Guide · Audit lineage
996102e0
Audit record map
A GazeAudit analysis is easier to review when the lineage of the research record is visible. This map shows how study inputs and researcher-owned scientific decisions connect to structural QC, declared specifications, execution evidence, robustness summaries, provenance, and the final publication record.
The record in seven stages
Preserve the canonical gaze table, coordinate system, units, participant/trial identity, and source provenance before analytical branching.
Write down the endpoint, exclusions, thresholds, uncertainty assumptions, specification factors, perturbations, and invalid combinations that are scientifically defensible.
Generate the study-QC record and investigate structural flags before they silently become scientific exclusions.
Run every valid predeclared specification and preserve technical failures instead of deleting inconvenient branches after outcome inspection.
Inspect specification curves, effect stability, and sensitivity summaries as empirical descriptions of the declared analysis space.
Record the exact software identity, manifests, source locks, scientific inputs/outputs, and bundle-level execution context needed to reconstruct the analysis.
Report only what the frozen protocol and preserved evidence support, including the uncertainty dimensions that were not tested.
The stages are ordered for traceability, not because every study needs every GazeAudit method. Use the smallest audit layer that answers the actual methodological question.
What moves through the record
Inputs establish the scientific contract.
canonical gaze table
coordinate system + units
participant / trial identity
endpoint definition
researcher-owned thresholds
declared specification factors
Start with the data-onboarding guide or the Audit planner.
The audit keeps the whole declared space.
study-qc/
specifications.csv
specification-curve.csv
effect-stability.csv
marginal-sensitivity.csv
pairwise-sensitivity.csv
Use the output-bundle explorer to inspect what each artifact describes and what it cannot establish.
Artifact → question → boundary
| Record layer | Representative artifact or record | Question it helps answer | Boundary to preserve |
|---|---|---|---|
| Study contract | canonical table + source identity | What data and units entered the audit? | Provenance does not establish data quality or causal validity. |
| Researcher decisions | endpoint, thresholds, factors, invalid combinations | What was scientifically declared before execution? | GazeAudit does not choose these decisions for the researcher. |
| Structural QC | study-qc/ |
Are there structural conditions that require investigation? | A flag is not a universal exclusion rule or validity judgement. |
| Specification execution | specifications.csv |
What happened across every valid declared branch? | The table does not identify one preferred specification. |
| Robustness summary | specification-curve.csv and effect-stability.csv |
How stable is the endpoint across the declared space? | Empirical quantiles and stability summaries are not confidence intervals, posterior probabilities, or causal evidence. |
| Sensitivity summary | marginal-sensitivity.csv and pairwise-sensitivity.csv |
Which declared analytical choices coincide with larger endpoint movement? | Descriptive sensitivity is not causal variance decomposition or mechanism identification. |
| Publication record | manifest, methods/report text, fingerprints, software identity | Can collaborators reconstruct what was run and reported? | Reproducibility does not expand the scientific scope of the claim. |
What may vary—and what should remain scientifically anchored
A robustness audit is useful only when the intended variation is distinguishable from accidental drift.
| Element | Typical role in the record |
|---|---|
| Scientific endpoint | Keep fixed across branches unless the protocol explicitly declares endpoint choice as a factor. |
| Specification factors | Vary only over defensible, declared alternatives. |
| Perturbation grid | Declare the tested uncertainty range before interpreting recovery. |
| Structural-QC policy | Record the rule and the researcher decision it triggers; do not convert a diagnostic into an automatic scientific exclusion. |
| Source identity | Preserve stable identifiers, checksums, manifests, or immutable retrieval metadata when material to the claim. |
| Software identity | Record the release or exact commit used for the analysis. |
| Reporting language | Match the evidence actually generated; do not upgrade descriptive robustness evidence into population or causal inference. |
From your first file to a publication record
Run the practical path
Take a canonical CSV through study representation, structural preflight, declared alternatives, and saved evidence.
Read the generated bundle
Inspect the purpose and interpretation boundary of each saved artifact before selecting figures or prose for a manuscript.
Bind the publication record
Preserve source identity, frozen choices, complete evidence, manifests, fingerprints, and exact software identity.
Use frozen evidence as a boundary check
GazeAudit’s own evidence map demonstrates why lineage matters. The three canonical cases reach different protocol-bound outcomes because they ask different questions under different frozen rules. Their labels are not transferable scores for new datasets.
- GazeBase preserves an
incompleteoutcome because a predeclared completeness prerequisite failed. - Korthals preserves
robust_negativefor the frozen AOI estimand under its measurement-error model. - Pedrotti/de Chambrier preserves
materially_fragilebecause the declared magnitude-and-direction recovery criterion failed under several perturbation families.
The practical lesson is not to seek one label. It is to preserve enough of the record that another researcher can see what was fixed, what varied, what failed, what was summarised, and what claim followed.
Fast routes
- Researcher workspace — project-level starting map.
- Audit planner — translate study conditions into governed method routes.
- First real audit — run the practical CSV-to-evidence path.
- Output-bundle explorer — inspect generated artifacts.
- Reproducible publication — build and verify the durable record.
- Evidence map — inspect frozen protocol-bound validation cases.