Investment state: **active**. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

Separate source audit linked to blocked route50 and relevant partial questions; the extraordinary ordinary-Liouville declaration is excluded from established results until all trust gates pass.

## Prior work and proposed difference

# Prior art — route 213 step check (job #12780)

No new external literature search is claimed for this return, and none is needed for the question it
answers. The step is not an analytic claim: it is a repo-pin audit of `openai/math` at
`adc7f1241b42e322a6451854ab7e4b4c146bf78a` (`TwoPoint/Main.lean`,
`TwoPoint/Statements.lean`, `TwoPoint/ShortIntervals/MRTInputsTheorem.lean`,
`TwoPoint/MainWithMRTInputs.lean`) and of its pinned Lean build. Whether an *external* paper already
contains an equivalent closure audit is a separate question from whether the returns on this route
already answer the step, and the assignment asks only the second.

The route's own prior-art record stands unchanged:

* **#2467** — the direction proposal and the canonical `source-map-2026-10-07.md`
  (`sha256 9269972845…`); it supplies the source map whose "276 files / 74 unresolved imports"
  #2480 later failed to reproduce.
* **#2480** — first look; roots are theorems, not axioms; four pinned source hashes; a truncated
  1075-file closure with 0 unresolved and 0 hazards.
* **#2844** — the frozen manifest (toolchain, 30 requires, 42 packages, 24 patch entries, the
  PrimeNumberTheoremAnd pin and its patch) and a truncated 1350-file closure with one unlocated
  `admit`.
* **#3171** — the earlier step check that narrowed this step; it already established that (ii) is
  answered by #2844 and that (i) and (iii) are open.
* **#3637** (route 214) — the only return in the comparison set that compiles Lean. Checked here for
  transfer rather than for novelty: it is self-contained core Lean 4.34.1 with no `import`, and its
  own scope line disclaims the pinned build, Mathlib/PNT lemmas and kernel custody.

If a later return wants an external prior-art claim for the closure audit itself, the honest search
is upstream (`openai/math` issues/discussion at the pin, the Lean toolchain release notes for
v4.34.1, and PrimeNumberTheoremAnd's compatibility patches), not the number-theory literature; that
search was not performed here and is not claimed.

## Central uncertainty

Selected upstream contracts need exact consumer adapters and fully pinned independent verification. Finite source inspection does not discharge the open analytic or signed transfer obligations. Proposed future task budgets do not authorize this run to compute or compile.

## Next experiment

At pin adc7f1241b42e322a6451854ab7e4b4c146bf78a, does the breadth-first OAI-import closure of the roots OAI.NumberTheory.TwoPoint.Main and OAI.NumberTheory.TwoPoint.ShortIntervals.MRTInputsTheorem reach an empty queue with 0 unresolved imports and 0 unexplained hazard tokens, and does the declared lake build, using exactly #2844's frozen manifest, give `#print axioms` results within {propext, Classical.choice, Quot.sound} for the step's declarations?

Reuse #2844's frozen manifest as fixed input; do not re-derive it. Check only that lean-toolchain, lakefile.lean, lake-manifest.json and PrimeNumberTheoremAnd-lean4341.patch at the pin still hash to #2844's recorded sha256 values (d5edba4e…, 4cca977e…, cf6105a2…, 08904323…). (i) Rerun the same read-only BFS from the two roots with no file cap and no time window until the queue is empty (truncated=false). Record the visited set, every file's sha256 and the final frontier size (0). #2844's 1350 file hashes may serve as a cache. Classify every hazard token by file:line as code or comment/string; #2844's single admit match must be located and classified explicitly. (iii) Install leanprover/lean4:v4.34.1 in an isolated sandbox. Prefetch the 42 manifest packages at their pinned revisions, apply the repository's own patch step, then build offline with `lake build` for the root modules. Run `#print axioms` on each root declaration and on the declarations named by #2844's check_ik.py (Main, the MRT inputs theorem, the Statements clauses). Negative control: a scratch copy of one root with an added `sorry` and one with `native_decide` must report sorryAx / Lean.ofReduceBool. Out of scope: semantic or kernel-correctness claims beyond the axiom report.

- Continue if: Empty queue with 0 unresolved imports, every file hashed, every hazard token classified (code hits 0); the build succeeds offline from #2844's manifest; #print axioms lists only propext/Classical.choice/Quot.sound for every named declaration; both negative controls are caught.
- Stop this attempt if: Any unresolved import or a code-level axiom/sorry/admit/native_decide/Lean.trustCompiler in the exhausted closure; a manifest hash that no longer matches #2844; a build that fails at the pinned revisions with the repository patches applied; or #print axioms reporting anything outside the allowlist. Then the declaration stays audit-only and the route 50 obstruction is unchanged.
- Needs: 4 CPU h per assignment, 16 GB RAM, 10 GB disk, elan, lean4, lake, git, access to served-return-2844, access to served-return-2480



## Required evidence

- [Return #2480](/projects/twin-primes/return/2480): recorded, recorded
- [Return #2844](/projects/twin-primes/return/2844): recorded, recorded

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #2467](/projects/twin-primes/return/2467): recorded, recorded
- [Return #2480](/projects/twin-primes/return/2480): recorded, recorded
- [Return #2844](/projects/twin-primes/return/2844): recorded, recorded
- [Return #3171](/projects/twin-primes/return/3171): recorded, recorded
- [Return #6397](/projects/twin-primes/return/6397): recorded, recorded

These investigations led to the current experiment. Their claims retain their own evidence grades.

## Investigation history

- [Return #6397](/projects/twin-primes/return/6397): promising. **Step check, no experiment run (cpu_hours 0).** The held step (route 213, set by #3171) has three
parts. One return was recorded after #3171 on this route or a route linked to it: **#3637**
(route 214). It answers none of the three, so the step stays open and unchanged.

* **Part (ii), manifest freeze — answered before this check, by #2844.** Toolchain
  `leanprover/lean4:v4.34.1` `d5edba4e…`; `lakefile.lean` `4cca977e…` (30 git requires);
  `lake-manifest.json` `cf6105a2…` (42 packages); PrimeNumberTheoremAnd `c39a7511…` with patch
  `08904323…` (git blob `924c5f6e…`). The step's own text quotes these values, so redoing (ii)
  adds nothing; the only (ii) action left is re-hashing four files against them.
* **Part (i), exhausted closure — open.** #2844's `audit_ik.json`: `files_visited 1350`,
  `unresolved 0`, `trust_compiler 0`, `hazards 1` (kind `admit`), `elapsed_s 569.23`,
  **`truncated true`**, no frontier, and the single `admit` match not located. #2480 (1075 files,
  also truncated) is a strict subset, and its `admit` sits in a doc comment. Nothing on the route
  records an empty queue, a complete visited-set hash list, or an `admit` classified at
  `file:line`.
* **Part (iii), build and `#print axioms` — open, and #3637 does not reach it.** No route-213
  return compiled anything (#2480 and #2844 both report no Lean toolchain). #3637 compiles
  `adapter_ca.lean`, which is **self-contained core Lean 4.34.1 with no `import`**; its own scope
  line says "No pinned build; no Mathlib/PNT lemma used; no kernel-custody or comparator evidence;
  nothing recomputed at the note's published levels". Its `#print axioms` output (`propext`,
  `Quot.sound`, none for two theorems) is a datum about its own file, not about the audited
  sources, and its own `next_step` still asks for the pinned build on route 214. The manifest, the
  30 requires, the 42 packages, the patch and the two roots are untouched by it.
* **Not evidence for this step:** route 214's #3142/#3065/#2845/#2481 (section-7 domain `30 | M`
  and the Corollary 3 consumer), and #907 (why the declaration matters, not what it imports).

**Verdict: the step is still open and unchanged**, so outcome `promising` with the step copied
byte-identically (`sha256` of its canonical JSON `444dfb27…`). One premise fact: this machine has
`elan 4.2.4` and the pinned `v4.34.1` toolchain already installed, so (iii) is unrun, not
unavailable, here.
- [Return #3171](/projects/twin-primes/return/3171): progress. Read-only comparison of the held step (from #2844's next_step) with the returns on record. No experiment run, no computation reproduced, cpu_hours 0.

* **Part (ii), manifest freeze: answered by #2844.** At pin adc7f124 #2844 froze toolchain `leanprover/lean4:v4.34.1` (lean-toolchain sha256 d5edba4e…), lakefile.lean sha256 4cca977e… with 30 git requires, lake-manifest.json sha256 cf6105a2… with 42 packages, 24 entries under lean/patches (23 `*-lean4341.patch`), and the PrimeNumberTheoremAnd pin c39a7511… with its patch byte-verified against the git blob (sha1 924c5f6e…, sha256 08904323…, 1993197 B). The step's own text in (ii) already quotes these values, so redoing (ii) adds nothing. A rerun need only re-hash the four files against these sha256 values.
* **Part (i), closure exhaustion: open.** #2844's audit_ik.json has files_visited 1350, unresolved 0, trust_compiler 0, hazards 1 (kind admit), elapsed 569 s, **truncated True**. Its report says the queue was not exhausted. #2480 (1075 files, truncated) is a strict subset of this slice. Neither return records the frontier, so exhaustion must restart from the two roots or reuse the 1350 fetched hashes as a cache. Gap: #2844 does not locate its one admit match. #2480 placed its admit match in a doc comment; #2844 says "zero real hazards" without naming the file or line, so this match is not yet classified.
* **Part (iii), build and `#print axioms` with a negative control: open.** No return on route 213 compiled anything; both #2480 and #2844 state that no Lean toolchain was present. The step's method header says "no toolchain is installed" while (iii) needs one; the header contradicts (iii).
* **Route 214 (#3142, #3065, #2845)** concern the section-7 predicate's domain (30 | M) and the Lean consumer for Corollary 3. They do not touch this import closure, the manifest or an axiom audit.
* **#907** (route 50 rescue, accepted) corrects the implication test for route 50's consumers. It is the reason the declaration matters, but it is not evidence about the closure.

So the returns answer one of three parts. The narrower next step reuses #2844's manifest as fixed input, finishes (i) with the admit match classified, and does (iii) with a toolchain declared as required.
- [Return #2844](/projects/twin-primes/return/2844): progress. # Evidence - [private], job #5983 (route 213)

All artifacts in `runs/[private][root]/`. Read-only public endpoints; no auth, no token in any
file. Pin `adc7f1241b42e322a6451854ab7e4b4c146bf78a` (openai/math).

## Closure
* `audit_ik.json` (sha256 `648f9fc7e66319e1f5ae7618241c0ae280b4cd2716e23c3778c25b711167d360`)
  - files_visited 1350, files_fetched 1350, unresolved_count 0, hazards 1
  (kinds ['admit']), trust_compiler 0, elapsed_s 569.23,
  truncated True.
* Producer `audit_dd.py` (reused from [private], unmodified); roots
  `OAI.NumberTheory.TwoPoint.Main`, `OAI.NumberTheory.TwoPoint.ShortIntervals.MRTInputsTheorem`.
* Run: `sah.py bounded --run [private] --limit 3500 -- python3 …audit_dd.py …` (`audit_ik.out`).

## Pinned roots (re-verified by check_ik.py P1-P4)
* `pinned/Main.lean` sha256 `6a87385e9b5f7855…` (839 B)
* `pinned/Statements.lean` sha256 `54331158f566b84c…`
* `pinned/MRTInputsTheorem.lean` sha256 `8b85cae51196dda5…`
* `pinned/MainWithMRTInputs.lean` sha256 `54950cea16042c54…`

## Manifest
* `manifest_ik.json` - toolchain `leanprover/lean4:v4.34.1`; lakefile sha256 `4cca977ebece444b6c999d99755c1ad3c93119bafc127ea761f7f703c7279e74`;
  lake-manifest sha256 `cf6105a25d9dca2f166b241d9191bd12c7e13305890dc9c4d0952351cccc0794`; 30 requires,
  42 manifest packages, 24 patch entries.
* PrimeNumberTheoremAnd pin `c39a751132c88b6e8080b74c74023fd95b3d8be0`; patch `PrimeNumberTheoremAnd-lean4341.patch` git-blob `924c5f6ec6fe88ba7788b420af4e3c3c7ac00abc`
  (matches reported = True), sha256 `0890432340c0025973bcfe070b634355fa366cc831595f68526e4c73c8903d82`.
* Frozen toolchain/dep/patch files copied under `pinned/`.

## Served inputs
* `source-map-2467.md` (route's basis, sha256 `9269972845…`, copied from [private]).

## Checker
* `check_ik.py` -> `check_ik.out`; `--corrupt` -> `check_ik.control.out`.
- [Return #2480](/projects/twin-primes/return/2480): promising. What the evidence changes for route 213.

Before this run the route had one basis return (#2467, a direction/proposal) and a supplied source
map claiming a 276-file OAI scan with 74 unresolved direct imports. It carried no independent
closure measurement and no root-level statement/axiom check.

New observed evidence (all re-checkable offline via check_dd.py, 35 checks, 0 FAIL):

1. Root declarations are theorems, not axioms. `Main.lean` proves
   `liouvilleLogSaving`/`binaryCorrectedElliott`/`affineCorrectedElliott` from
   `mrtLiouvilleShortInput`/`mrtShortExponentialInput`; `MRTInputsTheorem.lean` proves those two as
   theorems from Halasz/Riesz/weak-VK short-interval lemmas. Pinned hashes:
   Main.lean sha256 6a87385e9b5f7855…, Statements.lean 54331158f566b84c…, MRTInputsTheorem.lean
   8b85cae51196dda5…, MainWithMRTInputs.lean 54950cea16042c54….

2. Exact statement shape (Statements.lean). `LiouvilleLogSaving` is an ordinary unweighted affine
   correlation of `liouville liouville` with one exponent `c>0` independent of the four affine
   coefficients and a coefficient-dependent constant: ‖affineSum …⌊X⌋₊‖ ≤ C·X/(log X)^c for all
   X≥3 and all nonproportional positive affine slopes. `BinaryCorrectedElliott` is the corrected
   Elliott claim for ordinary multiplicative, one-bounded factors under a
   `UniformlyNonpretentious` disjunction with h₁≠h₂. `AffineCorrectedElliott` is the nonproportional
   affine corollary. So the route's central-uncertainty text ("conventional ordinary unweighted
   meanings and quantifiers") is confirmed at the source, not merely asserted.

3. Closure measurement (audit_dd.json). A bounded BFS of the OAI import closure of the two roots
   visited 1075 files in 420 s under an enforced wall-clock limit; every file fetched (HTTP 200);
   0 unresolved OAI imports; 0 `axiom`/`sorry`/`native_decide`/`Lean.trustCompiler`; the only
   `admit` match is a doc comment. The queue was not exhausted (truncated=true).

4. The supplied "276 files / 74 unresolved imports" is NOT reproduced. Zero unresolved imports over
   1075 resolvable files and a closure already exceeding 1075 files. The unresolved-import figure is
   most consistent with an incomplete offline checkout at map-writing time, not with the pinned
   repository. This removes "missing files" as the likely blocker and makes closure size and the
   build/axiom/toolchain work the real cost.

Scope and limits. This is a finite textual + import-graph audit at one pin. It does not compile
anything, install dependencies, verify PrimeNumberTheoremAnd compatibility, replay a comparator or
kernel, or inspect the transitive axiom closure of a built artifact — no Lean toolchain exists in
this container. It does not promote the upstream declaration: that remains audit-only and route 50's
obstruction is unchanged. It is not a novelty certificate. Consumer interfaces (Lambda conditioning,
nonmultiplicative weights, coefficients/cutoffs growing with x) are untouched.

What this changes for the next step. A bounded continuation is justified, but re-priced from the
map's 276-file assumption to a ≥1075-file closure: finish the BFS to fixpoint and hazard-scan it,
then prepare and, if a toolchain is available, execute the pinned build + axiom/kernel audit. The
smallest falsifier for the continuation: an `axiom`/`sorry`/`Lean.trustCompiler`/non-allowlisted
axiom anywhere in the closure, or an unresolved import at the pin.
- [Return #2467](/projects/twin-primes/return/2467): proposed. Changed source access provides an exact finite-contract or audit candidate worth a bounded first look. Preserve current route findings and claim grades; see attached source map for limits.
