Theta AI · Measurement engine
Theta AI converts a role-specific job description into a structured assessment model, estimates latent capability with IRT, and adds independently reviewable integrity evidence before any recruiter sees a recommendation.
01 · Role model
We map the responsibilities, decision conditions, risk exposure, and competency priorities in your job description. That role model governs the campaign, so candidates are evaluated against the work they will actually do, not a generic personality ideal.
02 · Graded Response Model
Theta uses a five-category Item Response Theory model rather than treating every response as equally informative. The engine estimates candidate theta, percentile, performance level, and item-level response behavior while accounting for how strongly each item separates levels of capability.
03 · Frozen ensemble
In synth-detectors-v1, XGBoost 0.35, LightGBM 0.35, and Random Forest 0.30 are combined; deep learning weight is 0.00. The frozen target-label cutoff is 0.495350 and the classification threshold is 0.5. This is configuration evidence, not a claim of real-world performance.
04 · Synthetic recovery evidence
For synth-detectors-v1 untouched synthetic validation, theta recovery was r = 0.850 (95% CI 0.794–0.900; 2,000 bootstrap samples; final validation run 35578002689). For synth-bank-v1 synthetic calibration (n=500), marginal reliability was 0.913. Neither result is real-world job-performance validity.
05 · Integrity engine
On synth-detectors-v1 untouched synthetic validation, precision was 0.900 (95% CI 0.794–0.977), recall 0.900 (95% CI 0.795–0.979), and specificity 0.991 (95% CI 0.982–0.998). These are synthetic validation results; a flag is never automatic proof of misconduct.
06 · Decision report
Recruiters receive theta, percentile, level, integrity context, cohort band, score summary, and an embedded HTML report. The output is designed to identify who advances to Titan and what still needs to be tested in conversation.
The point
Role context, fixed seeds, explicit thresholds, frozen model settings, versioned outputs, and a full report create evidence that can be reviewed after the decision.
