engineer-code

Decorate the function with @trace.bundle(trainable=True) so it behaves like a normal callable while participating in the Trace computation graph. Then define a feedback function from tests or scores and run the zero_feedback(), backward(...), step() loop.

Use trace.node(..., trainable=True) for editable values like prompts or thresholds, @trace.bundle(trainable=True) for optimizing a function's behavior as a unit, and @trace.model when several related trainable members belong to one cohesive object.

Yes. Trace accepts non-scalar feedback such as test failures, compiler errors, and natural-language critiques. You can also capture runtime exceptions via ExecutionError and feed them back into the optimizer instead of discarding them.

No. Trace is not the primary tool for generic runtime performance tuning. For slow ETL jobs, use profiling, algorithmic changes, or systems optimization first; Trace only helps when a repeatable feedback signal can improve behavior.

Hidden dependencies inside a bundled function can break tracing assumptions, and large graphs create context pressure for graph-aware optimizers like OptoPrime. Keep trainable surfaces narrow so feedback stays attributable and updates remain interpretable.