Skip to content

1.2 workflow product qualification

The 1.2 workflow release is qualified as an orchestration/product layer rather than as a new estimator programme. Scientific estimators composed by the workflows retain their earlier recovery, calibration, comparator and method-specific validation evidence.

Qualification design

The workflow programme is gated in four tranches:

  1. W1 — infrastructure: immutable configuration/result contracts, ordered step records, decision provenance, strict serialization and portable bundle export.
  2. W2 — sparse workflows: sparse FPCA, paired/asynchronous sparse MFPCA, sparse multilevel decomposition and sparse prediction.
  3. W3 — dense/regression workflows: common-grid FPCA, function-on-scalar, functional mixed effects and generalized functional regression.
  4. W4 — product qualification: recurrence orchestration, explicit workflow plotting, documentation/gallery integration, installed-wheel product analyses, performance envelope and final workflow API audit.

Each tranche must qualify on its pull-request head and then again on the exact protected-main merge commit before the next tranche begins.

Installed-package evidence

scripts/run_one_dot_two_workflow_product_qualification.py runs from an installed wheel outside the source checkout and exercises all ten workflow entry points on deterministic realistic synthetic eye-tracking fixtures.

The fixtures are not empirical human observations and are not used as known-truth recovery simulators. Their purpose is product integration: input → typed workflow → primitive result → audit records → reporting/tables → figures and reproducibility bundles.

The retained evidence includes:

  • a summary for every workflow contract;
  • decision and ordered-step tables;
  • generic workflow reporting;
  • representative dense, sparse and recurrence reproducibility bundles;
  • canonical workflow-generated figures;
  • runtime measurements;
  • a friction ledger with release-blocking status.

Performance envelope

The CI envelope is deliberately broad. It guards against accidental product regressions rather than claiming benchmark performance:

  • all ten workflows together must complete within 120 seconds on the qualification runner;
  • no single workflow route may exceed 60 seconds.

Dedicated estimator/performance workflows remain authoritative for method-specific computational claims.

API audit

WORKFLOW_API_AUDIT.json freezes the candidate 1.2 module surface and scientific policies. scripts/validate_workflow_api_audit.py requires exact agreement with eyetrajectoriespy.workflows.__all__ and rejects accidental root-namespace promotion.

The 1.2 candidate therefore keeps workflow APIs under eyetrajectoriespy.workflows; the established package root is not expanded by this release.

Release interpretation

Passing this qualification means that the supported workflows can be executed from an installed distribution with their declared scientific decisions, primitive outputs, reporting and provenance intact. It does not establish that one estimator architecture is superior to another and it does not relax any method-specific uncertainty or interpretation limitation.