# Phase 93 edge and on-device deployment

The implementation is provider-neutral and fail-closed. Apple CoreML/MLX,
Qualcomm QNN, MediaTek NeuroPilot, TensorRT, LiteRT/TFLite, ExecuTorch, MLC,
WebLLM and WebGPU SDK calls enter through adapters; the core never reports a
successful conversion without artifact integrity and quality evidence.

## Capability map

- `edge-compilation.ts`: ONNX optimization and all target compilers, INT8
  calibration requirements, validation, structured/unstructured pruning with
  fine-tuning, Axolotl QAT, teacher/student distillation, and supported SLM
  fine-tuning.
- `edge-runtime.ts`: per-class budgets, device profiling and Pareto selection,
  quality-aware device/edge/cloud fallback, complete offline bundles,
  AES-256-GCM in-memory decryption, WebGPU/WebLLM selection, and battery/thermal
  scheduling.
- `edge-operations.ts`: signed differential atomic updates, fleet telemetry and
  issue detection, device-vs-cloud promotion gates, and rollback-safe OTA CI/CD.

## Production invariants

Vendor adapters must pin SDK/compiler and target versions in artifact metadata.
Real target-device profile evidence is mandatory before distribution. Model
updates require a trusted manifest, patch/base/result hashes, staging and atomic
activation. Offline mode requires model, tokenizer and vocabulary integrity.
Encrypted cleartext is held only for runtime loading and must never be persisted
or logged. Device experiment promotion requires quality parity while also
improving latency and cost, and installation verification failure rolls back.
