Skip to main content

IndexFair › Methodology › Review filtering

/methodology · v2.1 · append-only · last updated 2026-01
/methodology/review-filtering

Review filtering

A deterministic filter removes signals that can't be scored or that look manipulated, before anything counts toward a number.

What it means

The filter applies four deterministic checks to every review before it is eligible to contribute to a score. Each check has a published category name and a published count of what it removed — the exact thresholds and detection logic are withheld so a bad actor cannot tune around the filter.

Inputs & data fields

FieldTypeMeaning
content_fingerprinthashContent-derived hash used to collapse exact and near-duplicate reviews.
posted_attimestampPosting time; used to detect coordinated burst patterns.
content_quality_signaljsonbPre-computed signals: length_chars, paragraph_count, sentiment extremity.
aspect_ratingsarrayExtracted aspect ratings; review dropped if none are gradeable.

Process & formula

Categories we disclose
01De-duplicate by content fingerprint
02Drop posting bursts & coordinated timing
03Length & sentiment anomalies
04No gradeable aspect detected

Withheld by design. We publish the categories and the counts of what is set aside — never the exact thresholds or weights, which would let a bad actor tune around the filter.

How it affects public pages

Example: 5,132 collected → 3,324 used (65%). Affects the funnel on /review-signals.

Worked example

Lumen Exchange · fake data
01Collected: 8,214 reviews
02De-duplicated: −412
03Burst / anomaly removed: −1,630
04No gradeable aspect: −540
→Used: 5,632 (68%)
Related