Ratings can lie, but gaze never does #
Every content platform dominating the market today stands on the same premise — that likes, dwell time, and completed views reflect genuine preference. That premise is wrong. The rational self gives a philosophy essay five stars; the instinctive self scrolls away in three seconds.
The gap between these two selves — the Two-Self Gap Δui — is precisely where genuine cognitive disposition hides. GQ-Engine measures this gap mathematically, decomposing the cognitive density of Korean text into cognition nodes — through a white-box pipeline, free from commercial generative-LLM bias.
J(u) = ⟨S̄u, R̄u, IV̄u, GP̄u⟩ — weighting the gap between explicit preference Sui and revealed preference Rui with text parameters.7 research reports converging into one argument #
Cognitive Engineering of Geulgage (WP-2026-01) gathers 7 research reports — 3 Foundation, 4 Theory — alongside an engineering spec, a business commentary, and a market analysis, organized in 5 parts: Foundation · Theory · Engineering · Business · Market. Cognition can be measured — and measured cognition can be returned.
GQ-Engine Architecture #
GQ-Engine is a white-box hybrid: a rule-based core eliminates commercial generative-LLM bias, while a KoSentenceBERT layer adds sentence-level speech-act understanding. Features are extracted via Kiwi POS-tagging, reweighted by the speech act they appear in, scored across the cognition nodes, then re-ranked by a statistical post-processor (Layer 4). Every verdict is traceable back to its features. Local Gemma runs an independent audit in the background — never on the response path.
G·LAB · Precondition for Measurement
Direct preference elicitation via Likert scales (star ratings) induces SDB (Social Desirability Bias). Respondents distort answers to appear more intellectual or moral — contaminating the collected Sui at the source.
G·LAB bypasses this bias through a micro-fiction-based SJT (Situational Judgment Test) methodology. Users choose between options for a character in a short story, unaware they are being measured.
Δui is contaminated. G·LAB securing the validity of Sui is the logical precondition for the entire gap-measurement system.Two-Self Gap · Δui
Explicit preference Sui is a normalized rating value; revealed preference Rui maps the weighted sum of dwell time, reading speed, and back-scroll onto a CDF.
When the gap Δui is positive, Intellectual Vanity emerges; when negative, Guilty Pleasure surfaces.
Text Parameters · Cdiff · Cprov
[0, 1]. Neuroscientifically, corresponds to the surface-feature set of the KRIT model predicting ACC (anterior cingulate cortex) activation probability.Antagonistic Fit · Exclusion Coefficient
When the dot product of a user's rejection-disposition weight vector and content node probabilities exceeds a threshold, the content is permanently removed from the recommendation pool. Not a simple filter — a neuroscientifically modeled exclusion algorithm based on aversion response mechanisms.
Analysis Pipeline · 5 Stages
A sentence is not scored in isolation. GQ-Engine reads each sentence's speech act first, then lets that context reshape how its features are interpreted — before any node score is computed.
NP pronoun, EC/ETM clause endings, EF sentence-final, VX/MAG negation, NNG nouns) drive 30+ raw features. Similes are gated on the JX particle tag, not surface keywords. Hardcoded constants removed in v1.5.0.ko-sroberta-multitask) embeds each sentence and matches it against per-node anchor sentences to assign one of 10 speech acts — confession, reversal, question, assertion, provocation, description, metaphor, argument, recollection, narration.metaphor_pattern_norm; a confessional one amplifies suppression_lexeme_ratio.dominant), in which case the raw ranking stands. See below./analyze returns tags synchronously the instant content is published, with zero analysis delay. The Gemma audit is fully decoupled via BackgroundTasks, so the user-facing latency depends only on the rule-based core.Layer 4 · Form-vs-Content Re-ranking
Well-written prose always scores high on the two form nodes (Fluentia, Refractio) — they crowd out the content nodes that reveal what a text is actually about. This is the halo effect of automated scoring. Layer 4 corrects it with a 3-stage classification, grounded in the inter-node covariance and bifactor analysis of report TECH-2026-04.
lyrical; Refractio in the top 3 → argumentative; neither → low_quality.true / false.true / false. Narrātio is deliberately excluded here, since it otherwise dominates every personal text.major / middle / minor. Layer 4 is skipped entirely when the top-2 score gap ≥ 0.15 (trigger: dominant) — a clear winner is never overridden.Statistical Validation
The headline golden-set accuracy (81.0%, 17/21) is one checkpoint, not the whole story. The node system is held to a standing psychometric battery so it stays valid as the corpus grows.
10 Lēctio Nodes · Cognition Node System #
The 10 Lēctio nodes defined in FOUND-2026-02 are axes that decompose how text stimulates the reader's neural-cognitive circuits. Each carries a weight in the [−1.0, +1.0] range on the 4 LECT axes. In the engine, 9 nodes are scored directly (2 form + 7 content); the 10th, Synthesio, is derived as the harmonic mean of 4 base nodes — a composite integration index rather than an independent measurement.
Normalization Formula · 4-Axis Persona
Final 4-axis disposition scores are computed from the 10-node raw score vector N ∈ ℝ¹⁰ and axis weight vector W ∈ ℝ¹⁰. One of 16 LECT codes is assigned based on the sign.
Santiago Protocol · 10-Day Adaptive Validation
Day 1 calibrates with 30 multiple-choice items; Days 2–9 present 20 sentences per day (160 items, rated + I/O scoring) updated nightly via MIRT-based Bayesian posterior estimation; Day 10 delivers results. Total: 190 items.
θu distribution.IVui · GPui and maximizing Δui variance.Moat · Unassailable Business Moat #
Market · Global Mental Wellness Crossroads
GQ-Engine's market is not e-books or web novels but the intersection of Global Mental Wellness × Premium Knowledge Subscriptions.
2025 · Emergen Research
Deep Reading Segment
Addressable Market
2025 National Reading Survey
TECHTONIC #
TECHTONIC is a small Seoul-based research team. At the narrow intersection of cognitive psychology, neuroscience, psychometrics, NLP, and systems engineering, we ask: can Korean text become a tool for measuring thought?
Pre-seed · GQ-Engine v1.5.0 (Kiwi + KoSentenceBERT)
Product app ↗ · Engine GQ-Engine (white-box hybrid + Gemma auditor)
Geulgage
A text-based reading subscription platform. Encounter docent content on classical music, painting, and literature in a deeply immersive app environment. GQ-Engine automatically derives cognitive tags for all content.