Technical doc 32.16.1.7. Model families in
@kuanyin/ai-ml-models; gating in@kuanyin/evaluation.
Kuan Yin specifies six model families. Each is documented with a model card
(intended use, training data shape, fairness constraints, known limits) and is
gated through @kuanyin/evaluation before it touches a person.
The six families#
| Family | Output |
|---|---|
| Intent classification | 25+ intent classes, multi-intent head, calibrated confidence |
| Emotional state detection | Arousal-valence + discrete emotions, trajectory, crisis module |
| Harm potential scoring | Multi-label severity + reversibility + cascade risk |
| Cognitive reframing | CBT distortion detection + intent/tone-preserving reframe |
| Behavioural pattern | Sequence model for escalation/recidivism/intervention selection |
| Community health | Graph-neural relationship model + conflict/raid/health trajectory |
Each family supports continuous learning, A/B testing, bias detection + mitigation, confidence calibration, attention-based explainability, and versioning with rollback.
Evaluation gate (@kuanyin/evaluation)#
model-evaluation.ts computes, against hand-verified known values:
- Classification — confusion matrix, precision/recall/F1/MCC, Cohen's kappa, ROC-AUC and PR-AUC, calibration (ECE/Brier).
- Regression — MAE/RMSE/R².
- Fairness — demographic parity, equalized odds, EEOC four-fifths rule, cross-cultural disparity.
- Robustness — adversarial-perturbation suites, edge-case runner.
- Operations — latency percentiles (p50/p90/p95/p99 + SLA), regression vs baseline, two-proportion A/B testing, human inter-rater agreement, and PSI/KL/JS drift detection.
A model release is blocked if any fairness constraint regresses, drift exceeds threshold, or the latency SLA is missed.
Privacy & federation#
Models train under the privacy regime in 13-privacy-implementation.md:
on-device feature reduction for typing dynamics, ephemeral content analysis, and
federated learning (dharma-analytics/federated-learning.ts) with FedAvg,
gradient clipping, (ε,δ)-Gaussian differential privacy, and secure aggregation —
so cross-community learning never requires sharing raw user data.
Serving#
Real-time families target sub-100ms inference (intent) and sub-50ms fast paths
(toxicity), with model quantization (fp32→int4) and edge deployment for
latency-critical analysis. See dharma-analytics/performance-sla.ts and
11-performance-tuning.md.