Scientific calibration extension: B6, F1 and F5 (unpublished)¶

This validation tranche builds on B6 draft PR #243. It is not a scientific qualification decision, a version qualification, or permission to publish 1.2. The 1.1.0 stable implementation and frozen root API are unchanged.

Evidence lineage¶

These are different workflow matrices. The source-head success records do not transfer to the new follow-on branch or to inferential qualification.

B6: prior-generated calibration and stress¶

python scripts/run_b6_population_calibration_grid.py --replicates 8 --draws 60 --warmup 120 --out build/scientific-b6

Three matched-prior/likelihood scenarios are: rank-1, rank-2 and rank-2 with heterogeneous irregular observation densities. Here the mean/loadings are drawn from precisely the Gaussian priors assumed by the sampler, the participant scores are standard Gaussian, and observation errors are sampled under the assumed iid Gaussian likelihood. Every replicate refits the full learned-population Gibbs procedure.

Three additional scenarios are model stress, not SBC: an exactly near-tied rank-2 population eigenspace on the declared evaluation grid, a true noise SD twice the fitted SD, and a true loading-prior SD twice the fitted SD. They must not be pooled with matched calibration.

Assess the 90% pointwise marginal posterior intervals for the population mean and population covariance diagonal at a preregistered time location. Report coverage conditional on successful fits and the number of attempted/failed fits. Retain a Monte Carlo interval across independent simulation repetitions; a point estimate from 2 or 8 runs is not evidence for acceptable nominal coverage.

A rotation-invariant subspace projector error is calculated from the population covariance eigenspaces. An exactly tied rank-2 eigenspace does not identify two individually ordered eigenfunctions; report the subspace rather than arbitrary loading signs/rotations.

One wholly new, untrained participant is generated per replicate. Posterior predictive marginal 90% intervals incorporate population covariance and observation error, evaluated at irregularly selected grid nodes. The five observations from that one participant are correlated; the pooled observation-level coverage tally has no binomial Monte Carlo CI and cannot be called joint trajectory coverage.

Current B6 limitations remain: fixed rank, fixed spline basis, fixed noise, fixed prior scales; no participant/trial hierarchy, no learnt eigenfunction identity, no registration, and no definitive chain mixing or SBC-rank study. An executable two-chain fit is not convergence evidence. A separate long-chain, repeated-seed Rhat/ESS/rank-SBC programme is still required.

F1: repeated full sparse refits¶

python scripts/run_f1_f5_scientific_stress.py --f1-replicates 8 --f1-permutations 199 --f5-replicates 25 --f5-bootstrap 199 --out build/scientific-f1-f5

All F1 trials invoke the actual sparse-MFPCA pooled-score permutation pipeline, not the isolated score-only permutation kernel. The scenarios include balanced groups, unequal allocation with heteroscedasticity, repeated participant curves with intact unit resampling, and low sampling density. Each scenario has a null and a known effect. Fit failures remain in f1-cases.csv; rejection fractions conditional on fitting are not unconditional false-positive rates when missing fits are selective.

The unequal-heteroscedastic null deliberately violates the strict permutation-exchangeability assumptions and must be treated as a failure stress, not as a null under which exact randomization inference is expected.

F5: weak dependence and block-length sensitivity¶

For independent curves and heavy-tailed iid curves, run independent-mode ordered whole-function CUSUM. For stationary AR(1) Gaussian ordered curves, run weak-block resampling at block lengths 2, 4 and 8 under moderate and strong dependence. The generated AR process has stationary marginal variance matched to the independent innovation scale. Assess no-break, single-break and two-break data-generating scenarios.

The current estimator reports a single best split. Two-break simulation is an adversarial case, not implementation or verification of multiple-break estimation. Repeated-patient/trial dependence, evolving variance and genuinely nonstationary nulls still need dedicated modelling.

Replication plan and interpretation¶

A first engineering CI run uses only two replicates per design to check execution, output shape and failed-case retention. It provides no size, coverage or power qualification. Before a scientific decision, preregister a fixed full grid, effect magnitudes, seeds, tolerances, long-chain diagnostic criteria and a computational budget. As an initial planning target, use at least 1,000 independent null replications per critical F1 and F5 null scenario, 500 alternative replications for power, and at least 400 model refits per matched B6 scenario, then expand as required by Monte Carlo uncertainty. These are suggested study sizes, not completed experiments or automatic passing criteria.

Model comparison must use independent truth-based baselines, including native PACE and appropriately isolated external methods, not merely compare plots from the same sampler. Quantify Monte Carlo uncertainty, multiple prespecified alpha thresholds, convergence failure and fit failure. Use a versioned analysis decision record before setting any scientific_inference_qualified flag to true.

The follow-on workflows archive cases.csv, evidence.json and sha256.txt. All publication interlocks remain false. Passing CI is a necessary engineering checkpoint, not a statistical qualification gate.

Pending B7 modelling¶

A genuine Bayesian planar MFPCA model must learn shared x/y loadings, cross-covariance and channel-specific observation error from separately observed irregular timestamps. Do not relabel the currently available fixed-population paired/asynchronous score benchmark as full B7. A separate implementation and qualification programme is required.