Model registry

Governance-weighted blend of the four retention voters (Cox survival, LightGBM, logistic regression, neural MLP). Voter weights are updated from observed placement outcomes by leave-one-out attribution (M5.1) and an EMA trust-region simplex update (M5.2); a Wald SPRT (M5.3) gates promotion of a challenger blend before it ships.

Fitted logistic-regression retention model predicting P(retain >= 90 days) with interpretable per-feature coefficients. The most accurate single retention voter on synthetic data, and the one that explains its own reasoning.

LightGBM gradient-boosted retention classifier predicting P(retain >= 90 days) from the candidate-placement feature vector. Captures nonlinear feature interactions the linear voters miss.

From-scratch numpy multilayer-perceptron retention model (manual backprop + Adam) predicting P(retain >= 90 days). No external deep-learning dependency.

Cox proportional-hazards survival model (lifelines) producing a full survival curve S(t) and a 90-day retention-risk read. Built on five refugee-support covariates: language-learning velocity, transit access, co-ethnic community distance, time since arrival, and funded-training match.

K-means cohort segmenter (Lloyd iteration with k-means++ seeding) that groups candidates by their support-feature vector to give caseworkers workable cohorts. Descriptive segmentation only — it never scores an individual for matching and is not a retention voter.

Credential confidence scoring (M1.1) with a live verification hook (M1.5) that runs a real HTTPS registry-reachability probe, a bilingual transferable-skill graph (M1.2), and a months-to-ready skill-gap estimate (M1.3).

Multilingual transformer sentence embeddings (intfloat e5) with an offline character-n-gram hashing fallback when the transformer is unavailable. L2-normalized vectors power semantic skill/title matching and approximate-nearest-neighbor candidate recall.

Self-improving document extraction: density-based routing (M3.1) chooses the parse path, field-level calibration (M3.2) tracks per-field reliability, and a correction-retrieval store (M3.3) injects past human corrections as few-shot guidance into the next parse.

Continuous feedback loop across four channels — match rating, placement outcome, parse correction, and translation correction. Every event updates a learner: placement outcomes flow into the governance window and move the retention-ensemble weights, while parse/translation corrections feed the extraction retrieval store. Exposed at POST /api/v1/feedback.

Eight-dimension match engine combining skills (0.27), experience (0.20), logistics (0.15), retention (0.15), language (0.10), education (0.05), culture (0.05) and interview (0.03), with an essential-skill coverage cap and a cross-function-family cap. The retention dimension reads the retention ensemble.

Softmax vertical router estimating P(vertical | NAICS code, job title, description) with per-vertical weight vectors, so a job is routed to the right industry vertical before matching.