The science of how information moves
Direction, change over time, and source dependence provide a richer view of information than volume alone.
In plain terms
Several AI models read a statement and say what it means. Arithmetic then counts how many separate sources are really behind it and decides. Nothing in the decision is a model’s feeling; every number can be recomputed by hand.
A research framework for information assessment
Lane Vector studies signals relative to a defined outcome, their rate and acceleration, and the dependence among their origins. TruVector applies that framework before an AI action. Both are part of InTellMe AI.
The research uses explicit statistical definitions and an engineering approach to measurement. Outcome prediction is evaluated against observed results; geometrical projection is exact under a declared linear model.
Meaning and calculation have distinct roles
Readers classify statements as supported, refuted, or unsure and separately assess instruction consistency. Embeddings check subject relevance. An independent numerical layer aggregates the readings, discounts shared origins, and evaluates Allow, Review, or Block conditions.
A sentence and its negation can have high embedding similarity. Subject relevance therefore remains separate from agreement and action authorization.
Source structure changes the contribution
For a known origin group of m observations, n_eff = m / (1 + λ(m − 1)). For positive λ, its contribution approaches 1/λ within that group. Recovering origins and estimating dependence are measured separately.
Five known groups of 100 at λ = 0.5 contribute about 9.9 effective weight units. Five hundred singleton groups contribute 500. These are analytical scenarios with declared source structures.
A first comparison
On a first set of sixty statements, read by four models, the reading-based decision was right far more often than the word-count gate it replaces. The set was an easy one, drawn from public records, and it tested statements only, not actions. The research page gives the numbers with their limits, and the harder sets that come next.
Research and applications
Projection is tested against volume alone. Acceleration is tested for earlier alerts at matched false-alarm rates. Reader dependence, calibration, subject guards, and decision thresholds are evaluated against explicit baselines.
Applications include retrieval, AI action assessment, search optimization, source analysis, and communication monitoring. Explore LaneVector SEO and other business applications.
Explore the calculations
Interactive calculations use declared inputs. The reading panel isolates aggregation; complete action decisions require the additional checks in WP-02.
Where the arithmetic goes inside a system that acts
A system that acts on text reads, agrees, and does something. Each is a place the same arithmetic runs before the next step is taken. These three panels are simulations; each states the observation that would prove it wrong.
The reading step, before and after
The models read, the arithmetic judges. Every “how alike” and “does it agree” number is computed on what the models read, never on the words they used. Counting stays outside the readers, where the inputs, weights, and thresholds can be inspected, and the decision stays a formula a person can recompute by hand.