DMB Frontier Leg: Dual Mesh Benchmark Raw-API Columns for Claude Opus 5.5 and GPT-6.1 Sol on a Frozen 20-Instance Corpus — 20/20 + 20/20 Decision-Tuple Correctness (Preregistered, Scoped)

Abstract

SQUIRL OS RESEARCH DISCLOSURE — READ FIRST: AI-assisted, receipt-first research program (open-book doctrine: audit-and-correction, versioned amendments, append-only public ledger). Errors are disclosed, not erased. This software is a prototype and is provided for educational and research purposes only. Not intended for production use, commercial deployment, or safety-critical environments. All payloads synthetic. DMB Frontier Leg — the Dual Mesh Benchmark gains its two frontier engine columns, completing engine coverage from flash-class to frontier on the frozen B1 corpus. Two engines never before measured raw-API on this corpus — Claude Opus 5.5 (anthropic/claude-opus-5.5) and GPT-6.1 Sol (openai/gpt-6.1-sol) — each processed the frozen 20-instance selection (CORPUS_B1, sha256 13295d19...fee59, first-20 rule, instrument byte-identical to the DMB-X runner), one fresh stateless OpenRouter request per instance, model_returned recorded per call. Result: both columns 20/20 correct on decision-tuple correctness, 20/20 strict-valid (zero normalization needed) — Opus 5.5 median 3346.2 ms (p95 3983.0 ms; one 8688.9 ms outlier disclosed), GPT-6.1 Sol median 2225.6 ms (p95 2916.5 ms). Deterministic mesh in-memory lookup reference: 0.000443 ms. That is the dual-mesh doctrine in one number pair: engines propose correctly at seconds-scale; the mesh adjudicates at microseconds. Falsifiers (all clean): FBF-1 raw logs preserved (40/40 responses verbatim in the bundle); FBF-2 corpus byte-identical, sha re-verified post-run; FBF-3 zero retries or omissions; FBF-4 no post-execution metric additions; FBF-5 synthetic corpus only, no PHI/PII/credentials; FBF-6 claims scoped to the two tested columns on the raw-API substrate. Spend $0.1712, no payment abort. Provenance disclosure (audit-and-correction): the preregistration document (sha256 ebaa3aab...ca43f5e) was authored and timestamped in the working workspace BEFORE execution; the instrument runner was committed publicly before execution (commit d802b83); the prereg document itself enters the public repository tree with this publication (forward-only addition, commit 87e855b). No metric was added post-execution. Scope caveat (load-bearing): sampled-baseline measurement of two engines via OpenRouter on the frozen 20-instance corpus, October 2026. Not an engine certification, not a universal performance claim, not a cross-substrate wall-clock equality claim (XPLAT-2/3 doctrine: raw-API legs comparable within their substrate only). Dual Mesh Benchmark status: MEASURED (scoped) — matrix v2.0 DOI 10.5281/zenodo.23321637; DeepSeek column 10.5281/zenodo.23345241; cross-engine expansion 10.5281/zenodo.23348675. Closes the three-process receipt trilogy with DOORS (action gates, DOIs 10.5281/zenodo.23607572 + 10.5281/zenodo.23608157) and SBR (bounded-learning mesh, SBR-1 DOI 10.5281/zenodo.23331224). Provenance: ledger entries 0322-0324, github.com/LLong2026/four-horsemen. DCAI community: zenodo.org/communities/dcai. Two-key curation: drafted and vetted by the system (Gabriel), published on the owner's explicit GO (chat authorization, 2026-10-23 23:04 CT).