Paper 3 · Section 10Learning
A concrete map of the learning problem
The findings separate architectural capacity, finite-budget production, calibration selection and interaction-law validation, with direct evidence that fitting effort repairs a material part of the original gap.
10 Conclusion
The original same-evidence comparison shows a useful finite advantage of standard probabilistic automata and controlled hidden-state models over the tested tree controls. A prespecified secondary portfolio without the bespoke group-aware candidates reproduces the full portfolio's broad performance. The paired fixture directions support that aggregate description while preserving severe mechanism-specific reversals.
The capacity check is elementary but diagnostically useful: the chosen general architecture can express every target. Under the original budget, produced individual successes cover 39 of 48 fixtures and calibration selection succeeds on 37. These figures are properties of the executed search, not a structural discovery impossibility. The separate post-hoc six-start, 600-update analysis repairs five of the 11 selected failures and produces passing candidates on seven. Two public-selection misses persist, including one deterioration, while both locks and both challenges still lack an individually passing expanded candidate. Three old events also certify obstruction of their expanded banks.
The challenge diagnosis refines the failure mechanism. The selected old models distinguish observed challenges, but mix responses that depend on the prepared bit; the tree exposes preparation, phase and last output directly. Increased fitting effort improves that discrimination without yet making the full challenge laws adequate. Thus capacity, finite-budget search, selection and record-law validation remain distinct, with an important part of the original gap now demonstrably budget-sensitive.
The original panels, context duplicates, calibration overlaps, secondary designations and pre-score repairs remain part of the reported evidence. The retrospective additions neither replace those results nor constitute new external confirmation. The next validation steps are methods review and independently authored benchmarks, not another rebranding of this same-author finite study as a universal interface-learning result.