Sequence analysis: validation scope and provenance

The September 2026 deep-search briefings did not identify a new scanpath similarity, HMM, semi-Markov or transition estimator whose validation justifies displacing the existing gp3sequencespy scientific surface. Accordingly, no speculative estimator is being introduced in this development tranche.

Boundaries of existing sequence inference

Sequences belong to participants, trials, sessions and sometimes datasets. A transition table derived from adjacent gaze events can share observations, stimuli and study context across train/test folds even when its rows differ.

Before machine-learning evaluation, externally certify:

• temporal-block separation, trial separation, session separation and participant separation as different properties;
• participant and trial identity in the canonical sequence table;
• event-detector and AOI mapping provenance;
• order, duplicated state positions, undefined states and exposure denominators;
• genuine within-person pairing versus unpaired cross-source data.

The GazeForge validation-scope certificates can independently audit sample/trial/participant/dataset split identities; its research programme extends session and temporal-block checks. Those validation tools do not automatically certify every sequence model.

Appropriate reporting

Use participant-generalization only when no participant ID appears in both training and test groups. Block-wise disjointness alone does not establish unseen-person inference. Separate transition probabilities from causal claims about attentional mechanisms; analyses of event-derived states should report detector and AOI sensitivity where material.

This page is documentation-only. It does not change the public sequence API, installed version, benchmark status or release readiness.