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.