Cognivec

The geometry of what goes unsaid

Cognivec measures the shape of meaning in text — and extracts signals that sentiment analysis and LLM summaries cannot see: hidden intent, structural blind spots, and early behavioral change.

Research reports · Contact

A day of news as a cloud of word vectors. The empty ring is a topological hole — a region of meaning the text circles around but never enters. Cognivec finds these holes, names their poles in plain words, and tracks them over time.

Method

Every text — a news day, a year of team chat, a sales call — becomes a cloud of high-dimensional word vectors. Standard NLP asks what the words say. Cognivec asks what shape the cloud has: its dominant axes, its curvature, its topological holes, and how that shape moves day over day.

The core finding, validated across three independent domains: the most predictive information in language is structural, not lexical. What a speaker avoids, how their meaning-space is organized, and how far their words deviate from a model's expectation — these are measurable quantities, and they forecast behavior.

Every signal is interpretable: a hole's poles are named in ordinary words ("people's names" vs. "the economic machine"), an axis reads as a spectrum you can quote. There is no black box between the measurement and the claim.

Evidence

All results come from our research corpus: 20 years of news (≈470k articles across UK and US sources), multi-year team communication archives, and public literary corpora. Market results follow a strict causal protocol — walk-forward out-of-sample testing, placebo controls, and pre-registration of confirmatory runs. Full experiment reports — negative results included — are in the Research section.

Why this is different

• Structure, not sentiment. Sentiment saturates and lags; geometry moves first. Our signals survive placebo tests that kill most text-based factors.
• Interpretable by construction. Every signal decomposes into named words and axes. A risk officer or an HR director can read the evidence directly.
• Runs on small open models. Compact open-weight encoders on commodity hardware — no API dependency, data never leaves your perimeter.
• Validation discipline. Pre-registered confirmatory tests, placebo controls, and a kept graveyard of 100+ killed hypotheses. We know exactly why the survivors survived.

Working with us

Funds and trading desks: a 6–8 week signal evaluation — daily signal files for a pre-agreed universe, judged in your own backtest and paper-trading stack, success criteria fixed in advance.

Organizations: a 4–6 week pilot on an archive you choose (team chat, support transcripts, sales calls), fully on-premise. Deliverables: a blind-spot map, aggregate risk indicators, and live conversation telemetry where you want it.

Research partners: the instrument set (hole detection, axis fitting, deviation profiling) as a collaboration — your corpus, joint publication, shared tooling.