Methodology

1. Community Payout Reliability Score

Each signed-in member may cast one toggleable vote per claim type on an opportunity. The reliability score uses only two of them as opposing signals:

• Positive: reliable_payout
• Counter: rate_risk
• (high_margin and strict_approval are displayed as separate tallies; they do not feed reliability)

We publish the Wilson 95% lower bound of the positive proportion:

The lower bound is deliberately conservative: a program voted reliable 8 times out of 10 scores ≈ 49%, not 80%. Small samples cannot look good.

Minimum-sample gate: below 3 opposing votes the score is NULL and the UI displays “insufficient data” instead of a number.

2. Confidence Score & Auto-Publication Gate

Opportunities discovered by background agents carry a confidence score defined as 1 − payout_anomaly_score, where the anomaly score comes from a structured audit of claims and vendor signals (0 = clean, 1 = certain fraud). Auto-publication requires both: audit verdict pass and anomaly ≤ 0.30. Everything else enters manual review.

3. Evidence States

Every extracted claim carries exactly one state: official_source, community_report, agent_extracted_unreviewed, moderator_verified, stale, disputed, or removed. Agent-extracted claims are labeled as such until a moderator verifies them — never silently promoted.

4. Dispute SOP

Letters move through four visible statuses — waiting_vendor, under_review, resolved, closed_unresponsive — with a 14-day vendor response window before a case is marked unresponsive. Vendors answer through verified Right of Reply; answers appear beside the original letter, never replacing it.

5. Traffic Arbitrage Context (IAB/MRC IVT)

Where ad-network programs are listed, traffic-quality notes distinguish Invalid Traffic per IAB/MRC taxonomy: GIVT (general — known bots, data-center origins) versus SIVT (sophisticated — hijacked devices, fabricated sessions). Programs with unresolved SIVT histories are flagged rather than delisted, because the record itself is the product.