Research evidence matrix — unpublished¶
Current research findings and original artifact links (10 October 2026): use the scientific evidence dashboard and original SVG figure gallery. The chronological matrix below preserves earlier source-level evidence and may describe prior branch checkpoints. It is not a statement that research estimators are installed from PyPI or merged into protected main.
Stable release versus prototypes
Stable 1.1.0 is available from PyPI. Research 1.2 (source 1.2.0rc2.dev0) is unpublished. The 27/27 engineering workflow success refers exclusively to PR #227 commit d48c939b01fbfa19f0e060902b6a61ebddbe7aa6. It does not certify newer commits or scientific inference.
Further exact-source scientific findings — 9 October 2026¶
Scientific qualification remains failed/unestablished, irrespective of green CI.
- B6/B7 exact conditionals: PR #257 independently constructed dense observation-row Gaussian reference posteriors; loading/mean/asynchronous shared-score conditional precision and mean tests passed on Python 3.11, 3.12 and 3.13. This is conditional-update algebra, not proof that the whole Gibbs posterior mixes or has nominal coverage.
- B6/B7 scale interweaving: source run #37976468861 completed 72 matched-prior genuine Gibbs refits with no exceptions. A mathematically reversible likelihood-invariant MH factor-scale proposal had no consistent R-hat or ESS improvement, and most rank-2/asynchronous comparisons deteriorated. The implementation remains opt-in and unqualified. Detailed results.
- F5 fresh independent nulls: source study #37975560959, SHA256 recovered in #37976469318, produced 1,800 real hypothesis tests, with no fitting exceptions and no recovery refits. At alpha .05, true phi=.8 null rejection was 99/200 = 49.5% for old block-4 but 16/200 = 8.0% for the reference using known oracle phi. This is an improvement in one conditional simulation regime, not a calibrated test for unknown serial dependence. Detailed model and limits.
- F1 higher-precision study completed: draft #259 completed an independent, full native sparse-PACE 4,000 null + 2,000 moderate alternative simulation programme (original complete source run #37977625803, 5/5 study jobs green; original combined artifact 11640464273), zero fit exceptions. At α=.05 the null rejection rates were 4.9%, 5.1%, 5.2%, 4.4% (1,000 independent datasets each) and moderate-effect detection 15.8%, 18.2%, 19.2%, 14.2% (500 datasets each). This suggests nominal size in the tested configurations, not universal validity or useful general power.
- Bayesian partially collapsed mean: draft #260 implemented opt-in exact score-marginalized B6 and joint x/y cross-channel B7 mean conditionals. Its original complete science run #37978244822 passed all 4 numerical/scientific jobs and 72/72 paired fits, zero failures; B7 rank-1 asynchronous median worst monitored R-hat improved 1.470→1.067 and minimum ESS 4.07→15.35, while B7 rank-2 paired remained poorly mixed (1.937→1.537, ESS 2.92→3.81). No adequate full posterior convergence or 90% coverage certification. The original sampler remains default. Mathematical protocol.
The earlier B6/B7 blocker #255 and F5 dependence blocker #256 stay open. Stable 1.1.0 is unchanged; 1.2 research and its source pages are not a production release or independently deployed public website.
| Route | Opt-in API | Engineering | Scientific status |
|---|---|---|---|
| F1 sparse independent groups | test_sparse_functional_groups() | Original #227 27/27 | Null size and exchangeability not qualified |
| F2 validation targets | summarize_gaze_validation_targets() | Original #227 27/27 | Descriptive, independent target provenance |
| F3 Eye-Tracking-BIDS | from_bids_eyetracking() | Original #227 27/27 | BIDS 1.11.2 narrow subset, not official conformance |
| F4 AOI geometry | compare_aoi_functional_geometries_holdout() | Original #227 27/27 | Not a constrained eigenfunction estimator |
| F5 whole-curve changes | detect_ordered_functional_changepoint() | Original #227 27/27 | Bootstrap p-values unqualified |
| F6 weighted L2 | fit_weighted_mfpca() | Original #227 27/27 | Declared weights, no optimal weighting |
| D4 sample-size/power | simulate_functional_study_power() | PR #235 exact-head 27/27 green (013cd05d) | Actual F1 experimental test; no certified power |
| D4 repeatability | fit_functional_reliability() | PR #235 exact-head 27/27 green (013cd05d) | Balanced moments, no generic functional ICC |
| D4 paired groups | compare_repeated_functional_groups() | PR #235 exact-head 27/27 green (013cd05d) | Sign-symmetric participant differences only |
| D2 quality linkage | link_gaze_validation_sessions() | PR #235 exact-head 27/27 green (013cd05d) | Identifier match, no clock or drift proof |
| D2 multi-stream BIDS audit | audit_bids_eyetracking_dataset() | PR #235 exact-head 27/27 green (013cd05d) | Not official bids-validator |
Open gates: F1/F5 full type-I error/power/serial-dependence calibration; F2 external validation evidence and session timing; F3 official BIDS conformance and external datasets; F4 constrained AOI eigenspace; F6 weight sensitivity; D4 reliability uncertainty and crossover modelling; D5 fully qualified examples and docs.
Method finder · Method contracts · Case studies
Retained PR #227 exploratory simulation data (not qualifying)¶
An exact-head workflow run #37870154333 retained the following 100-replication synthetic diagnostic results at alpha = 0.05:
| Pilot | Null rejection | Detection under simulated alternative | Methodological limitation |
|---|---|---|---|
| F1 isolated pooled-score permutation kernel | 0.01 | 0.77 | No sparse-PACE refit on each replication |
| F5 independent whole-curve CUSUM | 0.06 | 1.00 | Single independent-generating process and shift |
| F5 AR(1) weak-block CUSUM | 0.01 | 0.99 | One serial-dependence strength/block setting |
The F1 native-pipeline pilot completed only two refits per each of three regimes, all fitted successfully. The F4 held-out feasibility produced 144 sensitivity records. None constitutes scientific calibration. With only 100 simulation replications the null rejection estimates have wide Monte Carlo uncertainty; the 0.01 estimate has approximately 0.002–0.054 Wilson 95% range, and the 0.06 estimate approximately 0.028–0.125. The target nominal 0.05 is not ruled out by these pilots, nor proven by them. Larger design-gridded refit experiments remain mandatory.
New scientific research development (staged, not published)¶
The original F1–F6 PR #227 completed 27/27; stacked D1–D5 PR #235 completed 27/27; Bayesian foundation PR #237 completed 32/32; and native learned B6 PR #243 completed 27/27 on their separate exact GitHub heads. Draft #244 also completed its 28/28 exact-head engineering workflows. It adds six-scenario learned B6 population calibration and full-fit F1/F5 stress runners. Its short engineering run completed 12 B6 refits, 16 F1 full sparse refits, and 48 F5 stress evaluations without execution failure. None of those small-run totals certifies coverage, nominal null size or power.
Draft #245 introduces the learned joint x/y Gaussian latent functional factor posterior on native paired/asynchronous observations. The component columns remain sign/rotation unidentified; joint covariance is learned, but fully calibrated eigenfunction inference is not established. Its strict documentation contract was repaired on a later head; retain CI status by exact head, not by assuming older green matrices apply automatically.
The further B7/B9 research extension predicts genuinely new participants
using learned population posterior draws, a fresh shared x/y latent factor
and channel-specific observation noise. It does not borrow training
participants' fitted scores. See new participant prediction
and the B6/F1/F5 protocol.
Prediction intervals are pointwise, not jointly calibrated functional
bands. CI passing and synthetic pilot success never change
scientific_inference_qualified or publication interlocks.
Reproducible large-sample calibration preparation¶
The independent-shard research procedure supports versioned, seed-separated B6/B7/F1/F5 simulations, audits checksums, rejects repeated scenario/replicate keys and observed random-seed collisions, and preserves failed fits. See calibration batches. The availability of a repeatable study framework does not establish actual large-sample simulation evidence or scientific qualification.
Actual 13,600-attempt scientific calibration (9 October 2026)¶
Large empirical simulation did not qualify Bayesian or dependent F5 inference
Original full simulation #37965949039 performed 13,600 native fitted-model/test attempts in six independent shards with zero fitting exceptions. Bayesian B6 and B7 aggregate jobs passed; F1/F5 original aggregate had an evidence-schema defect, then recovery run #37969068128 verified and recovered all 9,600 archived F1/F5 tests without refitting models. This evidence is not 1.2 release authorization.
- B6 (400 genuine refits per scenario): under matched priors, nominal 90% population mean/covariance intervals covered 70.75%/77.00% (rank 1) and 62.00%/67.00% (rank 2), with zero exceptions.
- B7 (400 refits per scenario): paired rank-1 x/y mean intervals covered 61.50%/67.00%, joint x/y cross-covariance 74.00%; asynchronous rank-2 x/y means covered 57.50%/59.50%, cross covariance 67.75%. Held-out person average pointwise predictive inclusion was 86.41%–87.59%, not a simultaneous band. Severe undercoverage cannot be attributed to implementation versus poor MCMC mixing until long-chain diagnostic #253 and independent reference comparisons finish.
- F1 (300 attempts per truth/design): α=.05 null rejection 4.33%–7.67%; no nominal-size qualification or general-power claim. Alternative detection 98%–99.33% under the deliberately strong simulated effect.
- F5 (300 per scenario, α=.05): true strong AR(1) stationary no-change false rejection was 100% (block 2), 42.67% (block 4), and 2.33% (block 8). Dependence-calibrated p-values fail under the tested regimes and cannot be recommended as nominally calibrated.
See complete design-by-regime scientific evidence and governance issues Bayesian undercoverage #255 and F5 null-size failure #256. Scientific inference, stable API, deployment and publication gates stay false.
B5–B10 native Bayesian research track¶
Draft PR #237 adds experimental backend-independent posterior tools, a known-population conditional sparse score baseline (not a learned Bayesian FPCA estimator) and a fixed-noise Gaussian B-spline functional regression prototype. No architecture replacement, inference calibration or publication decision is established. See the Bayesian research catalogue and synthetic fitted-model gallery.
The B8 experimental Bayesian functional random-intercept Gaussian conjugate model (fit_bayesian_functional_mixed_effects()) accepts repeated participants, but conditions on fixed noise and random-effect prior SDs; no random slopes, nested trial effects, serial residual model or learned hyperparameter inference is yet qualified.