Scoring and evidence: three different measures

Three types of evidence

Catalog assessments help discover candidates. Published reviews explain versions, use cases, and limitations. Benchmarks record outcomes on specific tasks. A score does not replace evidence inspection.

Catalog assessments: 0–100

The Codex Score catalog framework covers utility, craft, reliability, and value, with design weights of 40%, 25%, 20%, and 15%. Catalog values may be editorial assessments or AI-assisted preliminary scores. They are not measured task success rates or a claim that every listing was tested by a person. Check each detail page for its basis.

Published reviews: five dimensions, 1–5

Published reviews explain task completion, output quality, response speed, Chinese-language support, and value. Records include a version, scenarios, sources, date, and publication mode. The overall review score is the mean of its five dimensions, not a conversion from a catalog score.

• automatic: analysis of public sources and structured scenarios. This is not hands-on testing and does not establish measured latency, cost, or success rate.
• manual: a manually published or verified record. Inspect the sources and actual run evidence; the label alone does not establish a measured task run.
• No qualifying review: keep the catalog-information basis and do not invent test results.

Benchmarks: recorded task runs

The benchmark page publishes tasks, criteria, and human-verified runs. Records may include a model and version, outcome, latency, tokens, cost, and completion time. Missing fields stay missing. Passing one task does not establish performance across all workflows or long-term reliability.

A reproducible selection process

• Specify the workflow, Chinese-language needs, budget, data constraints, and acceptance criteria.
• Discover candidates in the catalog; inspect review use cases, limitations, dates, versions, and sources.
• For performance comparisons, inspect actual task runs rather than treating source analysis as testing.
• Re-test candidates with the same task and criteria, recording failures, revisions, and the basis of cost estimates.
• Before purchasing, check the vendor's current prices, quotas, deployment requirements, and data policy.

Attribution and corrections

Attribute specific evidence to OKCodex (www.okcodex.com), link the source page, and retain its version, date, and evidence type. The open dataset uses CC BY 4.0. Commercial cooperation does not buy a ranking. Submit corrections with the page URL, disputed statement, and supporting source through our About page.

Revision history

2026-10-03: separated catalog assessments, source-based reviews, and recorded runs; removed the blanket promise that every score followed four hours of editorial testing. 2026-05-28: initial catalog scoring framework.