Research development — experimental methods and evidence¶
Research site ≠ published release
The current published stable package is eyetrajectoriespy 1.1.0. These pages document unreleased 1.2 research from separate draft GitHub pull requests. The methods are not necessarily present in the installed stable package or in main, and statistically unqualified inference must not be used as validated hypothesis testing. Documentation publication does not merge research implementations or authorize a 1.2 release.
This research library preserves successful, inconclusive and negative results, original source figures, reproducible simulation evidence and open scientific gates. Green CI verifies engineering contracts and completion of documented computations; it is not a scientific inference guarantee.
Scientific evidence dashboard · Original figures · Reproducibility from source · Full source evidence matrix
Research programmes¶
| Programme | Research capabilities under investigation | Scientific boundary |
|---|---|---|
| F1 sparse group inference | Participant-level sparse-PACE group testing; 6,000 actual fitted tests | Tested null size encouraging; sensitivity to moderate effects limited; broader exchangeability and covariance assumptions unqualified |
| F2 gaze measurement quality | Target-based error, precision, sampling loss, coordinate/unit provenance | Target, participant/session and hardware traceability require further validation |
| F3 Eye-Tracking-BIDS | Schema checks and eye-tracking stream provenance | Official validator, real BIDS datasets and multi-stream interoperability pending |
| F4 constrained AOI FPCA | Compositional log-ratio and projection approaches | Independent constrained estimator and uncertainty not validated |
| F5 functional change-point inference | Block/AR(1)/AR(2) nulls, independently trained nuisance estimates and stress tests | AR(2), MA(1), baseline-size and nonstationary null failures; not a calibrated unknown-dependence test |
| F6 weighted functional geometry | Unit-aware weights and quadrature | Independent reconstruction, weighting and scale sensitivity pending |
| B5–B10 Bayesian layer | Diagnostic evidence, learned B6/B7 population covariance, research prediction and specialized models | Poor native covariance mixing, small posterior reference study and absent rank-based SBC/coverage qualification |
See F1–F6 method inventory, experimental API contracts, and B5–B10 development overview. None of these pages promotes new stable root-level APIs.
R86: independently auditing the secondary validator's failures (21 October 2026)¶
R85 established that outcome-selective recovery can produce a positive fitted first-stage missingness coefficient even when the original source is MAR. R86 freezes a third-stage independently randomized synthetic audit of secondary validation failures, using a new complete source design with 32 primitive independent generating blocks per fixed unit/support/variance cell. The original R86 science workflow #39639148930 passed 12/12 jobs, and separate original R86-A archive audit #39640120223 passed 4/4. Independently replayed 256 original X/Y sources, 512 first-stage MAR/MNAR masks and 4,608 source/audit/estimator records. With q=0, the audit-fitted secondary gamma/delta are NULL, not identified; positive q supplies genuinely new synthetic Y. Under adverse original MAR true delta=0, fitted first-stage means are approximately −.002/+ .003 curves and −.001/−.003 participants at 25%/50% audit, contrasting with no-audit naïve +.66. These numerical fits are not field-recovered gaze, source-independent uncertainty or validated functional bands.
Read R86 research · methods · full source ledger and figures. W3 #289/#296 OPEN; stable 1.1.0 unchanged; version 1.2/PyPI RELEASE HOLD.
R85: independent falsification of selective external validation (21 October 2026)¶
R84's extra missing response information required an independent recovery law. R85 now falsifies that fragile boundary: a second, outcome-dependent validation ascertainment mechanism can produce spuriously positive first-stage selection estimates even when original gaze loss is MAR. 256 genuinely new complete X/Y synthetic source conditions, 512 paired first-stage MAR/MNAR masks, and 2,560 validation-weighting strategy evaluations were independently reproduced, using 32 new primitive source blocks per fixed cell. Original R85 scientific source #39629520741 12/12 PASS; actual independent R85-A source #39630863879 4/4 PASS. In the adverse original MAR source, naïve latent-selective recovery yields mean δ +.664 curves / +.630 participants, despite true first-stage δ=0. With original MNAR δ=1.1, naïve latent-selective recovery yields means 1.671 /1.523, whereas simulator-only true validation propensity yields 1.028 /.950. No real missing-gaze recovery instrument or 90% functional inference established.
R85 source-grounded research article · Secondary ascertainment methods · Evidence and original figures. Issues #289/#296 remain open, stable 1.1.0 protected-main inference unchanged, version 1.2/PyPI release HOLD.
R84: independent recovery of missing synthetic Y and conditional selection identification (21 October 2026)¶
R83 found that varying assumed δ on the same observed sample does not identify the missingness mechanism. R84 newly asks whether genuinely additional missing-response measurements help estimate one prespecified outcome-dependent selection parameter. Thirty-two NEW independent primitive blocks per fixed design/support/variance cell generated 256 complete X/Y conditions, 512 paired MAR/MNAR masks and 1,536 0%/25%/50% randomly validated source records. Original science workflow #39581392688 12/12 PASS, independent R84-A exact real archive #39582618020 4/4 PASS. At zero recovery δ is not estimated; at positive recovery MAR source means near 0 and MNAR means near 1.0 (true generator 1.1), but error is not universally improved by more recovery and the visible-only pilot remains model dependent. This does not imply feasible gaze remeasurement or valid functional 90% bands.
R84 full scientific article · Identification methods companion · Original numeric evidence/plots. Issues #289/#296 stay OPEN, main inference code unchanged, v1.2 release HOLD.
R83: independently audited nonidentified MNAR assumption sensitivity (21 October 2026)¶
R83 freezes five externally supplied outcome-dependent selection coefficient assumptions δ=-1.1, -.55, 0, .55, 1.1 on the same visible source-mask data, with a separately labeled simulator-only true-propensity oracle unavailable to field analysts. The 32 NEW independent generating blocks per fixed condition yield 256 X/Y source conditions, 1,024 paired masks and 3,328 analysis configurations, not 3,328 independent experiments. Original science #39575847765 12/12 PASS, independently coded R83-A actual original #39576738539 4/4 PASS. In the adverse MNAR condition, δ endpoint slope spread on jointly estimable grid points is .00453 curves / .00841 participants, but no MNAR parameter is identified; no general missingness-adjusted 90% confidence method follows.
Read R83 source-grounded study · Observed-only assumption methods · Original figure and immutable evidence ledger. Issues #289/#296 OPEN, protected main estimator unchanged, v1.2/PyPI release HOLD.
R82: independently audited observation-process feasibility (21 October 2026)¶
The next prospective falsification now changes the observation process, not the complete-source error covariance. Thirty-two new independent primitive source blocks per condition generate 256 true Student-t5+skew synthetic X/Y source conditions, each tested under paired complete, MCAR, recorded-input MAR and hidden-outcome-dependent MNAR masks: 1,024 source observation-process evaluations. Stable public regression still refuses NaNs. A separate descriptive available-case pointwise OLS demands n≥8 independent units and surviving both ±4 predictor anchors at every time/coordinate location; it does not claim any band or correction. Original source workflow #39571308151 12/12 science PASS and independent R82-A #39572315409 4/4 actual original archive PASS. In the two-anchor curve heteroskedastic MNAR source, 83.1% sample retention coexists with 33.56/62 refused grid positions. Numerical evidence is verified but cannot justify 90% MNAR-adjusted bands.
R82 full scientific article · MCAR/MAR/MNAR methods guide · Original source and three plots. W3 #289/#296 OPEN, protected main inference unchanged, v1.2 release HOLD.
R81: independent source-audited rare skewed unit contamination (20 October 2026)¶
The R81 protocol froze a combined unit-wise shared Student-t5 radial and centered rare Bernoulli asymmetric event (probability .08, amplitude .16, signed time profile), while keeping original W3, residual HC3 and previously fixed λ=.25/.50 covariance regularization unchanged. The known generating reference was correctly recomputed as a source-conditional non-Gaussian t5 plus skew mixture, with true variance adjusted for the Bernoulli shock. Original R81 #39490888383 12/12 science PASS; independent R81-A #39491187099 4/4 real source replay PASS, verifying 256/256 source cases, 0 refusals and all 1,278,720 raw maximum arrays. Adverse two-extreme-unit heteroskedastic whole 31×2 slope inclusion under frozen λ=.50 was 24/32 independent curves and 24/32 repeated participants, still not general nominal-90%-calibrated inference.
Read scientific results and plots · Methods explanation · Original numerical source ledger. W3 blockers #289/#296 OPEN, protected main estimator unchanged, release 1.2 HOLD.
R80: independently audited unit-level Student-t5 heavy-tail stress (20 October 2026)¶
R80 prospectively preserved the R79 Gaussian conditional covariance and the two frozen source-observable HC3 covariance regularizations while adding one Student-t5 radial scale per independent observation unit, shared across all 62 grid coordinates and participant trials. The source-known comparison was correctly changed to a non-Gaussian independent-unit conditional mixture, not a Gaussian reference. Original science #39487449656 passed 12/12 and independent R80-A #39487811433 passed 4/4 actual-original-archive jobs, verifying 256/256 source cases, zero refusals, 1,278,720 conditional maxima. In the adverse two-anchor design, λ=.50 whole-function inclusion was 23/32 independent curves, 24/32 repeated participants, not nominal-90%-qualified inference.
Read the R80 scientific article and original plots · Student-t source methods companion · Exact original evidence ledger. Inference blockers #289/#296 remain OPEN, v1.2 RELEASE HOLD.
R79: independently audited nonseparable coordinate covariance stress (20 October 2026)¶
Original R79 changed the Gaussian generating law's cross-coordinate correlation to q(t)=.15+.65t while keeping trial temporal AR1 rho .72, within-trial time-varying scale, and the two fixed HC3 regularization candidates unchanged. The original experiment passed 12/12 scientific jobs; the mandatory R79-A independent source replay passed 4/4, independently verifying 256/256 source cases, zero refusals and all 1,278,720 raw reference maxima. Under adverse two-anchor heteroskedasticity, λ=.50 attained 24/32 full-grid curve inclusions and 20/32 repeated-participant inclusions, still well below nominal 90%.
Read the R79 methods/results article · Understand nonseparability · Open all three original figures. Inference #289 and missingness #296 remain open, version 1.2 unreleased.
R78: independent temporal covariance falsification and three original plots (20 October 2026)¶
R78 prospectively held both previously tested HC3 covariance-shrinkage strengths fixed and replaced within-trial independent Gaussian innovations with time-varying Gaussian AR(1) (rho .72). The original experiment passed 12/12 science jobs and the separately implemented R78-A passed 4/4 actual-source audit jobs, independently reproducing 256 original source cases, zero refusals and 1,278,720 maximum-statistic values. In the two-anchor predictor-heteroskedastic setting, full-grid inclusion under frozen λ=.50 was 22/32 independent curves and 19/32 repeated-participant datasets. This is not generally nominal-90%-calibrated inference, and it says nothing about MNAR gaze loss.
Read R78 original study · Read independent source-audit interpretation guide · View the original vector figure ledger. Blockers #289/#296 remain open; no inference code or stable package release change.
R77: independent source-audited covariance regularization stress (20 October 2026)¶
R77's prospective HC3 covariance shrinkage research retained 256 fully observed new Gaussian source cases (32 independent primitive generating blocks per condition), with 12/12 original scientific jobs passed and 4/4 independent original-source audit jobs passed, reproducing all 1,278,720 conditional maximum-statistic values. Two fixed λ=.25/.50 candidates improved observed slope truth inclusion under extreme leverage and heteroskedasticity but did not establish calibrated nominal-90% inference. The original W3 issue #289 and response-dependent loss issue #296 stay open. View the two original R77 scientific figures.
New W3 research articles and original evidence plots (20 October 2026)¶
Numerical verification is not interval calibration. R71–R75 independent source audits expose substantial undercoverage of nominal-90% full-grid functional slope bands under high leverage and heteroskedasticity. R76 and R76-A independently reconstructed 512/512 complete source cases and 1,534,464 maximum-statistic values; increasing independent extreme predictor support coincided with improved estimated-HC3 whole-band truth inclusion in the tested Gaussian setting, but does not yield a qualified 90% procedure.
Why W3 simultaneous bands under-cover · R76 leverage support and lower-tail SE evidence · Original W3 figure and source-audit gallery
Research blockers #289 and #296 remain open. Their research implementations are separate draft PRs; publishing these articles and static SVGs to the website does not merge an estimator or release version 1.2.
New workflow-level scientific evidence¶
- W3 independent known-truth functional regression benchmark: 24 generating blocks, 48 workflow fits, independent OLS-oracle agreement within numerical precision, transparent participant-versus-trial accounting, and original source-derived recovery SVG. This is an implementation/recovery validation under simulated designs, not inferential calibration or authentic data.
Evidence worth reading first¶
- F1 — 6,000 full sparse-PACE null/power fits: failure-inclusive scenario results and a limitation-oriented power assessment.
- B6/B7 — independent PyMC/NUTS posterior: identifiable mean and covariance diagnostics, raw loading sign ambiguity and inadequate native ESS.
- B6/B7 — nine-fit longer-chain Gibbs versus independent NUTS: source-verified B6, B7 paired and B7 asynchronous covariance mixing diagnostics, original figure and failure-inclusive reference evidence. All native covariance chains fail exploratory mixing; both B7 NUTS runs pass the diagnostic screen without establishing nominal coverage.
- B6/B7 — six four-chain independent NUTS references (R1): two fresh datasets per design; all identified population diagnostics passed a prespecified exploratory mixing screen, with zero divergences, and original evidence/figure retained. Rank-two validity, simulation-based calibration and posterior interval coverage remain unqualified.
- B6/B7 — rank-two independent Gaussian/Woodbury likelihood identity: paired, asynchronous, rotation and near-tied dense Gaussian/Woodbury identities verified; this mathematical prerequisite alone did not check the PyMC graph.
- B6/B7 — actual rank-two PyMC full log posterior and gradients: four tested mathematical configurations pass independent dense-log-posterior and 30-gradient numerical checks; actual NUTS sampling, rank SBC and interval coverage remain unqualified.
- B6 — first independent-NUTS prior-SBC pilot: four independent datasets, eight primary truth ranks, one diagnostic-screen failure, and original source visualization; formal SBC uniformity and interval coverage remain unqualified.
- B6 — full-budget refit of the same problematic SBC dataset (R2): all six population-functionals passed after longer sampling on the identical dataset (worst R-hat improved from 1.0158 to 1.0053), but this is not an additional independent SBC replication.
- B6/B7 — first actual rank-two NUTS paired versus asynchronous pilot (R3): two actual four-chain posterior fits, paired screen pass and asynchronous mean-functionals screen failure (6/15); graph correctness does not establish rank-two computational reliability or posterior coverage.
- B6/B7 — joint-loading elliptical-slice experiment: a valid research transition that failed to improve covariance mixing consistently.
- F5 — independent-baseline unknown AR(1), training baseline sensitivity and AR(2)/MA(1)/nonstationarity falsification.
- Research figures: original checked-in scientific SVGs, including negative results.
Qualification and governance¶
Currently blocked: Bayesian population uncertainty #255 and F5 dependent-null inference #256. Other F/B/D methods remain under independent scientific review. A green branch matrix must not be reinterpreted as approval to merge experimental estimators, close scientific issues, or release 1.2.
Evidence policy: Every numerical statement should identify the underlying GitHub workflow, dataset count, unit of independence, failure count, comparison condition and interpretation limits. Simulation evidence is not a substitute for independent eye-tracking recordings, and illustrations derived from simulations must be labelled accordingly. See the scientific dashboard for original artifact IDs and links.