Dual Mesh Benchmark: Preregistered Latency and Integrity Measurements for Deterministic Anomaly-to-Playbook Resolution (Public Edition v1.0)

Abstract

Preregistered dual-mesh benchmark paper reporting three measurements of the Squirrel OS deterministic anomaly-to-playbook resolution layer: B1 lookup latency — in-process mesh resolver 200/200 correct, median 0.000443 ms, versus an LLM baseline (Automatic engine, fresh platform conversation per instance, wall-clock request-to-validated-tuple) 20/20 valid, 0 failures, median 4509.1 ms; measured ratio median/median 1.018e+07. B2 mesh capacity & collision integrity — 50 synthetic nodes ingested, collision hashes 50/50 unique, state integrity clean. B3 resolution correctness — 72/72 exact replay of the CCR-1 permitted healing-action journal, 0 false positives on negative probes. Binding disclosure: the B1 ratio characterizes the tested environment — deterministic in-process exact-match lookup versus this platform's API conversation loop (agent loop, inference, transport included). It is not a universal performance claim and neither confirms nor supersedes the historical production observation of 267x / 800ms->3ms (September 2026), which remains a historical observation of a different environment. Not model certification; not a production claim. Preregistration discipline: protocol and seeded corpus builder committed to the public append-only ledger before any measurement (corpus sha256 13295d19..., commit 2eb3879); metrics and falsifiers fixed at prereg; identical frozen corpus on both B1 legs; all 220 runs recorded, none removed; falsifiers FB-1..FB-5 all PASS. Synthetic data only: every benchmark record is simulated corpus; no production, customer, or personal data are involved; no PHI, credentials, wallet identifiers, or transaction amounts appear in any artifact. Patent-safe statement: architecture content is covered by filed provisional applications (64/157,915; 64/119,191; 64/114,746; 64/145,825). No unfiled method material is disclosed; the DCAI-EVT taxonomy and five-step finality method remain drawer-bound. AI assistance: benchmark design, execution tooling, and drafting were AI-assisted under human direction; all numbers are receipts anchored to the public append-only ledger; two-key curation: AI vet, human GO (October 11, 2026). Research home: DCAI community; public append-only ledger: four-horsemen (commit de0e68d). This software is a prototype and is provided for educational and research purposes only. It is not intended for production use, commercial deployment, or safety-critical environments. All systems are experimental and may contain defects on them.