Scoring method
No secret scoring.
Every published constant, penalty, bonus, last-updated rule, and source-quality rule is exported as policy data.
Constants
- independent_headline_recent_within_days: 120 · Independent-only headline evidence must be recent enough to avoid relying on stale context rows.
- coverage_report_visible_target: 90 · Coverage audit flags preset/model pairs at or below this visible-coverage target.
- latency_gate_percentile: 60 · Low-latency mode requires mean speed evidence at or above this percentile.
- finalist_limit: 4 · Shareable reports surface enough finalists to keep the strongest clean candidates and the leading hybrid/preview source-data candidate visible.
Scoring method
- Recommendation formula version: recommendation-fit-v2.0.0.
- Custom-ranking preview formula version: composite-preview-v1.1.0.
- Domain leaderboard formula version: domain-leaderboard-v1.1.0.
- Base score = source-balanced weighted mean of benchmark percentiles for the selected preset domains.
- Confidence penalty = spread penalty + last-updated penalty + source verification penalties + relay/backfill/manual/provider/preview/rollup penalties.
- Coverage bonus = max(0, exactCoverage - requiredCoverage) * 0.18.
- Efficiency adjustment = +6 budget, +3 mid, -1 premium, -4 frontier when efficiency preference is active.
- Reasoning adjustment = +4 for reasoning-track models, -1.5 otherwise, when reasoning preference is active.
Penalties and bonuses
| ID | Type | Value |
|---|---|---|
| benchmark_spread | penalty | max(0, spread - 24) * 0.18 |
| aging_evidence | penalty | max(0, 70 - freshness) * 0.12 |
| blocked_source | penalty | 1.8 per source |
| relay_source | penalty | 1.2 per source |
| warning_source | penalty | 0.6 per source |
| recent_parser_or_mapping_change | penalty | 0.9 per source |
| backfilled_row | penalty | 0.7 per row |
| relay_row | penalty | 0.8 per row |
| manual_only_evidence | penalty | 1.2 when no verified runtime row remains |
| provider_official_row | penalty | 0.9 per row |
| preview_evidence | penalty | 1.6 when preview rows contribute |
| track_rollup | penalty | 0.6 when model-family rollup contributes |
| exact_coverage_bonus | bonus | 0.18 per point above coverage floor |