Cognivec · Research

Experiment reports

Selected reports from the research program, published as generated — including the negative results. All conversation corpora below are public, PII-redacted datasets (AIxBlock automotive & medical calls, Bitext support chat).

Cross-Industry Conversation Analysis

May 2026 · 240 calls + 1,500 chat turns · automotive, medical, support

Semantic-physics features detect conversation phase transitions 2.5× above within-call shuffles, pinpoint critical micro-moments with 5–6× enrichment over baseline, and show the dominant semantic axis is universal across industries (99.9% PC1 correlation, automotive↔medical).

Honest Comparison: Cognivec vs an LLM

Negative result · May 2026 · 120 automotive BDC calls

Where our features lose: on closed-set soft-close vs hard-close classification the LLM wins decisively (68% disagreement, Cognivec systematically too lenient). The honest recommendation is a hybrid pipeline — geometry ranks calls for free, the LLM classifies the top slice. We publish results like this on purpose.

Sample Deliverable: BDC Call Review

May 2026 · 120 calls · what a manager actually receives

A worked example of the end product for a call-center manager: 120 calls triaged automatically, the 31 "soft close" calls that likely hold the week's lost revenue listed in priority order, with the evidence for each.

News Is Stochastic, Not Chaotic

Negative result · 2026 · Takens embedding + RQA + Lyapunov · UK, US, AU news

A clean kill of a seductive idea: if the news flow were a low-dimensional chaotic system, chaos-theory tools would forecast it. We ran the full apparatus — Takens delay-embedding, recurrence quantification, largest Lyapunov exponent (Rosenstein) — on three countries' news-novelty series. Lyapunov ≈ 0 across all three: the news tape is stochastic, and dynamical-systems forecasting does not apply. Report in Russian.

Foundational studies · originals in Russian

Character = Destiny: 183 Writers

2026 · 183 English-language authors, 1640–1896 · leave-one-out + permutation nulls

The core hypothesis tested at scale: a writer's semantic axis — the geometric mean direction of their language — carries their biography. Fate labels (tragic vs untroubled lives) were fixed before analysis; validation by leave-one-out with a 2,000-draw permutation null. The axis also encodes historical epoch (r = 0.71 with birth year). Report in Russian.

The Geometry of Talent

2026 · 50 Russian classics vs 55 amateur authors · AUC 0.88

Can geometry alone tell a canonical author from amateur prose? Each text is run through a small LLM in teacher-forcing mode, comparing the word actually written against the word the model expected. Twenty deviation features per text separate classics from amateurs with AUC 0.88 — and the signature of talent turns out to be consistent, structured deviation from expectation, not randomness. Report in Russian.

Psychological Map of Sign Classes

2026 · sign patterns of top-6 principal components · fit on 10 authors, tested on 10 held-out

Words cluster into discrete "sign classes" — patterns of pluses and minuses across the leading components of pair-difference space. Each class reads as a coherent psychological register, named by joint reading of its words, and the classes transfer across held-out authors. A vocabulary of style that generalizes. Report in Russian.

The Bible Through Vectors

2026 · KJV, Genesis + New Testament · same pipeline as team analyses

A demonstration of method universality: the exact pipeline used for team communication — gram matrix, PCA axes with logit-lens poles, the "nerve" (the direction that most distinguishes a corpus from neutral English) — applied to scripture. Companion visualization: the New Testament fingerprint. Report in Russian.

Associative Portrait (concluded experiment)

2026 · voice word-association test · multilingual-e5-large

A spoken word-association test turned into a personality instrument: free associations to 70 stimulus words, embedded and projected onto 17 robust psychometric axes built from 90+ antonym pairs, plus ~50 theme centroids and reaction-time analysis — yielding an interpretable "associative portrait" from a few dozen words. Consistent with published findings that embedding-scored associations correlate with standard psychometrics (r ≈ 0.3–0.6). The experiment is concluded; the public demo has been retired. Session-level reports contain personal data and are not published.

Markets line

FX Results Dossier

Results public · methods under NDA · graveyard included

Headline results of the markets research line: eight surviving signal families (topological hole-direction, curvature bridges, semantic novelty, hard-pairs portfolio and more) with their out-of-sample Sharpe ratios and p-values — plus the graveyard of 100+ killed hypotheses that gives the survivors their meaning. Constructions and protocols deliberately withheld; verification offered by free 8-week forward-test instead.