Plot gallery¶

A scientific figure should make the interpretation boundary visible as well as the result.

Research lifecycle¶

GazeForge research lifecycle

Use for: understanding where source preservation, QC, review, analysis, validation, and reporting sit.

Do not infer: that completing every stage establishes external or construct validity.

Evidence ladder¶

Evidence ladder from observed data through review and validation to bounded scientific claims

Use for: showing that observations, diagnostics, review decisions, analysis/validation evidence, and scientific claims are distinct layers.

Do not infer: that a higher layer is automatically justified because a lower layer exists.

QC diagnostics¶

Synthetic gaze quality-control diagnostics

Use for: identifying quality patterns requiring interpretation or review.

Do not infer: that a QC flag is automatically an exclusion.

Event diagnostics¶

Synthetic event diagnostics

Use for: inspecting event segmentation behaviour.

Do not infer: that a visually plausible threshold is universally valid.

AOI scanpath¶

Synthetic AOI scanpath

Use for: communicating reviewed spatial or semantic sequence structure.

Do not infer: latent cognitive states from sequence geometry alone.

Dynamic AOIs¶

Synthetic dynamic AOI example

Reviewed temporal support and bounded interpolation for dynamic AOIs

Use for: distinguishing reviewed temporal support from bounded interpolation.

Do not infer: that extrapolated boxes outside reviewed support are validated AOIs.

Validation design¶

Event validation workflow

Use for: explaining why grouping and held-out unit identity belong in the validation claim.

Benchmark evidence¶

Lund2013 derived-60 evidence¶

Lund2013 derived 60 Hz benchmark performance

Hollywood2EM derived-60 evidence¶

Hollywood2EM derived 60 Hz benchmark performance

Do not infer: that derived-rate evidence is equivalent to native-device evidence.

Location-scale modelling¶

Location-scale workflow

Use with hierarchical location-scale models, residual calibration, and bootstrap Monte Carlo precision.

Observation and denominator states¶

Observation states separating observed zero, missing, absent-by-design, undefined, and right-censored outcomes

Use for: explaining why zero, missingness, absent-by-design states, undefined quantities, exposure, and right-censoring must remain distinct in a statistical handoff.

Do not infer: that a convenient numeric encoding is scientifically interchangeable with the underlying observation state.

Use with Denominator, exposure & censoring, Missing-data assumptions, and Analysis handoff.

Provenance chain¶

Provenance chain from source bytes through transformations, review, analysis derivatives, software identity, and reporting

Use for: showing where hashes, transformation specifications, review ledgers, analysis derivatives, and software identity belong in a reproducible record.

Do not infer: that matching hashes alone establish measurement, construct, model, or external validity.

Use with Research evidence bundle, Reviewer & replication handoff, and Reproducible reporting.

Held-out validation boundaries¶

Matrix of participant, stimulus, source-token, and dataset held-out validation boundaries

Use for: matching the held-out unit to the population or source boundary that the validation claim is intended to address.

Do not infer: participant-disjointness from opaque source tokens, or universal validity from a single held-out dataset.

Use with Event-model validation clinic, Cross-dataset events, and Validation evidence guide.