Thresholds, Not Timetables: A Bounded-Snapshot Protocol and Pre-Registered Predictions for Restoring Oversight of Open-Weight AI

Thresholds, Not Timetables is the fourth paper in the Knowledge Thermodynamics Recursive Planning Science (KTRPS) series and the first designed to be scored against future evidence. It specifies a bounded, dated, reproducible sampling protocol for a single-date public-record snapshot of the divergence of the AI platform landscape into a closed, provider-gated stack and an open-weight, self-hostable stack. A pre-registered bounding analysis covers deployments the public record cannot show. The protocol is demonstrated on an illustrative candidate snapshot and coded under the KTRPS commitment register and five-dimension correction-capacity instrument. The paper performs an ex ante threshold estimation for two pre-specified scenarios of restoring oversight of deployed open-weight systems in the European Union over a 36-month horizon: artefact-level oversight (Scenario S) and oversight through compute and hosting infrastructure (Scenario Sβ). Seven boundary-conditioned predictions are pre-registered with comparators, scoring rules, statuses and probabilities. They concern thresholds, feasibility boundaries, binding constraints and comparative predictive value, not calendar dates. The comparative claim (Prediction 1) is scored by a ridge-penalized out-of-sample comparison and a Bayesian model comparison, with decision thresholds on both sides and a pre-deposited analysis of the design’s operating characteristics. The paper reports no outcome data. The candidate snapshot contributes no data to scoring; predictions are scored against the formal snapshot collected on the registration date. Keywords: KTRPS; open-weight models; AI governance; human oversight; pre-registration; predictive evidence; threshold proximity; correction capacity; commitment register; bounded snapshot protocol; Bayesian model comparison; selection-bias bounding Additional descriptions Methods: Pre-registered, single-date public-record snapshot. Source pages are archived and SHA-256 hashed. Ridge-penalized cross-validation and Bayes factors under a default g-prior mixture at three registered scales. Notes: Companion white paper: The Institutional Mechanics of Open-Weight AI Oversight. Where the two differ, this protocol governs.