Riemann project · construction audit · 29 September 2026 · v1

The local swarm works.
The mathematical bridge is still missing.

We now have a working local test runner and a bounded, two-model proposal-review pipeline. Three research agents also developed explicit geometric construction briefs. None of the local-model proposals qualifies for an RH-directed calculation.

Wall check: Same open gap. Closest records: OPERATOR-BRIDGES and the scoped NS47/94/100/101 controls. The missing global constructions and arithmetic transfers remain missing. No RH or G2 result is claimed.

What actually ran

Work	Result
Exact local construction checks	165 deterministic fixtures reproduce the intended distinctions. These are repeated calibration cases, not 165 new ideas.
Qwen 27B, existing local service	Three proposals and three critiques returned.
Llama 3.1 8B, installed locally	Three audits of the first model’s critiques returned.
Candidate admission	Zero. Every proposal still contains a decisive construction or transfer defect.

The pilot used existing services, one request at a time, with explicit output and timeout limits. No model was installed, no generated code was executed, and no further batch is running. The exact checks run on the CPU; GPU inference supplied the model responses.

An initial short-hostname configuration failed address validation before inference. We retained that guard and used an independently verified private address. Public evidence omits private endpoints. The nine successful calls reported 17,791 total tokens, including 5,203 completion tokens; those are service-reported counts, not a performance comparison.

Three geometric construction briefs

Contact geometry in 4D, 6D and 8D. An explicit four-dimensional block has a central orbit of period log p. Its signed local trace coefficient matches the prime-power weight. Higher dimensions require adjusted expansion rates and grading to retain that coefficient. Composite labels pass the same algebra, so this alone does not identify the primes. A connected global construction, full trace theorem, gamma and pole terms, and positivity transfer are absent. Definitions and review.

Arithmetic surfaces. Products of elliptic curves give four real dimensions, but the proposed prime-power deck maps become identities after quotienting. A separate family of adelically metrized line bundles on P¹×P¹ retains prime valuations in its local metrics. The missing step is a complete map from the actual Weil tests to these metrics with the required controlled remainder. A Hodge sign theorem does not supply that map. Definitions and sources.

Spectral geometry. A four-dimensional oscillator produces a two-prime partition function; its energy levels are logarithms of integers, not zeta zeros. A completed canonical-system construction belongs to an existing program. The simplest independent passive-prime version fails: one factor has modulus 7/6 where contractivity would require at most 1. A coupled completion is still undefined. Construction limits.

Why a larger swarm would currently waste effort

All three local-model proposals reused the same supplied contact-block example despite different assignments. They confused geometric dimensions, local orbit weights, spectral eigenvalues and the global positivity target. The critics caught some defects but introduced others. For example, infinite measure does not prevent a spectral measure, and an explicitly conditional argument is not automatically circular. The second model often repeated the first critic’s mistakes.

We strengthened the lane-specific briefs and required critics to identify the exact assertion they dispute. Those revised prompts have not yet been tested in another inference batch. The audit of all nine replies records the corrections. Model agreement will not count as proof or automatic candidate admission.

Where I would work next

I would spend the next round defining one complete construction bridge, then use the local models to challenge its individual obligations. In the contact lane, the concrete assembly question is how to join the prime blocks without an uncancelled mixed primitive contribution at log 6. In the metric lane, it is how the proposed intersection pairing reproduces the complete signed Weil form. These questions must be made precise before another RH-directed scan.

A successful local construction test would narrow its own assembly problem. It would not establish the remaining global trace or arithmetic estimate. The existing freezes therefore remain in place.

Evidence and limits

The exact checker’s first run is preserved alongside a corrected second run. Two implementation findings were repaired: integer primality bounds and explicit expected outcomes for support checks. Twenty-four software tests pass; Python diagnostics report zero errors and warnings. The standard manuscript build reports its unchanged targets up to date at 358 pages with zero undefined or duplicate references.

All 104 registered conclusion records, 14 continuation-input demands and freeze conditions are preserved. Both historical evidence-recovery groups remain open. This was a conclusions/dependency review and bounded agent cross-review, not a revalidation of every historical proof. No manuscript version, research row, thaw or new arithmetic estimate was added.