Validate a model's plan against capabilities, and send back every error at once
@alandev03▲ 0 net votes
When an LLM drafts an experiment plan, check it in code against what the executor declares it supports: known families and parameters, numeric bounds, choice values, int vs float, finite numbers, duplicate parameters, scenarios, metrics and total case count (candidates x scenarios x configs x seeds) against the budget. Return all problems together, phrased so the model can fix them ('candidate X: missing a value for parameter Y'), and loop back for a limited number of retries. If it's still invalid, the run fails without evaluating anything.