Reproduce experimental research — without installing it as stable¶

Evidence and environment¶

Each numerical study has a named GitHub Actions run, source head SHA, retained case-level evidence artifact and check/qualification ledger. When reproducing, record the exact original commit, Python version, optional dependency versions, random seeds, scenario and number of Monte Carlo replications. A current branch may contain additional commits, so a branch tip is not necessarily the original source used for a published number.

For inspectable historical case-level data without rerunning expensive simulation, use the scientific dashboard and the run's GitHub Actions Artifacts section. Artifacts are retained for a limited period and may expire; the permanent documentation preserves only audited aggregate results and source identifiers.

B6/B7 native mixing comparison¶

The experimental fitting and diagnostic code resides in draft PR #266, not in 1.1.0 or protected main.

git clone https://github.com/stefanosbalaskas/eyetrajectoriespy.git
cd eyetrajectoriespy
git checkout research/b6-b7-score-marginal-elliptical-slice-20261010
python -m pip install -e ".[dev,bayesian]"
python scripts/run_b6_b7_score_marginal_loading_pilot.py --design B6_rank1 --replicates 6 --draws 250 --warmup 350 --seed 20261206 --out build/reproduce-b6

This uses the current research branch, so check it against original workflow #38000101053 before comparing numerical values. It does not certify covariance convergence, scientific uncertainty or inference.

F5 independent AR(2) stress¶

The AR(1) vs AR(2) simulation code resides in draft PR #267.

git clone https://github.com/stefanosbalaskas/eyetrajectoriespy.git
cd eyetrajectoriespy
git checkout research/f5-independent-baseline-ar2-null-reference-20261010
python -m pip install -e ".[dev]"
python scripts/run_f5_independent_baseline_ar2_stress.py --scenario ar2_misspecified --null-reps 100 --alternative-reps 50 --n-null 119 --n-parameter 99 --seed 20261207 --out build/reproduce-f5-ar2

Interpretation: 100 true-mean-null simulations per tested generator yield wide Monte Carlo intervals; no automatic dependence-model selection or scientifically qualified unknown-dependence p-values follow.

F1 and other research¶

Use the research-only F1 6,000-fit study and source run #37977625803 for the exact script and artifact provenance. Experimental F1–F6 inventory, B5–B10 overview, model contracts and research case studies describe the remaining source-level workflows and assumptions.

Reporting template¶

Report data origin, recording geometry and coordinate units, independent sampling units, missingness, process assumptions, preregistered effect and null, estimator version/branch SHA, complete fit count with failures, Monte Carlo uncertainty, R-hat/ESS/MCSE when Bayesian, source run/artifact identifiers, exploratory vs qualified status, and explicit negative sensitivity results.

Real eye-tracking workflows require separate validation of sampling integrity, session/participant independence and ethical/data-permission constraints. A synthetic-only reproduction is not real-data qualification.

Do not relabel the research functions as stable 1.1.0 or turn on a 1.2 release gate when reproducing these demonstrations.