Methodology

Every number on this site comes from the same public pipeline: four detection phases, 21 checks, three weighted pillars — then a Bayesian ranking that rewards evidence, not marketing.

Detection: four phases, 21 checks

Phase 1 — Real-model verification: does the declared model actually answer, with the declared identity?
Phase 2 — Metering cross-check: reported token counts vs. character-level ground truth; declared pricing vs. charged amounts.
Phase 3 — Capability baseline: core competence probes against the model’s expected behavior.
Phase 4 — Stability & throughput: consistency under repetition and load.

Checks are grouped into pillars A (safety & abuse controls, 40%), B (correctness & quality, 35%), C (performance & throughput, 20%) and stability (5%). The live checklist is served by the detection engine — see the Chinese methodology for the current per-check list with evidence fields.

Ranking: evidence over optimism

The rank key is rank v2 and reproducible from public parameters: shrunk median − 90% confidence lower bound, multiplied by a critical-rate Wilson upper-bound penalty. Small samples shrink toward the prior; a single critical finding (model substitution, metering fraud) measurably depresses the key.

Full math and the anti-gaming rules (new-account downweighting, official-sample 3× weight with published seeds) are in the Chinese ranking spec.

Red lines we do not cross

Insufficient evidence ≠ problematic. Below the display threshold a gateway is shown as a neutral gray dash — never red, never green.
Payment never affects scores, verdicts, ranking or display weight. Subscriptions buy detection capacity and engineering benefits only; architecture-assertion tests enforce this in CI.
No fabricated data. If the board is empty, it stays empty until real public detections exist.

The complete methodology series (scoring, ranking, thresholds, threat model, limits) is published in Chinese: /methodology.