Measurement accountability: latency, event plausibility, and validation ladders

eyeprocesspy.measurement_accountability_11 adds three diagnostics that make uncertainty and validation claims explicit without replacing the package's existing pupil, timebase, multimodal, or grouped-validation engines.

Pupil latency sensitivity

pupil_latency_sensitivity() returns a sustained-threshold onset, maximum-slope tangent onset, and piecewise breakpoint together with estimator spread, robust baseline noise, signal-to-noise ratio, and a small parametric resampling audit under the observed sampling regime. The result includes a latency_resolvability label rather than presenting one latency as algorithm- or hardware-independent.

from eyeprocesspy.measurement_accountability_11 import pupil_latency_sensitivity

result = pupil_latency_sensitivity(
    time_s,
    pupil,
    event_time=0.0,
    baseline_window=(-0.5, 0.0),
    search_window=(0.0, 2.0),
    simulations=500,
)
print(result["estimates_s"])
print(result["estimator_spread_ms"])
print(result["latency_resolvability"])

This is an independent, transparent sensitivity harness. It is not a verbatim reproduction of any published implementation.

Event-marker plausibility is not clock synchronization

event_marker_qc() evaluates whether independently estimated channel offsets corroborate a nominal event. It returns confirmed, plausible, ambiguous, or implausible, a consensus offset, and robust uncertainty. It never changes timestamps and explicitly records that no clock-drift correction was applied.

from eyeprocesspy.measurement_accountability_11 import event_marker_qc

qc = event_marker_qc([0.012, 0.018, 0.016], tolerance=0.050)

Use the package's timebase/alignment tools for synchronization and drift correction; use this diagnostic to audit event/annotation plausibility or conduct timing-uncertainty sensitivity analyses.

Validation ladder

validation_ladder() separates five evidence stages:

1. acquisition QC;
2. analytical QC;
3. construct check;
4. within-person evidence;
5. held-out-person generalization.

A generalization claim is not_supported unless the held-out-person stage passes. This prevents personalized/within-participant calibration performance from being reported as out-of-person generalization.

from eyeprocesspy.measurement_accountability_11 import validation_ladder

ladder = validation_ladder(
    {
        "acquisition_qc": "pass",
        "analytical_qc": "pass",
        "construct_check": "pass",
        "within_person": "pass",
        "held_out_person": "not_assessed",
    },
    claim="generalizable",
)

Methodological provenance

These additions were motivated by the September 2026 measurement-methods surveillance tranche, especially work on pupil-latency benchmarking, multimodal construct-validation ladders, and physiological validation of uncertain timestamps. Relevant primary references include DOI 10.1038/s41598-026-68921-9, DOI 10.3389/fnrgo.2026.1911259, and DOI 10.1007/s12028-026-02637-6.