CSL-Meter - Does your AI plan serve the model it claims? Measure it.

Maker here. The reading that made me build this in the open: a $10/mo coding plan claiming kimi-k3.1-pro whose served distribution, across four independent harness framings, matches claude-opus-5.5's baseline far better than Kimi's (which ranks 4th–5th). One caveat I want on the record before anyone else raises it: our golden baselines are still being crowd-calibrated, so we publish "distribution ≠ claimed baseline", not fraud verdicts. If you have first-party API access, your calibration captures would genuinely move this project forward. Questions about the method very welcome.

CSL-Meter is a public observatory for AI subscription plans and API offerings. It runs real probe batteries against official endpoints, fingerprints the served model's response distribution, de-biases it across four client harnesses, and scores every plan for consistency, substitution risk and latency — with raw distributions attached to every reading.

Honest by construction: simulated data is labeled in the data model and can never be counted as verified; insufficient samples produce no report at all; submitted keys are never stored. Open source (Apache-2.0), pure Python stdlib, self-hostable in minutes.