TruVector

A checkpoint between interpretation and action

Language models interpret meaning. Coordinates check subject relevance. Arithmetic combines recorded readings and source weights.

Input and structured readings

Record the statement, retrieved passages, source assignments, and the authorized instruction and proposed action. Readers return support, refutation, or uncertainty with confidence and a reason. Separate action readings return yes, no, or partly.

Numerical aggregation

Origin-group contributions are divided into per-reply weights, then combined with confidence correction and subject factors. Normalized support S, refutation R, and uncertainty U remain visible. D = S − R measures direction: how far the readings lean toward support or refutation. C measures certainty: how concentrated the readings are on one answer, which is not the same as how often that answer is right.

Reference decision policy

In words: if too few readers answered, stop and say why. If the readers say the action does not do what was asked, or they refute the statement with confidence, stop. Go ahead only when the readers support the statement with confidence, the support comes from at least two separate origins, the readers’ reasons are about the statement, no reader is asserting without support in the material, the action is about what was asked, and the readers say it does what was asked. Everything else goes to a person.

In symbols: fewer than K usable stance readings produces Block with a quorum reason. With quorum, A_no ≥ 0.5 or D ≤ −0.5 with C ≥ 0.5 produces Block. Allow requires D ≥ 0.5, C ≥ 0.5, effective origin weight ≥ 2, reason subject score ≥ 0.6, no unsupported-reader flag, action subject score ≥ 0.7, and A_yes ≥ 0.5. Other outcomes are Review. The full rule is WP-02 §7.

The proposed action-input policy additionally requires K usable action readings for Allow. Zero-weight sets with stance quorum route to Review. WP-02 identifies these policy completions and their implementation-conformance requirements.

Reconstructable output

The record contains model and prompt versions, raw readings, source assignments, dependence parameters, weights, S/R/U, D/C, subject scores, action readings, flags, quorum, thresholds, and the decision reasons. It distinguishes substantive opposition from unavailable inputs.

Workflow integration

Retrieval weighting informs passage selection or annotation. Reader aggregation preserves dissent and uncertainty. Action assessment checks instruction consistency alongside application permissions. Origin recovery, delivered context, confidence calibration, and adversarial robustness are evaluated separately.