Example · Interpretation & reporting

Reporting-language rewrite

This exercise starts with seven synthetic over-claims and repairs each one by identifying:

1. the actual evidence object;
2. the denominator that must remain visible;
3. what the evidence can support;
4. what it cannot support;
5. a bounded rewrite;
6. the limitation that belongs next to the claim.

Rewrite 1 · Structural review is not “bad data”

Evidence

Assume structural preflight reports:

status = review
issue_codes = ("coordinate_nonfinite", "timestamp_duplicate")
coordinate_issue_rows = 4
duplicate_timestamp_rows = 6

Too strong

GazeAudit identified 10 bad observations that were removed.

Why it fails

• review is an inspection state, not an invalid-data classification.
• coordinate and duplicate counts can overlap;
• the denominator is not stated;
• removal is a researcher decision, not a package action;
• duplicate timestamps can be a valid representation feature.

Bounded rewrite

Structural preflight returned review because non-finite coordinates and duplicate participant × trial timestamps were observed. The corresponding row-level diagnostics were inspected against source and preprocessing provenance before any study-specific retention, repair, or exclusion decision.

Limitation

Structural review identifies representation-level conditions and does not establish their cause or scientific invalidity.

Rewrite 2 · ready_under_policy is not universal quality

Evidence

Assume:

policy_name = preregistered_primary
status = ready_under_policy
trial-scope retained = 96 / 100 trials
participant-scope retained = 48 / 50 participants

Too strong

The dataset passed GazeAudit quality control and was validated for analysis.

Why it fails

ready_under_policy means the observed units satisfy the declared active structural policy.

It does not test:

• calibration quality;
• event detection;
• AOI validity;
• missingness assumptions;
• endpoint validity;
• inferential model assumptions.

Bounded rewrite

Under the preregistered primary structural-readiness policy, the study returned ready_under_policy. Trial-scope preview retained 96/100 trial units, while participant-scope preview retained 48/50 participants. The primary application scope was fixed separately in the analysis protocol.

Limitation

Readiness status is policy-relative and does not establish broader measurement or inferential validity.

Rewrite 3 · unassessed is not “all included”

Evidence

Assume:

policy = ReadinessThresholds()
readiness.status == "unassessed"

Too strong

All participants met the readiness criteria.

Why it fails

There were no active criteria.

No readiness decision was made.

Bounded rewrite

The readiness report returned unassessed because the declared policy contained no active structural-readiness criteria.

Limitation

This state provides no evidence that participants would satisfy a different declared readiness policy.

Rewrite 4 · Stable direction can coexist with unstable magnitude

Evidence

Synthetic eight-branch estimates:

+0.006, +0.011, +0.019, +0.028,
+0.041, +0.057, +0.073, +0.091

All estimates are positive, but the magnitude varies substantially.

Too strong

The effect was robust across all analytical choices.

Why it fails

The sign is stable, but the magnitude changes by more than an order of magnitude.

A single “robust” adjective hides that distinction.

Bounded rewrite

All eight declared synthetic specifications produced positive estimates, but magnitude ranged from +0.006 to +0.091. Direction was therefore less sensitive than quantitative magnitude within the declared specification space.

Limitation

Whether that magnitude spread is scientifically material requires an independently justified substantive scale and does not follow from sign stability alone.

Rewrite 5 · Six successful branches do not become a six-branch denominator

Evidence

8 declared
7 valid
6 successful
1 valid technical failure
1 invalid before execution

The six successful estimates are positive and tightly grouped.

Too strong

All valid analyses produced consistent positive effects.

Why it fails

One valid branch did not produce an estimate.

The correct execution denominator is 6 successful / 7 valid, not 6/6.

Bounded rewrite

Six of seven valid specifications produced endpoint estimates and all six available estimates were positive; one valid branch failed technically. The available estimates were directionally consistent, but the declared valid audit remained incomplete.

Limitation

The unobserved branch estimate is unknown and cannot be treated as positive, null, or irrelevant.

Rewrite 6 · Sensitivity is descriptive, not causal attribution

Evidence

Assume marginal_sensitivity() returns its largest ratio for AOI definition and the largest pairwise non-additive pattern for detector × AOI definition.

Too strong

AOI definition caused most of the uncertainty, and detector choice interacted significantly with AOI definition.

Why it fails

The GazeAudit diagnostics are descriptive properties of the realised specification set.

They are not:

• causal variance decompositions;
• shares that must sum to one;
• inferential interaction tests;
• p-values or significance tests.

Bounded rewrite

Descriptive marginal sensitivity was greatest across AOI-definition levels, while the strongest pairwise non-additive specification pattern involved detector × AOI definition.

Limitation

These diagnostics localise variation within the declared specification set and do not establish causal attribution or inferential interaction effects.

Rewrite 7 · Frozen validation labels do not transfer

Evidence

The frozen Korthals protocol has canonical outcome:

robust_negative

Too strong

Negative AOI effects are robust in GazeAudit, so our negative effect can be classified as robust_negative.

Why it fails

robust_negative is a protocol-bound label for one frozen validation record.

It is not a reusable result category.

Bounded rewrite

Under the frozen Korthals target-tracking AOI protocol, the canonical validation outcome was robust_negative. That record provides protocol-specific validation context and is not used as a classification rule for the present study.

Limitation

A new study requires its own declared endpoint, uncertainty model, decision space, and conclusion rule.

Compare the seven repairs

A reusable rewrite sequence

When a sentence feels too strong:

Step 1 · Find the noun

What evidence object supports it?

Examples:

• structural report;
• readiness report;
• branch ledger;
• stability summary;
• sensitivity table;
• validation case.

Step 2 · Find the denominator

Write the numerator and denominator explicitly before writing the adjective.

Step 3 · Find the strongest literal statement

What can be said without interpretation?

Example:

Six of seven valid branches produced estimates.

Step 4 · Add bounded interpretation

Example:

All six available estimates were positive.

Step 5 · Add the missing-evidence boundary

Example:

The audit remained incomplete because one valid branch failed technically.

Step 6 · Generalise only as far as the study supports

Do not extend beyond:

• the declared policy;
• the declared specification space;
• the tested measurement model;
• the observed cohort;
• the fixed validation protocol.

Methods rewrite pattern

Weak

We applied strict QC and robustness testing.

Better

Structural preflight was inspected before a researcher-declared readiness policy was evaluated. The policy’s trial- and participant-level cohort consequences were previewed before the application scope was fixed. Analytical robustness was then evaluated across the declared valid specification space using one common scientific endpoint, with failed or unexecuted valid branches retained in the execution record.

This still needs the study’s actual criteria, thresholds, specification factors, endpoint, timing, and software identity.

Results rewrite pattern

Weak

The results were robust and insensitive to analysis choices.

Better

Across [successful] of [valid] valid represented specifications, estimates ranged from [min] to [max], with [positive/negative/exact-null pattern]. [Magnitude/direction] was [relatively stable/sensitive] within the declared specification space. [Failure/not-run accounting] remained visible in the execution denominator.

Limitation rewrite pattern

Weak

Some limitations remain.

Better

The robustness audit covers only the declared analytical and measurement dimensions. It does not quantify uncertainty from [untested dimensions], and its across-specification summaries are descriptive rather than sampling-theory or posterior uncertainty intervals.

Use the reporting center

Open the Results Interpretation & Reporting Center and filter by:

• Structural QC;
• Readiness;
• Robustness;
• Sensitivity;
• Frozen evidence.

Each card preserves the same rewrite structure used in this exercise.

API and reference routes

• Evidence and denominator vocabulary
• audit_study_qc() pathway
• evaluate_analysis_readiness() pathway
• run_specs() pathway
• effect_stability() pathway
• Reporting robustness
• Claim-boundary reporting
• Validation matrix

Reuse boundary

Reuse the rewrite procedure.

Do not reuse:

• synthetic counts;
• thresholds;
• effect values;
• evidence labels;
• claim strength;
• limitation scope.

Those belong to the actual study record.