yuji-tt Discovery Loop
Use the smallest subset needed. User instructions override defaults; a routine implementation with a clear objective usually needs execution, not this loop.
Reconstruct the real goal and constraints; distinguish them from the proxy and current optimization target. Keep objective mismatch and solution insufficiency alive until evidence discriminates them. Audit only consequential uncertainty, then deepen the strongest plausible direction with a causal prediction, baseline, ablation, and kill criterion.
Compare evidence to those predictions. Classify the limitation as objective mismatch, missing information, representation limitation, algorithmic limitation, optimization limitation, incorrect causal hypothesis, evaluation artifact, fundamental constraint, or insufficient evidence. Multiple explanations may remain live. Read decision loop when selecting a diagnostic or deciding whether to pivot.
Continue deepening while a bounded informative test can distinguish a plausible mechanism from alternatives. A flat score or exhausted patience alone does not establish a structural plateau. When valid tests expose a persistent local limitation, search distant mechanisms by structural signature, translate the best mapping into a minimal target intervention, and test against the baseline under comparable resources.
Use objective-audit, research-deepener, or structural-transfer if installed and relevant; never assume sibling skills are installed. Without them, apply the corresponding compact steps above and the self-contained reference. Do not invoke all skills mechanically.
Update the objective only with goal-alignment evidence, the hypothesis with mechanism evidence, or the method with implementation evidence. Stop when the requested decision is resolved, the agreed experiment budget is consumed, or needed evidence is unavailable. Return a decision record with the next discriminating action, not an endless loop or a list of untested ideas.