OTIP-for-LLMs Pilot A Preregistered Specification for Commitment Integrity Under Rule Change and Proxy Pressure

This document presents a preregistered pilot specification for OTIP-for-LLMs, a diagnostic evaluation protocol designed to test whether large language models can maintain or correctly revise explicit commitment artifacts under rule change and proxy pressure. The protocol is positioned as a diagnostic overlay for LLM evaluation, not as a benchmark of hidden cognition, consciousness, or “true understanding.” Its purpose is to determine whether post-shift failures can be decomposed into competence failure, update failure, and integrity failure, and whether this decomposition provides diagnostic value beyond standard accuracy-only and faithfulness-only evaluation. The preregistration locks the pilot design before empirical evaluation, including the Tax Code Sandbox v0.1 microworld, the machine-readable Commitment Object schema, experimental phases, baseline suite, scoring rules, statistical plan, logging policy, and explicit falsification criteria. The protocol also includes safeguards against new Goodhart effects, such as calibrated abstention, strict separation between frozen and revised commitments, and algorithmic scoring without manual repair. This release should be read as a methodological document. It does not report benchmark results. Its contribution is to define a falsifiable and reproducible pilot framework for testing commitment integrity under intervention in LLM systems.