Observational Convergence Across Independent Diagnostic Frameworks: Three Coincidences, the Nulls They Survive, and the Boundary They Do Not Cross

Four independent AI-diagnostic frameworks (TAT-7, Cophy, HeartFlow, U/D/A/H) were run on the same test document and converged on three things: a detection threshold of 0.3 within 0.02 across three frameworks, a correlation of r = 0.985 between two frameworks' drift trajectories over 31 ordered steps, and five flip points at identical positions across four frameworks. This report states each convergence against its own null, and is explicit about what they do not establish. The three results are deliberately not combined into a joint probability: they are computed on one shared dataset, so they are not independent events and their product bounds nothing. The correlation null remains conditional on a drift estimated from the trajectories themselves, which is marked as pending rather than claimed. And the sample for the field-health claim is one document, so 31 steps are 31 measurements of a single object rather than 31 tests of the hypothesis; separating "field health" from "shared text" requires an experiment that varies the stimulus while holding the instruments fixed. What survives is the observation itself, which is substantial: a driftless null gives P(|r| ≥ 0.985) of about one in a million at n = 31, so the correlation is not an artefact of autocorrelation alone. The paper also records its own open questions rather than resolving them, including the measurement-model assumption of independent per-framework noise, which four readers of one document do not obviously satisfy. Version 1. The drift-conditional analysis, once run on the real trajectories rather than on a synthetic reconstruction, is intended for a later version.