Dynamic Architectures

Use this contract for the concrete task. Apply a short relevant check for a small change; expand investigation when the failure, risk or requested artifact warrants it. Resolve facts from the repository and runtime before imposing a process. Delegation is optional and should answer a bounded unresolved question.

Mutation and training contract

Record the host/attached-module interfaces, permitted mutations, parameter/compute budget and old-task evaluation set. Separate how the network trains from the policy deciding whether to mutate it.

• Specify which parameters and buffers may change in each lifecycle state. Freezing owned weights means preserving values, not erasing them with a mask.
• Trace gradient paths through host and new modules. detach, no_grad, parameter freezing and optimizer membership have different effects; test the intended paths.
• Account for optimizer state, schedulers, checkpoints and distributed synchronization when parameters are added or removed. A shape-valid forward pass is insufficient.
• Check output shape, normalization, device/dtype and initial influence at attachment points. Use a declared blending/warmup policy when gradual integration is needed.
• Define lifecycle transitions, contribution/stability gates, hysteresis and rollback state. Preserve old-task performance evidence while assessing new capability.
• Compare growth with static and simpler adaptation baselines under matched resource budgets. Stop growing if it adds cost without demonstrated benefit.

Evidence and output

Produce a lifecycle/interface specification or a focused repair with parameter-ownership map, gradient/state checks, integration conditions and rollback/replay evidence. Record what was tested and what remains a research assumption.

Fault-specific references

• Old tasks regress: continual-learning-foundations.md.
• New-module training changes the host unexpectedly: gradient-isolation-techniques.md.
• Capacity or lifecycle thrashing: dynamic-architecture-patterns.md, ml-lifecycle-orchestration.md.
• Gating, grafting, merging or interface mismatch: modular-neural-composition.md.
• Integration shock: progressive-training-strategies.md.
• Adapter-method selection: peft-adapter-techniques.md; LLM application/data/tuning workflow belongs to LLM specialist.

An RL mutation controller's action/reward/governor design belongs to morphogenetic RL. Generating a novel candidate graph belongs to structure synthesis. Do not require those workflows for fixed adapter tuning.

Optional references

All sheets below are in this directory. Choose a sheet because its checks or examples help the task; there is no requirement to read the catalog in sequence. Verify time-sensitive APIs and numerical/performance claims before relying on examples.

• continual learning foundations
• dynamic architecture patterns
• gradient isolation techniques
• ml lifecycle orchestration
• modular neural composition
• peft adapter techniques
• progressive training strategies