Model evaluation

How models perform as panelists.

Every model will be run against the same benchmark suite — cases carrying planted defects — then scored on how many of the planted defect categories it surfaced, averaged across cases. No run has been published yet; the rows below are placeholders.

Suite standard-v1 — suite content not yet published; no run exists Cases not yet published Suite metric category recall vs. planted defects Accuracy metric not yet defined Source first-party runs — none yet
# Model Accuracy Cost / corroborated finding Latency
01 claude-sonnet-4.6 94.2% n/a 1.2s
02 gpt-4o 92.8% n/a 0.9s
03 gemini-2.5-pro 90.6% n/a 1.5s
04 llama-3.3-70b (local) 88.1% n/a 4.1s
05 claude-haiku-4.5 85.4% n/a 0.4s
06 qwen-2.5-coder-32b (local) 82.0% n/a 3.3s

Illustrative data for the concept mockup — not a published benchmark.

When this board carries real numbers, they will be first-party measurements: the suite run by the maintainer across each model, not scores submitted by their vendors or users. That is the choice made here rather than an upstream requirement — the suite contract only guarantees a submission is suite-sourced, not that whoever sent it actually ran anything. It is why there is no submission endpoint.

No run has been published yet, so every row above is a placeholder — the suite, its case count, and the run date will appear here once one has.

Cost reads n/a on every row because nothing has been measured — the mockup carried per-1K-token prices, which are a different quantity, and relabeling them would have swapped one metric for another without saying so. The Accuracy column is the remaining open question: the suite scores category recall, and what fills "Accuracy" gets settled before any real run is published, not after.

Methodology

How the suite scores a model

01

Seat the model

Each model in the roster reviews every case in the suite as a single panelist, emitting findings in the same standard format a real review produces.

02

Same ingestion path

Each case's diff goes through the same ingestion the production pipeline uses, so a model is scored on exactly the payload it would see on a real review.

03

Score vs. planted defects

Each panelist's findings are matched individually against that case's planted defect categories — no clustering step — giving category recall. Precision is not scored: planted defects are only a subset of what a reviewer may legitimately find.

Recall is one input, not a verdict. A cheap, fast model that catches fewer planted defects is still a strong panelist when paired with a slower, pickier one — because the reconciler scores them together. Use the table to staff a panel, not to crown a winner.

$atcr doctor

Verify your own roster against every configured endpoint before you spend a review on it.