Directional Transition Model Calibration Note v0.1 A Two-Model Public Canary Screening Pilot for Norm-Dynamic Transition Edges

Directional Transition Model Calibration Note v0.1: A Two-Model Public Canary Screening Pilot

This calibration note reports a small two-model screening pilot for the Directional Transition Model (DTM) and its associated Public Canary Benchmark. The purpose of the note is not to validate the full transition model, rank models, or infer hidden internal states. Instead, it documents an early calibration checkpoint: whether small user-friendly canary cases can distinguish between different visible transition patterns in model outputs.

The study compares two publicly accessible model runs across a calibrated set of DTM canary cases. The cases probe whether bounded claims remain stable, become strengthened, are requalified, are carried over into nearby contexts, propagate into family-similar domains, or recover after explicit correction. A neutral contamination sentinel is included to test whether the benchmark induces over-attribution or theoretical leakage into an unrelated planning task.

The main finding is a partial recurring profile rather than a full validation. Same-style continuation, explicit recovery, and family-similar propagation appear as the most stable signals across the two runs. Surrogate legitimation and weak-handoff carryover become clearer after calibration, while delay erosion, relay requalification, and near-frame propagation remain fragile or prompt-sensitive. The contamination sentinel remains clean.

The accompanying supplement provides the supporting material for inspection and reproduction, including raw outputs, scoring summaries, calibration files, and reproducibility notes. The artifact is intended as a methodological waypoint — a small calibration buoy for subsequent work on directional transition dynamics, norm preservation, bounded claim handling, and Anti-Goodhart evaluation of language-model behavior.

DOI: 10.5281/zenodo.19810178