Dual Mesh Benchmark DeepSeek Column: 20/20 Conforming Resolutions, 8.5 s Median, by a Previously Non-Conforming Engine After Preregistered Repair

Abstract

Preregistered benchmark leg, not a certification. DeepSeek v4-pro (via OpenRouter, model_returned verified per call) resolved the first 20 instances of the frozen CORPUS_B1.json synthetic anomaly corpus (sha256 13295d19...), the identical selection rule as every prior B1 baseline leg, one stateless request per instance, wall-clocked request-to-response. Results: 20/20 strict-valid conforming tuples (zero normalization needed), 20/20 correct against expected playbooks including exact platform playbook_ids, median 8,538.0 ms, p95 11,864.0 ms, total API cost $0.0089. Comparisons are directional only (XPLAT-2/3 doctrine: raw external API is not the conversation-loop substrate): the v1.0 mesh lookup median on the same corpus is 0.000443 ms; the v1.0 Automatic conversation-loop baseline median is 4,509.1 ms. The raw-API DeepSeek median exceeds the full-loop figure, attributable to reasoning-token generation rather than platform orchestration. Accuracy-per-dollar is excellent; latency under reasoning defaults is not this engine's lane. No wall-clock equality claim across substrates is made. Honest-results note: this column completes a documented repair. The Oct 11 CCR-4 DeepSeek capture was NON-CONFORMING (fabricated tuple, wrong array shape) and remains on the public tally. The preregistered containment re-run (CCR-4X leg I: conforming 9/9, forged APPR-9923 authority rejected) and this benchmark leg followed only after that loss was posted publicly (publish-first repair doctrine). Limitations: sampled baseline (n=20), single engine and route, October 2026; synthetic corpus only; providers varied by call (StreamLake, Parasail) per OpenRouter routing; no stress/load or live-healing testing; not a model certification or universal performance claim. Dual Mesh Benchmark: this record is itself a scoped DMB measurement; all other DMB claims remain NOT MEASURED here. Bundle contains the receipt (Markdown), machine-readable receipt JSON, full evidence archive (prereg, runner, results, all 20 raw responses and extracted contents, SHA-256 manifest). Registration and evidence: public four-horsemen ledger (prereg commit aa751dd before execution; receipt commit 9bc6851). Research home: DCAI; portal: DSOS Science. Lineage acknowledged: Nyquist, Hartley, Shannon, Kolmogorov. No new mechanism is claimed. 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. Version note (v2): this version supersedes v1 solely to redact verbatim operator chat text from the preregistration per the operator-privacy directive (Oct 13, 2026). Authorization is recorded as GO + timestamp only. No measurement, result, or claim changed.