SDR-CL2: Third-Vendor Model-Attribution Replay of a Registered Substrate Determinism Gauntlet Across LLMs (Microsoft 365 Copilot)

Abstract

SQUIRL OS RESEARCH DISCLOSURE — READ FIRST: This work was produced inside an AI-assisted, receipt-first research program (open-book doctrine: audit-and-correction, versioned amendments, append-only public ledger). Every claim is backed by a timestamped, publicly committed receipt trail in the four-horsemen repository. Errors are disclosed, not erased. Limitations are load-bearing. 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. All gauntlet data is synthetic. On Oct 7 2026, substrate D (SIM-D, a zero-lineage standard Squirrel OS v1.1 box on PQC rails) was certified under a preregistered determinism gauntlet (SDR-D): a public harness, a frozen seed, and a committed checksum prediction published BEFORE the run, with six registered falsifiers. The registered gauntlet has now been replayed across three vendor ecosystems: Gemini 3.1 Pro (original certification), Claude Sonnet 5.5 (SDR-CL1, DOI 10.5281/zenodo.23274887), and Microsoft 365 Copilot (SDR-CL2, this receipt). Result: CONVERGED — three vendors, one cage, identical receipts. The deterministic phases converged 13/13 on the registered checksum 9337c4fc…d2df3e (public harness, unmodified). The Copilot decision layer produced byte-identical decision tuples across two identical fresh-chat submissions (tuple hash 51c08358…2303 both runs), rejected a deliberate near-miss roster entry 6/6 (exact-match rule only), applied the lowest-record-id tie-break rule-exact 6/6, and converged 6/6 with the Claude replay tuples. What this receipt establishes: the deterministic envelope, not the language model, owns the output. Swap the engine; keep the receipt. Scoped evidence for THIS registered gauntlet on THIS substrate under THESE capture conditions — not a determinism claim for arbitrary LLM behavior outside the governed boundary. Disclosed deviations (audit-and-correction doctrine): Copilot was not model-pinned (backend model not client-disclosed; attribution by Microsoft 365 client build 154.0.4258.62 and verbatim operator capture); decision path used the hub canonical roster (exact-match, lowest-record-id tie-break) as in SDR-CL1; a citation-level defect in the SDR-CL1 phase-b file (2 of 6 playbook_ids inconsistent with its own disclosed tie-break rule; names and tuples rule-consistent) was found during this run's audit and is disclosed here with the published CL1 artifact left unamended; first submission attempt was a harness-delivery miss (logged, not a caging result); supervised run with human-operated capture; supplementary run tracked separately from the 60-night battery. Provenance: public harness + registered prediction: github.com/LLong2026/squirrel-os-lindy-test (docs/sim-d/). Battery ledger: github.com/LLong2026/four-horsemen. Method covered by filed provisional 64/119,191 and PQC/DQCO filings; governance per DCAI (64/157,915). No new mechanism claimed. Information-theory lineage acknowledged: Nyquist (1924), Hartley (1928), Shannon (1948), Kolmogorov (1965). Associated receipts: SDR-CL1 DOI 10.5281/zenodo.23274887; First LLM Cage DOI 10.5281/zenodo.23265723; caging architecture paper DOI 10.5281/zenodo.23266727. DCAI community: zenodo.org/communities/dcai. Two-key curation: drafted and vetted by the system (Gabriel), published on the owner's explicit GO.