Graphbrain Neural Evolution Pack v2

This extension adds architecture-inspection, layer-surgery, evaluation, and retraining contracts so the Graphbrain-style symbolic/neural framework can operate on real TensorFlow, Keras, PyTorch, ONNX, and related networks inside TritFabric.

Contents:

• NETWORK_SURGERY_SPEC.md — canonical architecture for inspection, surgical modification, retraining, and governed promotion
• TRITFABRIC_RETRAINING_FLOW.md — execution flow for distributed training/evaluation in TritFabric
• schemas/ — machine-readable JSON Schemas for network artifacts, probes, patch plans, retrain jobs, eval suites, and reports
• examples/ — example patch plan and retrain job
• code/model_surgery_api.py — Python skeleton for adapters, patch plans, and evaluators

The governing idea is to treat each model as a first-class governed artifact with:

• a stable architecture representation
• a probe report over graph topology, layers, tensors, and trainability
• a surgical patch plan with auditable intent
• a retrain job with data, budget, identity, policy, and twin scenario bindings
• an evaluation report that combines ML metrics with governance, twin, and graph-quality checks

Recommended order of implementation:

1. Implement framework adapters that can emit NetworkArtifact and ArchitectureProbe.
2. Gate all modifications through LayerSurgeryPlan.
3. Execute retraining only through RetrainJob on TritFabric.
4. Promote or reject a candidate only from ModelEvaluationReport.
5. Keep symbolic Graphbrain/claim/provenance checks in the loop for high-risk or graph-mutating actions.

Additional code stub:

• code/tensorflow_keras_adapter_stub.py — minimal adapter example for inspecting Keras models and conservative freeze/unfreeze patching