Neuro-symbolic runtime monitors for drug discovery LLM agents (Mozi framework) will reduce unconstrained tool-use violations by >40% when enforced via cryptographically verifiable authorization (e.g., Datalog policies with ZK-proofs), as measured by compliance failure rates in high-stakes molecular docking workflows.

An LLM's reading of the literature — not computational verification.

VOID BY THE PRE-REGISTERED DENOMINATOR, NOT 'CONSISTENT'. Zero discordant pairs at every threshold, and both controls behaved -- a blocked row's hypothesis planted as a pass on the same script was detected, and the same hypothesis on a different script was correctly not counted, so the pairing code works. But the zero is STRUCTURAL rather than evidence of consistency: of 261 blocks, 209 are script-attributed, while of 28 passes, ZERO are. One side of the comparison is empty, so no discordant pair could exist at any threshold. The pre-registration said a zero would mean 'the monitor is at least self-consistent on this corpus' -- that attribution is wrong and is withdrawn. An absent comparison is not a negative result.

Method: Pre-registered self-consistency audit, 27 Sep 2026, pure local computation over the 2,585-row archive; no model calls. Discordant pairs = same routed script, hypotheses >=0.82 similar, one blocked and one passed; swept at 0.72 / 0.82 / 0.92 with a planted positive control and a cross-script negative control. analysis/policy_monitor/ · Result: inconclusive