CLAIM: in-repo HPE checkers do not read the pre-wrapper recipe

[hevlyo]

Status: CLAIM. Seed d9b349359a3c923c0212b9514fe130b3, scope key e3ffea67887eeb4c396c, base b5ea56e0a5ac4552166253b1fa4be0753bb38c30. Selected record unchanged.

Object. The hrec conjunct of hpe_word_exact: factor_iff_pre requires the executed pre-wrapper recipe R' to equal R for the certified pair (E, U). The two in-repo checkers are scripts/verify_hpe_lean_binding.py and research/hyperplane-end-lean/code/check_hpe_pairs.py.

Not the same as the finished semantic port of the three Clifford pairs (discussion #1394, PR #1415) or the HPE axiom-probe audit (discussion #1593). Overlap on 2026-10-15 covered 782 open PRs and 1329 discussions. Those matches are other formal-interface threads.

Method. State the boundary lemma, then show a synthetic 38-helper word whose pre-wrapper records name E:FULL is accepted by both checkers, while a phase flip and h += p are rejected. No Lean rebuild and no emitted run.

Stop. One binding gap. No new kappa, no selected-frontier edit.

Prepared by Riquelme Soares (@hevlyo) with substantial AI assistance. Apache-2.0.

[hevlyo]

DONE. Head 94d77d2cf19b5fbf3d14cd145fc5fd6ad8d24bd6. Package research/hpe-prewrapper-binding. PR #2706

factor_iff_pre says the wrapped block equals C₇ ⊗ id if and only if the executed pre-wrapper recipe equals the certified (E, U) recipe. Both in-repo checkers accept a synthetic 38-helper word whose pre-wrapper records name E:FULL, and they still accept when that record is deleted. A phase flip and an h += p gate are rejected.

Limits: the witness is not an emitted word. This does not rerun Lean, hpe_pre_wrapper_resolved_to_full_recipe, or the selected baseline. No new kappa and no selected-frontier change.

Prepared by Riquelme Soares (@hevlyo) with substantial AI assistance. Apache-2.0.

[petalas]

Status: DONE / RELEASED, 2026-10-15 19:54 UTC, priority revision 5.

Author-checked finite result: actual event-to-literal bindings pass for all 30 non-cold selected k8 HPE helpers in both orientations. Public artifact PR2748, exact commit 2af9c0fc46492a4ce56ac20ab14cc40e2f6544e0. Ten tests with 360 per-helper controls, the registered replay and a clean committed-source export pass locally. This complements the existing production event audit and Hanson Sun's helper450 binding; it is not a production-defect claim. Scalar restoration, serialization correspondence, native/physical admission and a new kappa remain outside scope. Next useful task: an explicit Lean event word and correspondence proof. No active claim.

Original claim preserved below.

---

CLAIM, 2026-10-15 19:40 UTC, priority revision 5. Expires 2026-10-15 21:40 UTC.

I am taking a scoped verification follow-up to Riquelme Soares's PR #2706 at 94d77d2cf19b5fbf3d14cd145fc5fd6ad8d24bd6, from current main bd1b3bc7174cddbfa4e30ac073dc134f48aa6e77.

Question: can a small opt-in checker bind the actual LOCAL-PROGRAMS.json.gz pre/post residual-adapter events, their chronology and their selected recipe keys to the retained Lean pair literals? The original production hpe_audit_local.py already checks wrapper references and ordering. Hanson Sun's k8-hpe450-binding also checks those events, with the literal/row/column result scoped to helper 450. I will preserve that distinction, not claim a production defect or duplicate the Clifford identity audit.

Method/domain: independently parse the actual emitted event grammar, cover every non-cold HPE helper in both pinned k8 orientations, and reject omitted, duplicate, wrong-terminal, misplaced and phase-corrupted wrapper controls. Reuse only the source producers to regenerate input and the literal reader to identify the existing Lean data. No imported mathematical checker is the oracle. Expected artifact is a source-pinned successor package and focused PR. No frozen checker, source pin, frontier or supplier-admission change.

Evidence status at claim time: LOCAL-ONLY. Both genuine k8 event streams have been regenerated and match their published pins. Full verification remains to be implemented.

[petalas]

DONE / RELEASED, 2026-10-15 19:54 UTC, priority revision 5.

Public artifact: PR #2748, exact head 2af9c0fc46492a4ce56ac20ab14cc40e2f6544e0, package research/hpe-wrapper-event-binding/.

Both genuine selected k8 orientations pass the new independent event-to-literal check for all 30 non-cold HPE helpers, 24,767 actual events each. The pre/post events select 3:998 for 16 helpers and 2:999 for 14 helpers; both selected monomials equal the retained Lean pair data, including phase. Ten tests pass, including 360 per-helper deletion/terminal controls and real CLI optimized rejection. Registered Python 3.12 replay, 70 research-workflow tests, lint, committed-blob equality and a fresh committed-source export replay also pass locally.

This does not diagnose a production defect. The original hpe_audit_local.py already checks emitted references and chronology. Riquelme Soares's PR2706 identifies the lightweight metadata-only boundary; Hanson Sun's helper-450 package supplies the pinned regeneration harness and the prior chronology/single-helper binding. Their credit and the formalization/compiler/selected-word lineage are retained in the package.

Limits: no scalar endpoint or isolated-column claim for the other helpers, no cold/FULL-ended helper coverage, no fresh Lean build, complete hrec, physical hosting, supplier admission or new kappa. Next useful proof task is an explicit Lean event word with private COPY/CLEAR lifetimes and serialization correspondence, starting from the existing helper-450 evidence. This claim is released.

[petalas]

Status: DONE / released, 2026-10-16 13:28 UTC, priority revision 5. Author-checked result at PR2944, head 2c97782a83da48230f1cffca73c5a7ad58ada31e.

Both complete selected k8 programs now have source-bound concrete Lean WellLabelled and final-label certificates. Four kernel roots and the committed read-only export pass. This is a label instance for translated words; the serialized Python/JSON semantic bridge and operator/physical supplier obligations remain separate. Evidence, credit and next formal handoff. No new kappa and no active reservation.

Original claim follows.

---

CLAIM / 2026-10-16 / priority revision 5. LOCAL-ONLY at entry. Expires 2026-10-16 15:10 UTC.

Question: can both complete selected k8 emitted programs be translated to the retained Lean Event grammar and checked as concrete WellLabelled instances, including private COPY/CLEAR and all wrapper tags?

Base 9ee73fe. Follow-up to Sagar Thakrar/PR1964 Python-only H1 audit and Hanson Sun/PR1611 and PR2748 event bindings. Generic COPY/CLEAR conjugation is already in LemmaG.Stream and is not new work. Cold/noncold operator factorizations are also covered elsewhere.

Method/domain: regenerate both pinned 24,767-event k8 streams, translate every semantic event to the existing formal constructors with exact rational coefficients, bind the translation to original event hashes, and kernel-check concrete label admissibility. Test missing/moved adapters, wrong private-copy frames and wrapper changes. Bounded attempt, roughly 30 minutes plus verification; stop or narrow if full kernel reduction is too costly.

Expected artifact: focused source-bound translation/checker package and PR. This is H1 only. Exact RecipesExact, scalar restoration, operator semantics, hosting, all-size, precision/tape, full supplier replay and kappa remain outside scope. Substantial automated assistance; no independent human review claimed.

[petalas]

FINDING / DONE / RELEASED, 2026-10-16 13:28 UTC, priority revision 5. Author-checked finite formal instance; no independent human review claimed.

PR2944, exact head 2c97782a83da48230f1cffca73c5a7ad58ada31e, adds research/k8-lean-label-instance/. Lean proves Stream.WellLabelled and exact final labRun labels for both complete translated selected k8 programs: 24,757 semantic events each, including private COPY/CLEAR and all wrapper tags. The generic event semantics are reused. Kernel reduction uses only the three retained axioms and no native_decide.

Both complete emissions match the selected hashes. The final driver freshly reproduced all four roots in 420.1 seconds. Ten actual-program semantic controls, ten focused tests, 26 onboarding tests and Ruff pass. All 27 committed package/input files match, and a read-only Git export without Git passes the replay. Internal reviews checked all generated events/checkpoints and found one receipt-snapshot defect, now fixed. Hosted CI is separate.

This is H1 for the translated fixed words. Python/JSON semantic correspondence, exact recipe instantiation, scalar endpoints, physical hosting, all-size, precision/tape and supplier admission remain outside this package. Kappa is unchanged.

HANDOFF: join this concrete instance to existing exact recipe and helper450 endpoint evidence in one formal theorem. Start with explicit label operators and serialized-event correspondence; another numerical operator replay does not close that bridge.

Credit Sagar Thakrar/PR1964, Hanson Sun/PR1611, DreamingOfClouds, Jeremy Kothe, icekylinx and the retained compiler/word/packing lineage credited in the README, plus PR2748. Substantial AI assistance is disclosed; no active claim remains.

[thakrarsagar]

DONE / RELEASED — concrete Lean unit row for selected k8 helper 450. Sagar Thakrar, with substantial OpenAI Codex assistance; priority revision 6, 2026-10-17.

Public artifact: PR3357, exact head c6a4a024ef3eb9e3296152e8bb92485660138c10, package research/k8-helper450-lean-row/.

Lean now proves ∀ j : Fin 1024, Stream.S K8.Forward.events 450 j = if j = 450 then 1 else 0 for the complete existing PR2944 forward word. The stronger state theorem restores scalar coordinate 450 for every initial state, including arbitrary private-coordinate contents. This supplies the concrete forward hS premise; it is not a new serializer or a cold-private assumption.

The exact 24,757 semantic events and all 14 generated sources remain byte-identical to PR2944. Two opposite ADDs at semantic positions 13,944 and 16,501 cancel because four kernel-checked target-avoidance predicates preserve the required coordinates. The proof uses the original 97 blocks, with prefix/middle/suffix lengths 13,944 / 2,556 / 8,255.

Fresh Lean 4.21 replay passed in 591.5 seconds with the default stack: 21 source compilations, five main axiom roots, three rejected false claims (wrong sign, intervening CLEAR, aliasing), and positive relabel/row-versus-column controls. Generic roots use propext, Quot.sound; concrete roots use propext, Classical.choice, Quot.sound. Registered isolated Python 3.12.15 replay also passed both commands and seven independent source/arithmetic control groups. The ordinary registered replay checks source/receipt bindings; fresh kernel invocation is separate and documented.

Credits: Hanson Sun's PR1611 numeric binding; Nick Petalas's PR2944 complete typed words, translator, generator and labels; Cinna Davis's PR3331 composition lemma; the retained Lemma G/HPE/compiler lineage. Separate agent review is not independent human review.

Limits / handoff: the Python/JSON adapter remains outside the kernel. Backward hS, concrete RecipesExact, label interpretation, physical ownership/footprints, and the full selected cold replay remain separate. A unit row alone does not establish isolated dirt-column restoration. No new κ or selected-frontier change. The claim is released; the package lists concrete next questions.

Final overlap refresh covered 263 open PRs, 2,761 closed/merged PRs and 2,111 Discussion records with full reported pagination and no warnings; live #2703 was checked separately. This coverage is not a novelty certificate. Canonical main was refreshed at submission; the fork is synchronized and the contribution branch rebased.

[thakrarsagar]

DONE / RELEASED — source-bound forward recipe dispatch. The scoped claim opened at 23:04 UTC is complete; the broader supplier obligations below remain open.

Public artifact: PR3377, exact head 78c4bbdaf2542ebea60fc585ad62395224f8ff2f, package research/k8-lean-recipe-dispatch/.

The Lean theorem proves RecipesExact (tagOp tables model) (dispatch tables model) K8.Forward.events for the unchanged 24,757-event forward word, conditional on the 2,888 used child-operator identities. The model supplies independent frame, monomial and core operators; the dispatcher executes the source-selected monomial/core recipes. A separate theorem binds all 5,064 relabel descriptors in exact order. The source/target, residual-wrapper and identity cases are discharged by algebra.

Evidence: fresh Lean 4.21.0 compilation of 34 modules, all 97 block checks, 20 theorem/control axiom reports, and four rejected false claims (inverse direction, noncommuting order, terminal frame, child composition order). Only propext, Classical.choice, and Quot.sound occur; the child-count theorem uses none. Eight independent Python control groups pass under isolated optimized CPython 3.12.15, including literal helper450 2:999 versus FULL key 2, missing/duplicate occurrences, stride 712 versus 1024, malformed ranks and source corruption. The registered replay explicitly validates a stored receipt; the separate driver performs a fresh kernel run.

Remaining limitations: the used child identities and serialized Clifford operator denotation are not proved here. The Python/JSON adapter, native executor, physical ownership, common-chart interpretation, backward word and complete supplier theorem remain explicit. No new numerical recipe result or selected κ is claimed.

Attribution: PR1781, PR1964 and PR2048 already supply numerical checks and the extended-label interpretation; PR2378/2422 supply complementary monomial refinements; PR2944 supplies the exact typed word, codec and generator. Full predecessor attribution and exact heads are in the package. New work by Sagar Thakrar with substantial OpenAI Codex assistance; agent review is not independent human review.

Next useful handoff: bind one retained child record's independent operators to an exact serialized Clifford denotation, including global phase, then discharge its child identity. A physical invocation should also name its ownership/restoration contract before being connected to this dispatcher.

{
  "schema": "integer-mult-coordination/v1",
  "kind": "DONE",
  "task": "formal-interface-lemma",
  "scope": "Complete selected forward recipe dispatch and ordered descriptor binding, conditional on 2888 used child identities",
  "base": "6f100d4bb5083f8c35d96fd95a539811c62f245c",
  "expiry": null,
  "artifact": "https://github.com/CrocSwap/integer-mult-bounds/pull/3377"
}