2.1 KiB
Confidence-prioritized investigation scheduling
The controller gives every active, eligible hypothesis one bootstrap turn in numeric ID order. After that it chooses by qualitative confidence: strongly_supported > supported > plausible > unassessed > weak > weakened. Equal levels use numeric hypothesis ID (H1 before H2 before H10), not rotation. These ranks are ordering values, not probabilities or a new model-owned budget.
Selection is recomputed between turns. A strengthening hypothesis can continue; a weakened hypothesis yields to a higher-ranked alternative without losing its remaining budget or evidence. Initial turns generally issue queries: their results are semantically assessed on a subsequent turn, so the bootstrap does not claim that all hypotheses are already evaluated after three calls.
Existing eligibility checks, positive action costs, rewards, iteration safeguards, terminal dispositions and final pending-result assessments remain in force. An interrupted/pending active turn resumes as before; it is not preempted or rewritten. Existing cases adopt the scheduler at their next selection boundary. Archived states, contexts and model request hashes are not changed. The historical cursor field is retained for file compatibility but no longer determines selection.
Completion remains separate from selection
The model may finish its selected hypothesis with remaining credit, or reject it. Supported confidence requires stored evidence. This does not itself globally finish the case. The controller transitions to BUILDING_DIAGNOSIS after all hypotheses have terminal accounts and pending results have been handled, whether through model finish/reject, budget exhaustion or controller limits. No rule requires spending every allocated unit. A weak hypothesis is deprioritized, not automatically rejected just because another ranks higher.
The final diagnosis evaluates all alternatives, root-cause evidence and adequacy of a countercheck. Only an accepted supported diagnosis can proceed to a coding change request; inconclusive cases go to review. An automatic global early finish on a single high-confidence declaration is not introduced by this scheduler.