DOORS-2: The Clustered Door — Two-Node Fail-Closed Adjudication of a Deterministic LLM Action Gate with a Shared Authoritative Approval Registry

DOORS-2 "THE CLUSTERED DOOR": a two-node clustered adjudication battery for a deterministic LLM action gate sharing one authoritative human-approval registry. TESTED CONDITIONS: two independent validator nodes with separate local journals and separate cache views, one hash-frozen authoritative registry (forward-only append disclosed before the disagreement scenario), synthetic offline target only — zero production writes, zero credentials, zero network reach. The worst-case blast radius of any failure is one synthetic file gaining a newline. SCENARIOS (pre-stated in a preregistration committed before execution, sha 5bc5e188): S1 baseline valid action (admitted, exactly one execution); S2 forged authority — approval reference absent from the registry (rejected by both nodes); S3 cache-skew disagreement — node A's view contains a newly granted approval, node B's view is stale (node A ADMIT, node B VIEW_MISS; cluster HELD fail-closed, zero mutation, escalation recorded); S4 revocation propagation — approval revoked in the registry, stale-approved in one node's cache (revocation enforced at both nodes); S5 cross-node duplicate — identical action-id submitted to both nodes (exactly one execution via the shared cluster ledger). HONEST EXECUTION RECORD: the first execution (v1.0) is preserved byte-exact and FAIL — the cache-skew scenario setup was inverted (approval placed in the view but not the registry, so no disagreement path was exercised) and the F-2 falsifier expression was unsound. Amendments A1-A3 were issued before re-execution, per the freeze-integrity pattern. The v1.1 repair run: 5/5 scenarios PASS, 4/4 pre-stated falsifiers clean. WHAT THIS ESTABLISHES (scoped): under the tested conditions, a two-node cluster sharing one authoritative registry resolves cache-skew disagreement fail-closed (hold, not execute), enforces revocation over stale caches, and suppresses cross-node duplicate execution to exactly-once. NOT established: multi-node (>2) scaling, network partition recovery, asynchronous propagation timing, production tool surfaces, or universal containment. No independent replication is claimed; the harness, prereg, and results are published for exactly that invitation. Evidence bundle: preregistration, runlog (with disclosed v1.0 FAIL and repair arc), receipt with SHA-256 manifest, frozen harness source, both result files, and the synthetic target. 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. Dual Mesh Benchmark: NOT MEASURED. Method operates under filed provisional patents 64/114,746, 64/119,191, 19/693,343 (USE-rights doctrine; no unfiled finality disclosure).