Execution and statistical integration plan

Engineering decisions recorded on 2026-10-21. Keep NumPy as the core dependency and optional integrations behind extras. The shared execution roadmap and baseline protocol describe the numerical packages and representative measurements.

Preserve package responsibilities

trade-study owns generic replicate orchestration, scores, result tables, comparisons, optional surrogates and execution services. A simulator may use NumPy, JAX, Torch or CuPy internally while returning explicitly declared host results. This does not require making every runner/table operation GPU-native. Inference-specific priors, measurement models, sensitivities and posterior updating belong in OP Infer or a caller-supplied adapter.

No Rust rewrite or full Julia port is committed by this plan. Existing Julia repositories are inactive design scaffolds. Native array transfer and host serialization boundaries must remain explicit. Generic runners should retain failures and stable task identities regardless of a simulator's backend.

Fisher feasibility screening

The bounded Fisher adapter supplies a scalar result to existing constraint machinery and preserves unavailable outcomes. It consumes an explicit information artifact or caller-owned screen instead of adding a compiler/inference dependency to the core. The artifact records parameter coordinates/scales, observation noise, design, nuisance treatment, reference parameter regime and numerical tolerances.

Use measured-output parameter sensitivities, not the instantaneous RHS state Jacobian. Structural gauge freedoms need declared treatment; a known singular nominal Fisher matrix is not grounds for rejecting every otherwise useful design. Local numerical rank is a screen, not a structural-identifiability proof. Estimated screen uncertainty needs an actual sampling/replication contract before a probabilistic feasibility claim.

The implemented fixed-replicate allocation and finite-horizon bounded-score methods remain available. Recoverable failed-attempt artifacts and gate-specific power/false-failure qualification retain their separate scope. The Fisher guide also distinguishes SBC, effect recovery, predictive checks and simulation-support evidence; each needs its own declared reference and denominator.

Predictive stacking and dimension penalties

The existing Bayesian path delegates to ArviZ predictive stacking. Preserve that objective. Subtracting IDOF * log(n) from LOO scores is not an established correction and must not become an automatic behavior of stack_bayesian. Nominal parameter count, a reported gauge-quotient dimension, identifiable dimension and singular-model information penalties are distinct quantities.

A proposed penalized alternative first needs a named objective, normalization, unit of comparison, definition of sample size, assumptions and independent qualification. Identical predictive distributions should give the same ensemble prediction under duplicate representations; their individual weights can be nonunique. Deduplication/grouping is a separate policy from complexity penalties.

See predictive stacking and singular-model information criteria.

Surrogate acquisition and EVSI

Expected improvement over a current best score is a Bayesian-optimization acquisition target. Decision-theoretic expected value of sample information requires actions, utilities, a distribution of future observations and posterior updating before optimizing the action. A Gaussian surrogate alone does not make general EVSI analytic.

Keep those APIs and evidence separate. A generic acquisition may consume a surrogate, while OP Infer or the caller supplies statistical updating and utility semantics for EVSI. The current RF tree spread is a relative disagreement measure, not calibrated posterior uncertainty. Qualification needs analytic decision fixtures, retained simulation failures and Monte Carlo error before claims about decision value. See the EVSI formulation.

Sequence

The high-priority conditional table-sensitivity increment is implemented: pooled varying omitted inputs are refused, observed-stratum conditioning and retained scope/validation are available, and direct discrete contrasts reuse existing paired helpers. Configuration-factor grouping and explicit dataset/replicate partitions answer distinct validation questions. The Sobol budget records actual callback costs before consumers price another screen. Remaining statistical and solver extensions keep their separately qualified scope.

First reconcile overlapping issues and establish numerical-package baselines. Then integrate forward/provenance and shared statistical interfaces. Add Fisher screen adapters after sensitivity contracts are qualified. Keep penalized stacking and EVSI implementation gated on their separately specified targets. Forecasting consumers reuse trade-study execution and scoring rather than introducing a second replicate runner.