TruVector

The science behind the checkpoint

Lane Vector studies direction, change over time, and source dependence. TruVector applies that framework through structured readings and an explicit decision policy.

Interpretation before aggregation

Readers classify support, refutation, or uncertainty and separately interpret the proposed action. Embeddings check subject relevance. A sentence and its negation can remain close in an embedding space; a high cosine therefore does not establish agreement or authorize an action.

Dependence-aware weighting

For m observations in a known origin group, n_eff = m / (1 + λ(m − 1)). For positive λ the contribution approaches 1/λ within that group. Document and reader origins are tracked separately; origin recovery and dependence estimation are distinct evaluation tasks.

Signal dynamics

Directional projection is exact under a declared linear conditional-mean model. Finite differences describe rate and acceleration with explicit sampling-noise propagation. Forecast benefits are tested against volume baselines on future observations at matched false-alarm rates.

Reported study results

The preliminary balanced 60-statement comparison reported 14 correct decisions for the earlier word-based gate and 58 for the replacement under the original labels. Two labels are under independent review of their exact wording; the rule adopted is that a statement is labelled as written, not by the subject it is about. The study evaluates statements, not actions, and does not estimate deployment accuracy.

The five-model embedding comparison motivates the subject guard. Item-level reproduction and independently labeled holdout comparisons establish generalization. The full framework is at Lane Vector research.