Method / evidence and interpretation

Show the work.
Keep the limits visible.

Arcophos independently analyzes original HealthBench through its paper, official release, data card and pinned reference code. We separate source facts, historical author-reported measurements and our own mathematical explanations.

Count the right unit

Conversation themes and criterion axes have distinct denominators, both retained in the analysis.

Make arithmetic inspectable

Calculator examples use invented abstract criteria and the published normalization rule; they contain no released benchmark conversations.

Keep provenance visible

The displayed results are from May 2025. We did not perform the reported evaluations or create HealthBench.

What this resource cannot establish

  • The calculator does not grade model answers.
  • Historical repeated-run variability is not uncertainty about clinical outcomes.
  • Predominantly synthetic conversations are a designed evaluation sample, not a random real-patient cohort.

Original analysis and reported evidence

We distinguish source facts from our interpretations. Benchmark creators are credited in each dossier. Reproduced result rows are selected historical measurements from an identified paper version and configuration. They are not new Arcophos runs, a complete leaderboard, or a claim about the latest models.

When papers, repositories, or releases disagree, the dossier records which source governs the displayed claim. The evidence JSON exposes these facts, source IDs, and version notes for inspection. Dataset access and licensing are reported separately.

Analytical tools

Coverage explorers organize published task categories with our annotations; the number of displayed entries is not a performance score. Calculator inputs are explicitly illustrative. Published scoring formulas are applied to those inputs, while sampling assumptions and aggregation boundaries remain visible. These tools cannot establish clinical efficacy.

Using the worksheet

The worksheet turns the editorial guidance into a planning brief. Its completion count records items you have addressed. It does not estimate model quality, validate safety, or certify a system. The exported file includes selected and unresolved items so the remaining work stays visible.

Corrections and updates

When a source or interpretation changes, the affected guide and its update date should change together. Send the page URL, the disputed statement, and a primary reference through Arcophos contact.