~10.4 GB RSS for it). UOR Atlas: its CI shards need the parts of one run and a 12 GiB budget; if it does not fit I will cite the CI run rather than report it as run. I will say per leg which is run here and which is cited.Generated by Claude Code
Bind four README capability sentences to what their certificates esta…
afflom commentedon Oct 21, 2026
Gate record 2:
examplesleg andtest, run here (evidence for the audit, not the discharge)Same host and head as record 1 (Linux x86-64, 4 cores, 15 GB, no swap;
claude/issue-34-closure@48eb261ea, which differs fromb8ff7bef6only by README prose).test(workspace unit tests,repo-conformanceexcluded)examplesleg =LEXLEAN_EXAMPLES_PART=without:uor-atlas, width 2, 12 GiB budget, committedverify-hints.json(the CI profile)What the
examplesleg ran. 21 projects, each formatted, locked, checked, built and verified with real Lean 4.32.1,leancheckerreplay and the exact axiom policy, and each project's normalized verification records compared withexpected/verify(§29.5): the 16 examples other thanuor-atlas, pluscompiler,stdlib,bootstrap,std-models,std-reasoning(the gate treats those as examples, §28.6). Per project (verify-examples: <name> took N s):production-coveragefold-lawsbootstrapstd-reasoningreasoningrecursionstdlibcollectionsmodelsstd-modelsproductionsemantic-1.1helpdeskpeano-arithmeticcompilerrecursive-data,propositional-logic,higher-ordernat-add-zero,list-induction,language-1.2Times are for this 4-core box at the CI profile and are not CI's. The projects ran one after another here; in CI the leg is one job.
Not run here:
uor-atlas. Its CI verification is sharded (.github/verify-shards.json) over several jobs of one run and joined; I did not reproduce the shard/join run. I will cite the latest greenrelease/v0.4run for it (see the final record). It is not reported as verified by this session.Still to run, in this order:
repro, then thebddparts alone (bootstrap-check,bootstrap-validate,compiler,production,verification,preservation-lists,preservation,rest).Generated by Claude Code
Drop the nat-add-zero README: SE-07 holds the literal example's root …
afflom commentedon Oct 21, 2026
Gate record 3:
reproand thebddparts, run here, with the CI run that discharges the restSame container as records 1-2 (Linux x86-64, 4 cores, 15 GB, no swap). Head:
claude/issue-34-closure@d5f6c405d(see the README note underverification). The parts ran one at a time;preservationran alone. Per.github/vv-rows.json.repro(check-reproducibility)rest: 1921 s total for the legbddpartbootstrap-checkbddpartbootstrap-validatebddpartcompilerbddpartproductionbddpartpreservation-listsbddpartrest(library tests, doctests, meta-gate, remaining IDs)restlegbddpartverificationconformance_se_07(90 passed)bddpartpreservationconformance_sp_02(13 passed)The two failures, and what they are.
conformance_se_07was my change:examples/nat-add-zero/README.mdis a file the project root "gains", andSE-07(§25.6, confined staging) holds the literal example's root to a closed set (.lexlean,NatAddZeroHost.lean,lake-manifest.json,lakefile.toml,lean-toolchain,lexlean.lock,lexlean.toml,src). I removed that one README (d5f6c405d, [v0.4] #34 closure: example READMEs and README claim wording #250).conformance_se_07alone then passes (0.72 s). No other case failed on a README: the 16 other example READMEs are in the tree for every part above. The wholeverificationpart is being re-run ond5f6c405dand will be recorded; until then this row stays FAIL for the part as run. Consequence for the capability table: the literal example cannot carry a README without editing the allow-list incrates/conformance/src/cases/security.rs:849, which is a gate change and so belongs here, not in the PR.conformance_sp_02failed withLLS8002: "child_timeout_msexceeded byleanin phaseLexLeanPreservation.RustLemmas: configured 300000, observed 300008 ms". That is a resource limit on this box, not a failed proof or a semantic mismatch: 8 ms over the project limit, on the compile of one Lean module, on 4 cores with no swap. In CI the same part is green in 3790 s. I am not counting it as run-and-passed. It is being re-run alone; the result goes in the next record. Finding for the maintainer:child_timeout_ms = 300000leaves this module essentially no headroom on slower hardware, and the failure mode (a resource limit reported as a gate failure on a conformance case) is exactly what R2/R9 intend; I changed nothing about it.What discharges the criterion. The criterion is the gate on the release candidate. For the legs this box cannot discharge, the evidence is CI run 39520454225 (pull_request, head
53c7a8f60, green, 28 jobs):b8ff7bef6and53c7a8f60have the same tree (2a203993ac09…both), so it verified the tree of this line, and the later push run39529604597only recorded that. Its jobs, allsuccess:plan;just vvforexamples(2261 s),rest(1921 s),bdd-bootstrap-check,bdd-bootstrap-validate,bdd-compiler,bdd-preservation,bdd-preservation-lists,bdd-production,bdd-verification; all five hosts of §8.3 (x86_64-apple-darwin,aarch64-apple-darwin,aarch64-unknown-linux-gnu,x86_64-pc-windows-msvcas "build and non-Lean tests", Linux x86-64 in thejust vvlegs); and the UOR Atlas jobs (prefix, cert4, cert4-replay, census-a..d, pairs-a..d, orbit, closure, join), which I did not run here and cite only.just vvlegs (run 39520454225); this session's legs abovebuild and non-Lean tests (…), allsuccessin the same run; not run hereWhich is which. Run here and passed:
fmt-check,model,spec-links,lint,test,features,inventory,golden,deny,check-package,repro,examples(21 projects),bddpartsbootstrap-check,bootstrap-validate,compiler,production,preservation-lists,rest. Run here and failed:bddverification(fixed, re-run pending),bddpreservation(resource limit, re-run pending). Cited, not run:uor-atlasshards and join, the four other hosts, and the CI-measured times above.Generated by Claude Code
afflom commentedon Oct 21, 2026
v0.4 capability table: positive end-to-end artifact, negative fixture, gaps
Base
release/v0.4@b8ff7bef6, register 436. Rows follow the work graph of #15, plus #91/#92 (still landing). Method and limit. Positive = a committed project thatverify-examples(or the project's own gate) verifies with real Lean; negative = a committed directory undertests/negative/(285 of them), a committed negative certificate undercompiler/certificates/, or an inline defect a conformance case plants and the diagnostic code it asserts. I read the directory names and the case arms by hand; the automatic scan I tried could not tell inline negatives from helper-function ones, so "inline" below means I found the asserting arm, not that I audited each plant. Planted-defect records for gates are in VERIFICATION.md (§27.9) and were not re-derived.examples/language-1.2,examples/semantic-1.1; 1.0/1.1 examples byte-stable underexpected/language-1.2-construct-under-1.1,semantic-module-v1-under-1.2,semantic-module-v2-under-1.1,lock-schema-language-mismatch[-1.1],cross-version-glossary[-1.1],package-language-mismatch,unsupported-language-version,malformed-language-versionexamples/recursive-datarecursive-type-positivity,-nonuniform,-uninhabited,-forward,-type-arguments,recursive-mutual-uninhabited,recursive-constructor-mismatch,recursive-match-foreign-constructorexamples/higher-orderlambda-capture-extra/-missing,lambda-type-mismatch,application-arity,application-non-function,escaping-closure[-in-data/-parameter/-structure],production-higher-order-escapeexamples/recursionrecursion-*(cyclic measure, false/forged/missing evidence, non-decreasing, wrong argument, mutual …),mutual-group-noncontiguous,mutual-label-shared,polymorphic-recursionexamples/collectionscollection-*,graph-duplicate-edge,graph-invalid-referenceexamples/production*(extraction inverify);tests/lcnf/registry + oracleextraction-rejected,extraction-authority-drift,extraction-uncompiled-externalexamples/production,examples/production-coverage,examples/recursionproduction-*,eligibility-report[s-total]-limit,executable-calls-formalcompiler/(fixtures/, rendered and compiled by the pinned rustc, TC-07)compiler/gnaf/(60 requests) with kernel-decided answersuncovered_scope, …), inlinecompiler/certificates/(73 certificates, kernel-checked verdicts)examples/project carries a GNAF certificate (grep), so no example makes an optimality claimtests/production-export/{production,production-coverage,helpdesk,stdlib,bootstrap}byte-compared packagesexamples/{production,production-coverage,models,reasoning}(certificates A, B, E inverify)certificate-rejected,certificate-b-rejected,certificate-e-rejected,preservation-drift,certificate-{generation,size,heartbeat,total}-limit,certificate-resource-exhaustedlean4-proddifferential oracle (#26)tests/oracle/renderings/(7 files)tests/oracle/renderings/(e.g.*.wrong-on-one-input.rs), OR-04; known-difference tables, OR-05bootstrap/(5 components, certified Rust packages,tests/production-export/bootstrap/)rust::{lower,print,package}and steps 5-7 of #92 wait on #91 stage 4tests/production-export/*vialexlean production build;examples/helpdeskproduction-request-malformed,-root-unknown,-root-not-executable,-target-unregistered,-unknown-target,-expectation-stale,-named-type-boundary, …examples/modelsmodel-*(artifact digest/length/role/shape/symlink…, forged/false/vacuous evidence, composition junctions, float descriptor, bypass)examples/models,tests/text-laws/(native laws module),examples/helpdesk(classifier-adjacent text path)cases/language_models.rs)examples/reasoningreasoning-*(accept/extract/rule bypass, forged trace, unsound rule, nonterminating, under 1.1 …)examples/helpdesk(+std,std-models,std-reasoning)source-package-outside-workspace; EX-09 namesLLC0105,LLR3009,LLR3007,LLT4008as inline refusals;model-artifact-*crates/conformance;helpdeskis the application fixtureexamples/uor-atlasatlas-level-conflationcargo xtask release-check1.0.0naming vs. a v0.4 line is the maintainer's at the version bumpSummary of gaps. Four v0.4 capabilities have no committed negative fixture directory, only inline planted defects: TC (#21), GN (#22), RB (#24) and LM (#30). Whether the DoD's "relevant negative fixture" accepts an inline plant asserting a registered code is the maintainer's call; if it does not, each needs a directory under
tests/negative/(R3: scenario and test already exist for these rows, so only fixtures plus theEX-07coverage list would move). Pending, not gaps: #30 slice 4, #91 (RB-15/16), #92 steps 5-7.Related finding (gate-affecting, for you).
examples/nat-add-zerocannot carry a README:SE-07(crates/conformance/src/cases/security.rs:849) holds the literal example's root to a closed file list, and the conformance cases copy that project (P::example()). #250 therefore has 16 example READMEs, not 17. R3 order if you want one: no new row; edit the allow-list in theSE-07arm, plant a stray root file and record the failing signature in VERIFICATION's falsifiability table.Generated by Claude Code
19 remaining items
afflom commentedon Oct 25, 2026
Maintainer, rehearsal result: run 40221304393 (on #318's second head,
1d6acf575) finished with eight jobs green (tag identity, crate readiness, the five host binaries) and one red: its release gate'sjust vvfailed inconformance_rp_18only, the RP-18 oracle reading the oldcrate.ymllines, which #318's third commit fixed; the tag commit93fbe7df9carries that fix and its own gate passedrp_18. Nothing else in the release workflow failed. To have the whole workflow proved at the exact tag commit, one more rehearsal is dispatched onrelease/v0.4at93fbe7df9; it publishes nothing (sha-93fbe7dimage tag only) and does not hold the tag: the tag can be pushed at any time, and if its run ever fails in a job, that job is re-run without re-tagging.Generated by Claude Code
afflom commentedon Oct 26, 2026
Hold the tag: do not push
v0.4.0at93fbe7df9yet.Rehearsal 3 showed the release workflow's
just vv (release evidence)job cannot finish: on one hosted runner thebddrecipe alone took 4 h 14 min (run 40221304393), and rehearsal 4 (run 40244034460, at the tag commit) is about to hit the six-hour job limit in the same job. A tag pushed on93fbe7df9would fail there, andimage,manifest, andpublishall wait for it.The fix is in progress on
release/v0.4: the release gate reads theci.ymlrun of the tagged commit that was the whole gate (the sharded gate that already runs on every push to the line and passed on93fbe7df9, run 40233071771) and records it as the §30.3 evidence, instead of re-running the gate on one runner. SPEC §9.4/§30.3 and the RP-18 oracle change with it. When it lands, the tag commit moves to the new head ofrelease/v0.4; I will post the exact commands here again, and rehearse once more first.Generated by Claude Code
afflom commentedon Oct 26, 2026
The
v0.4.0tag commit isaa6593e026d92f4696fe970715c81528ad2c37daonrelease/v0.4: the merge of #321 (release notes for the cut that carries H5 and the release-workflow fixes), after #320 (the release gate reads the gate run of the tagged commit; rehearsal 5, run 40292173816, was the first fully green rehearsal: tag identity, release evidence in under a minute, the five host binaries, crate readiness, both images and the multi-architecture tag). The register is 457 and the three compiler-semantics IDs are the ones the locks pin after H5 (CHANGELOG0.4.0).Rehearsal 6 is running on this commit now; I will post one line here when it is green. Then the tag, which only you can push:
Push the tag only;
release.yml'spublishjob creates the GitHub Release (binaries, image, crate, SBOM, checksums, gate evidence) and uploads to crates.io only ifCARGO_REGISTRY_TOKENorCRATES_IO_TRUSTED_PUBLISHINGexists, saying so otherwise. Landings onrelease/v0.4are held until the tag exists, so the commit does not move again;mainreceives the line through a PR after the tag.Generated by Claude Code
afflom commentedon Oct 26, 2026
Rehearsal 6 on the tag commit
aa6593e026d92f4696fe970715c81528ad2c37dais green (run 40304904506: every job passed, the two tag-only publish steps skipped as designed). The tag is ready to push with the commands above.Generated by Claude Code
humuhumu33 commentedon Oct 26, 2026
The
v0.4.0tag was pushed at 12:17Z onaa6593e026d92f4696fe970715c81528ad2c37dawith the commands above (annotated tag object699dd53a96524b62a7320b2e0cf96048b14de832). The tag run, 40321415513, failed at its last job,publish the release, stepRead the release criterion over the assembled set.Every other job passed: tag identity, release evidence, the five host binaries, crate readiness, both images and the multi-architecture tag (the development pre-release upload was skipped as designed).
Failing criterion.
cargo xtask release-checkrefused withsource-tagandversion-output(the two the step forgives under §2.3) and with one it does not:Cause.
crates/model/src/release.rs:453requires the literalvv: the acceptance gate passedinrelease/vv-evidence.txt. Since #320,.github/release-gate.shwrites that file from the shardedci.ymlrun of the tagged tree (here run 40292537147 ond985f9561, the second parent, every job passed) plus the legs'vv.txttranscripts. None of the 29 leg transcripts prints that line; only the single-runnerjust vvdoes. #320 changedrelease-gate.sh,ci.yml,release.yml, SPEC, VERIFICATION and RP-18, notrelease.rs. The rehearsals skippublish the release, so this criterion first ran on the tag.What was published.
ghcr.io/uor-foundation/lexlean:0.4.0(with0.4.0-amd64,0.4.0-arm64) and:latest.What was not. No GitHub Release, so none of the binaries, SBOM, checksums or gate evidence, and no crates.io upload: the step stopped before
Decide how the upload is authorised. The run reportsHAS_CARGO_REGISTRY_TOKEN: true, so the next run that passes this step uploads the crate.The tag is untouched. The fix and whether the tag moves are yours.
Generated by Claude Code
afflom commentedon Oct 26, 2026
Thanks for the diagnosis; it is exact. The fix is #323 (
release-gate.shends the evidence with the linerelease-checkreads; the criterion is one function,repo_model::release::ci_evidence, with its own test, andRP-18holds the script's output to it, so the publish job can no longer refuse what the rehearsals accepted). Its gate is running.The tag moves. The publish job runs the workflow of the tagged commit and downloads the evidence its own gate job wrote, so a re-run of the failed job would refuse the same file;
aa6593e02cannot be released as tagged. When #323 is merged and rehearsal 7 on its merge commit is green, I post the new commit here, and the tag is re-made from it:Nothing was published under the first tag (no GitHub Release, no crates.io upload), so the move loses nothing; the image tags
0.4.0andlatestare overwritten by the next run. Landings onrelease/v0.4stay held until the tag exists.Generated by Claude Code
afflom commentedon Oct 26, 2026
#323 is merged and rehearsal 7 on its merge commit is green (run 40339000887: every job passed, the two tag-only publish steps skipped as designed; the evidence job named gate run 40325802734 and the file ends with the line
release-checkreads, checked byRP-18and the criterion's own test).The
v0.4.0tag commit ise0b74b81e6f49c83596a957d232c8383a5db1a25. Move the tag:Nothing was published under the first tag, so nothing is lost. I watch the tag run through
publish the releaseand post the result here.Generated by Claude Code
afflom commentedon Oct 26, 2026
LexLean v0.4.0 is released.
v0.4.0ate0b74b81e6f49c83596a957d232c8383a5db1a25(release/v0.4).publish the release(cargo xtask release-checkaccepted the evidence; the CI evidence is the gate run of the tagged commit, per release-gate.sh ends the evidence with the line the release criterion reads #323).lexlean 0.4.0(published 19:20Z).Next: release/v0.4 is merged into
mainby PR (the line at its tag, SPEC §9.4), the hold on release/v0.4 landings is lifted (patch work on the line continues as 0.4.x; feature work goes to release/v0.5), and the second forward merge into release/v0.5 follows.Generated by Claude Code
afflom commentedon Oct 26, 2026
Closing: v0.4.0 is released from
e0b74b81e(release run 40382770933, every job green including publish; GitHub release with 15 assets; crates.iolexlean 0.4.0;mainat the tag via #324).Against the acceptance criteria:
just vvfrom a clean checkout on the normative host: the release'svv-evidence.txtis the gate run of the tagged commit's tree (every job, everyjust vvleg, Atlas join), checked byrelease-checkat publish.leancheckerreplay and the exact axiom policy, reproducibility and golden checks: all jobs of that run and of the release run (release-check.txt,MANIFEST.sha256,release-identity.json, SBOM, checksums).Remaining v0.4-milestone work (#91 heap half, #92 step 4, #30 slice 4) lands on
release/v0.4as 0.4.x patch releases per the owner's rule on #92, not as a condition of this closure. #256 (#251 retitles, #252 fixtures) stays queued for 0.4.1. Umbrella #15 carries the release note.Generated by Claude Code
humuhumu33 commentedon Oct 26, 2026
v0.4.0 is published
The tag was moved to the commit the maintainer named, and the tag run went through
publish the release.Tag.
v0.4.0now points at commite0b74b81e6f49c83596a957d232c8383a5db1a25through annotated tag object22698a589f9f4d91b65c498f9de057c98fc4a4dc. The old tag object699dd53a96524b62a7320b2e0cf96048b14de832(commitaa6593e0) is orphaned; nothing had been published under it.Run. 40382770933 (
push, tagv0.4.0, commite0b74b81) completed success. Every job passed —tag identity,release evidence, the five host binaries,crate readiness, both images,multi-architecture tag, andpublish the release; the development pre-release upload was skipped as designed.Release. v0.4.0, published
2026-10-26T19:20:36Z, 15 assets:lexlean-0.4.0-binaries.tar.gz,lexlean.crate,sbom.json,vv-evidence.txt(194,960 bytes),release-check.txt,release-identity.json,version-output.txt,compiler-semantics-id.txt,MANIFEST.sha256,checksums.txt,SPEC.md,CONFORMANCE.md,ERRORS.md,LICENSE-APACHE,LICENSE-MIT. The image isghcr.io/uor-foundation/lexlean:0.4.0@sha256:d53355cb3c678ad342e77502d3a43e450a0fe2064315d8d3ecbea5d231547c08.The criterion. The shipped
release-check.txtnames only the two criteria §2.3 defers at this version:ci-evidenceis no longer among them, so #323's fix (release-gate.shending the evidence with the linerelease-checkreads) is confirmed on the publish path, not only in rehearsal.version-output.txtreadslexlean 0.4.0/language 1.0/compiler-semantics 07859acb5a756221018b7f65e7f725d99b00a694670eecc7cd264832ba1333fb/lean-toolchain leanprover/lean4:v4.32.1.Open points. Whether the crate also reached crates.io is not stated in the release — RP-18 permits a release without the upload. The release object records
target_commitish: maineven though the tag points ate0b74b81, which is onrelease/v0.4. The forward merge ofrelease/v0.4intomainis next.afflom commentedon Oct 26, 2026
Both open points are settled:
lexlean 0.4.0is on crates.io, created2026-10-26T19:20:26Z(the registry's/api/v1/crates/lexlean/versions), so the upload step ran, not only the release without it that RP-18 permits.target_commitish: main: cosmetic. A release object made from an existing tag resolves to the tag's commit (e0b74b81e) regardless of that field, andmainhas since received the line at the tag through main receives the v0.4.0 line: release/v0.4 at e0b74b81e (the tag) #324 (mergeb9793961c), so the field now also names a branch that contains the commit.Generated by Claude Code
humuhumu33 commentedon Oct 26, 2026
Follow-up to the line above — the crates.io question is settled: the crate was published.
crates.io/api/v1/crates/lexleanlists one version,0.4.0, not yanked:That checksum is the same string as the
lexlean.crateline of the release'schecksums.txt, so the crate on crates.io is byte-for-byte the crate the release attaches — the equality step 12 ofpublish the releaseasserts (the crate cargo would upload is the crate this release attaches).The job's own steps on run 40382770933:
and the log reads
Uploading lexlean v0.4.0→Uploaded lexlean v0.4.0 to registry crates-io→Published lexlean v0.4.0 at registry crates-io, withHAS_CARGO_REGISTRY_TOKEN: true. So the first upload took the crates.io environment secret; trusted publishing (#260, RP-18) was not exercised, its two steps correctly skipped.Both open points from my earlier comment now stand as: this one settled, and the release object's
target_commitish: main— which is moot for consumers, since the tag itself namese0b74b81.