ASAREE's vision (vision.md) describes "analyze results" as nonparametric
regression (GPR, Random Forest, kNN, Spline/LOESS) over collected
factor_values/metric_values — producing performance surfaces, factor-importance
rankings, and optimal configurations with uncertainty. That's genuinely
generic across any ASAREE experiment: fit a flexible model over whatever
factors and metrics an experiment collected, no assumptions about factor
count, level count, or design type (binary/full-factorial or otherwise).
We're not building that first. The spinal_surgery use case needs a specific,
already-designed statistical methodology (effect-coded ±1 fixed-effects
factorial model, Freedman-Lane permutation tests with max-statistic FWER
correction, BCa bootstrap non-inferiority testing against a reference
condition with Holm correction, homogeneity/heteroscedasticity diagnostics)
replicated as an ASAREE feature so the use case's demonstration for an
upcoming bioinformatics journal application note reflects what ASAREE
actually does. That's a narrower, use-case-specific analysis path, not the
generic capability the vision describes.
This issue tracks building the generic nonparametric-regression capability
as a separate, follow-on feature once the use-case-specific path is done —
so the vision's "analyze results" claim is fully backed by a genuinely
reusable primitive, not just the one methodology one use case happened to
need.
See project_plan/core_asaree_use_case.md (ARES repo) §10 for the fuller
design-vs-analyze discussion this decision came out of.
ASAREE's vision (vision.md) describes "analyze results" as nonparametric
regression (GPR, Random Forest, kNN, Spline/LOESS) over collected
factor_values/metric_values — producing performance surfaces, factor-importance
rankings, and optimal configurations with uncertainty. That's genuinely
generic across any ASAREE experiment: fit a flexible model over whatever
factors and metrics an experiment collected, no assumptions about factor
count, level count, or design type (binary/full-factorial or otherwise).
We're not building that first. The spinal_surgery use case needs a specific,
already-designed statistical methodology (effect-coded ±1 fixed-effects
factorial model, Freedman-Lane permutation tests with max-statistic FWER
correction, BCa bootstrap non-inferiority testing against a reference
condition with Holm correction, homogeneity/heteroscedasticity diagnostics)
replicated as an ASAREE feature so the use case's demonstration for an
upcoming bioinformatics journal application note reflects what ASAREE
actually does. That's a narrower, use-case-specific analysis path, not the
generic capability the vision describes.
This issue tracks building the generic nonparametric-regression capability
as a separate, follow-on feature once the use-case-specific path is done —
so the vision's "analyze results" claim is fully backed by a genuinely
reusable primitive, not just the one methodology one use case happened to
need.
See project_plan/core_asaree_use_case.md (ARES repo) §10 for the fuller
design-vs-analyze discussion this decision came out of.