Analysis Methodology
The Lindr Engine uses an ensemble approach to personality detection, combining lexical, syntactic, and semantic signals to build a high-fidelity behavior profile of every LLM response.
The Three Signal Pillars
Signal 01: Lexical Analysis
We analyze the specific word choices and token distribution of the response. Our engine extracts psycholinguistic features including pronoun patterns, emotion markers, certainty indicators, and cognitive complexity signals.
// Each dimension maps to validated linguistic markers
Signal 02: Syntactic Profiling
Sentence structure conveys personality as much as word choice. Lindr measures average sentence length, punctuation density, and readability indices (Flesch-Kincaid). A "Formal" persona will show significantly different syntactic markers than a "Casual" one.
Signal 03: Semantic Embeddings
This is our most advanced signal. We use local, high-performance embedding models to measure the "vector distance" between a response and reference exemplars of specific traits. This allows us to detect subtle behavioral nuances that keyword matching might miss.
Ensemble Fusion & Confidence
No single signal is 100% reliable. The Lindr Engine uses a Weighted Voting Fusion algorithm. Each signal is assigned a weight based on its psychometric validity and reliability characteristics.
Calibration
Scores are calibrated using curated reference exemplars for each personality dimension, ensuring 50 represents a balanced trait expression.
Reliability Score
If signals strongly disagree (e.g., Lexical says high openness but Semantic says low), the Confidence Level for that session is automatically downgraded to "Low".
Transparency Commitment
Unlike many "AI Evals" that use an LLM-as-a-judge (which introduces bias and high cost), Lindr relies primarily on deterministic linguistic algorithms and local inference. This ensures that our monitoring is consistent, auditable, and incredibly fast.