Nonlinear K4 Pareto Promotion Report

Source artifact rendered for reading.

Back to post Raw Markdown

This report compares the K4 spread state-density candidate, the K4 spread late predictive-y candidate, and the bootstrap particle-filter reference.

Candidate Metrics

Pattern	Row	seeds	state NLL	cov 90	var ratio	pred-y NLL
intermittent_sinusoidal	PF n512 reference	3	3.140	0.847	0.865	0.348
intermittent_sinusoidal	K4 spread state-density	3	2.640	0.786	0.619	0.362
intermittent_sinusoidal	K4 spread late pred-y	3	2.643	0.780	0.609	0.361
random_normal	PF n512 reference	3	4.288	0.819	0.807	0.521
random_normal	K4 spread state-density	3	3.278	0.747	0.477	0.575
random_normal	K4 spread late pred-y	3	3.308	0.745	0.473	0.571
sinusoidal	PF n512 reference	3	3.300	0.845	0.847	0.452
sinusoidal	K4 spread state-density	3	3.163	0.830	0.619	0.514
sinusoidal	K4 spread late pred-y	3	3.185	0.832	0.617	0.510
weak_sinusoidal	PF n512 reference	3	2.946	0.844	0.899	0.303
weak_sinusoidal	K4 spread state-density	3	2.724	0.918	1.082	0.329
weak_sinusoidal	K4 spread late pred-y	3	2.727	0.910	1.053	0.328
zero	PF n512 reference	3	2.843	0.871	0.899	0.263
zero	K4 spread state-density	3	2.757	0.935	1.190	0.263
zero	K4 spread late pred-y	3	2.757	0.935	1.190	0.263

late pred-y is marked as a secondary candidate when it improves pred-y and costs no more than 0.03 state NLL.

Pattern	pred-y gain	state NLL cost	PF pred-y gap	recommendation
intermittent_sinusoidal	0.001	0.003	0.013	late pred-y secondary
random_normal	0.004	0.030	0.050	late pred-y secondary
sinusoidal	0.004	0.022	0.058	late pred-y secondary
weak_sinusoidal	0.001	0.002	0.026	late pred-y secondary
zero	0.000	0.000	-0.000	keep state candidate

The K4 spread state-density row remains the main promotion candidate.
The late predictive-y row is useful as a secondary/Pareto row, but its pred-y gains are small relative to the PF reference gap.
The next research step should target the predictive normalizer directly rather than adding another scalar predictive-y weight.