# Manual Labeling Protocol: PageSpeed Cloaking Census

## Label universe
Each sampled site receives one site-level label after profile-paired crawls and ablation crawls:

- Confirmed PageSpeed cloaking: stable, directionally score-improving payload difference between real-user and Lighthouse/PSI-like profiles, not explained by benign variation, and supported by at least one mechanism indicator.
- Suspicious PageSpeed cloaking: material directional differences persist, but mechanism evidence or controls are insufficient.
- Clean/no evidence: no material directional difference, or differences are non-directional and consistent with normal variance.
- Benign differential serving: difference is explained by responsive design, geo/language, consent state, bot protection, A/B testing, outage, cache, or declared user-agent content negotiation.
- Inconclusive: repeated runs or controls fail, page unavailable, or artifact quality is insufficient.

## Minimum confirmation rule
A positive label must satisfy all four criteria:

1. Direction: the Lighthouse/PSI-like profile receives a materially lighter or more score-favorable payload than the real profile.
2. Stability: the directional pattern appears in at least 2 of 3 paired runs, with order randomized.
3. Specificity: A/A and B/B controls show smaller differences than A/B differences; ablations localize at least one fingerprint feature or mechanism.
4. Materiality: at least one preregistered materiality threshold is crossed, e.g. JavaScript bytes down >=30% and >=100 KB; total transfer bytes down >=25% and >=250 KB; third-party request count down >=50%; DOM/text similarity below threshold; or LCP candidate manipulation indicators.

## Human adjudication
Two independent reviewers label every high-confidence positive, a stratified sample of low-confidence positives, threshold-near negatives, and random negatives. Disagreements go to a third reviewer. Report Cohen's kappa, per-class precision, and a false-positive adjusted prevalence estimate.

## Reviewer blind
Reviewers see paired artifacts labeled Profile A/B, not the actual fingerprint condition, to reduce expectation bias. Mechanism fields are revealed only after the initial content-difference judgment.