Your AI is about to act on something it read. TruVector checks it first.

The problem

AI systems now act on information nobody checked.

Readers interpret meaning. Arithmetic evaluates the decision.

Structured support, refutation, and uncertainty are combined with source weights. Embeddings guard subject relevance; separate readings assess instruction consistency. The reported 60-statement comparison produced 58 correct decisions versus 14 for the earlier gate under the original labels. This preliminary statement-only study is documented with its scope in the research papers.

The arithmetic

Move the numbers yourself.

This is how text becomes numbers and numbers become a decision. Four steps, then the three places inside an autonomous system where the arithmetic runs. Steps 01 to 04 each work a formula from the Lane Vector working papers; 05 to 07 show that arithmetic at work inside a system that acts. Drag a control and the result follows.

01Reading

Mapping Words to Coordinates

A computer cannot do math on words. So every word, sentence or paragraph is turned into a list of numbers: a coordinate in a space with hundreds of dimensions. Two of those dimensions are drawn here. Meaning becomes distance. Words that mean similar things land close together, and a whole sentence becomes one arrow to one point.

At the starting sentence, “The queen addressed the court from the throne,” the whole sentence becomes one point and lands beside king, queen and throne. The positions on this map are schematic: they show how an embedding arranges meaning, not the numbers of one named model.

text → x ∈ ℝᵈ An embedding model turns text into d numbers, usually 384 to 1,536 of them. Two are drawn here. Which model and which basis must be stated before any angle means anything. WP-01 §2.

02Direction

Vector Cosine Similarity

Once text is an arrow, two arrows can be compared by the angle between them. cos θ runs from 1, the same direction, through 0, unrelated, to −1, the opposite direction. Volume says how much text there is; the angle says how much of it points at the question. One limit to keep in mind: between two pieces of text, this angle only says whether they are about the same subject. Whether one supports or refutes the other comes from a reading, not from the angle.

At the starting angle of 60° and a volume of 10,000, cos θ is 0.50, so 5,000 of the 10,000 counts toward the question. A plain count would report all of it.

cos θ = (A · B) / (‖A‖ ‖B‖) and the expected contribution is V ‖w‖ cos θ with ‖w‖ = 1. Exact when the outcome is a straight-line function of the signal. WP-01 §2.

03Counting

Astroturf Detector Simulation

An AI agent searches a topic and finds 500 accounts all saying the same thing. Each dot is one account, and it looks like a crowd of ordinary people. Switch on Astroturf Mode to see the same crowd when it is fake: 500 accounts that look separate but are run by 5 operators. λ is how alike the accounts run by one operator are, estimated from three things: how alike the wording is, how synchronized the timing is, and how much the accounts’ networks overlap.

A real crowd of 500 people who each speak for themselves counts as 500 sources. In Astroturf Mode at the starting coordination level of 0.50, the same 500 accounts run by 5 operators count as about 9.9 sources, and adding more fake accounts would not help: each operator tops out at 2.0.

n_eff = m / (1 + λ(m − 1)) worked out for each origin, then added up. λ is a declared illustration parameter here and is estimated in a study. A correctly identified group approaches 1/λ for positive λ. WP-02 §3.

04Timing

Acceleration Signal Test

Is a story really speeding up, or is that just noise? Acceleration is the change in the change of a count, and the noise in that estimate is known before any data arrives. Check too often and the noise drowns the answer. Check less often and the estimate settles onto the true line.

At the starting settings (checking every 2 samples, noise of 20 per check, a true acceleration of 2) the noise in the estimate is about 12.2 and the signal-to-noise ratio about 0.16: below 1, so the estimate is mostly noise. Widen the interval and it settles onto the true line.

sd(a) = √6 · σ / h² for independent noise. Correlated noise requires covariance-aware propagation; interval-dependent covariance can change the scaling. WP-01 §3.

Where it goes inside an autonomous system.

An AI agent reads, agrees, and acts. Each is a place the same arithmetic runs before the next step is taken. These three panels are analytical illustrations; the research directions below say what would prove each one wrong.

The reading step, before and after.

The research path from here: the models read, the arithmetic judges. Reader labels carry support and refutation. Embedding comparisons guard subject relevance. Source weights and thresholds determine the aggregation policy. Counting stays outside the readers, where its inputs, weights, and thresholds can be inspected, and the decision stays a formula a person can recompute by hand.

Where the research has to go next.

Each direction is stated with the result that would prove it wrong. The first three are about measuring meaning, the crucial parameter; the rest are about counting it.

• Which model assigns the coordinates, and what the angle can carry Every number in step 01 depends on the model that assigned it, so the model is named and pinned in every record. Across five embedding models the angle separated “same subject” from “unrelated” perfectly and put a sentence within a hair of its own negation; so the angle guards subject and the readers’ stances carry direction. It would be disproved as a fixed division of labour if a model’s angle separates “same point” from “opposite point” as well as a reader does.
• Calibrating a reader’s confidence A reader that says 0.9 should be right about nine times in ten. For each reader, how often a stated 0.9, 0.7 and 0.5 is right is measured on a labelled set, and a correction fitted on half is tested on the other half. It would be disproved if the correction does no better than the raw confidence on the half it never saw; then alternative confidence policies are compared with label-only weights.
• λ between readers How alike are ten runs of one model, one model under two prompts, and two models from different makers? The within-family correlation sets the discount in step 08. It would be disproved as a grouping if runs inside one model family are no more alike than runs across families; then common dependence and alternative grouping policies are evaluated.
• Measuring λ for each kind of material λ for syndicated news is not λ for scraped product pages or for text written by a model. It would be disproved as a stable quantity if the value measured on one half of a collection falls outside the interval measured on the other half.
• Finding the origins Copying more is capped at 1/λ. Making one origin look like several is not. It would be disproved if planted copies from a single controlled origin ever lift the count above that cap.
• Agents that agree with each other Copies of one model, one prompt or one retrieval run share an origin too, and sharing can be nested several levels deep. It would be disproved if a discount with a level for each kind of sharing does no better than a single λ on agent groups whose shared inputs are known.
• Direction against a plain count Whether weighting material by its angle to the question predicts the outcome better than volume alone. It would be disproved if the weighting adds nothing on material held back from fitting.
• Acceleration as an earlier warning Whether a story that is speeding up can be flagged sooner than a volume threshold flags it. It would be disproved if the lead time disappears once both are held to the same false-alarm rate.
• Fitting the bars once the readings are right Today’s bars were fitted to word counts. The allow and block bars in step 08 are refitted on a labelled set, half for fitting and half held back. It would be disproved as a rule if the bars chosen on one half mislabel more of the other half than a plain majority of readers does.
• The angle between an instruction and an action Coordinates put a sentence close to its own negation, so the angle in step 07 can say whether an action is about what was asked, not whether it does what was asked; that is read separately. It would be disproved as a useful guard if actions on another subject slip past the bar as often as actions on the right one.

A crowd is cheap to manufacture. Independence is not. That is the difference TruVector measures before an AI reads, before it agrees, and before it acts.

Company

Built by an engineer who measured things for a living.