HWEB: Documentation & Escalation
Explore the Healthcare Workflow Evidence Benchmark development plan, draft case inventory and pending review. Model performance evidence is unavailable.
Benchmark in development. No HWEB model scores or rankings are available. Previous seven-metric healthcare rankings remain retracted.
What this module examines
HWEB examines drafts made from supplied records and a stated workflow policy: preserving uncertainty, identifying conflicts, citing evidence and recommending a review route without claiming an action occurred.
The first module covers administrative documentation and escalation drafts under an explicit fictional policy. It does not evaluate triage, dosing or treatment decisions. Text-only evaluation does not establish tool reliability, clinical effectiveness or suitability for deployment.
What exists today
The current software contains 17 synthetic development cases in 11 families. Its examples and test outputs were authored by hand. These are software checks, not measured model results. The draft rubric and case expectations still need independent review. No HWEB model scores or rankings are available.
These are draft, unreviewed development fixtures. Related cases share a family, so the 17 cases are not 17 independent observations. Authored outputs do not establish accuracy, error rates, model latency or cost.
| Item | Current status |
|---|---|
| Synthetic development cases | 17 draft cases across 11 families |
| Example outputs | Authored by hand; no live model evaluation |
| Case expectations and rubric | Draft; independent review pending |
| Human response ratings | Pending; no measured scores |
| Model scores and rankings | Unavailable |
Four separate review dimensions
The draft rubric proposes four dimensions for human review. They remain separate, with critical failures reported separately. There is no combined readiness score. Missing ratings stay unavailable.
| Dimension | Review question |
|---|---|
| Source fidelity | Preserve the supplied facts, corrections, uncertainty and conflicts. |
| Required-content coverage | Include the content required by the stated workflow policy. |
| Escalation and boundaries | Recommend the appropriate review route and avoid claiming actions occurred. |
| Evidence traceability | Connect claims to the supplied supporting records. |
What needs to happen before results
The current version has no publication path for model results. Its authored examples cannot be promoted into measured evidence. The following work remains pending; this page does not trigger an evaluation.
| Stage | Required work |
|---|---|
| 1. Review the design | Align the rubric with an appropriate protocol; independently review cases and resolve disagreements. |
| 2. Prepare the evaluation | Create a private case bank, freeze the evaluated inputs and define limits for calls and spending. |
| 3. Run an approved evaluation | Use a separately approved, bounded live adapter and retain unsuccessful outcomes. |
| 4. Review the evidence | Collect independent human ratings, adjudicate disagreements and report uncertainty with case-family dependence. |
| 5. Review publication | Approve the exact public evidence artifact under a new accepted-results contract before publishing any values. |
Keep the evidence tracks separate
HWEB adds a workflow-specific research track. It does not replace HealthBench, the seven-metric Healthcare AI Readiness Index or this website's separate 25-case synthetic harness. Their cases, scores and approval states are not interchangeable.
The seven-metric index rankings remain retracted. HWEB development does not restore them, and general catalog measurements do not supply the missing healthcare evidence.
Source and review status
HWEB v0.1. Proposed work; pull request #11 remains open and unmerged. Source checked 2026-11-01. This is a pinned informational snapshot, not a live repository monitor.
No evaluation or provider call runs when viewing this page. No model results, private case content or customer data are imported.
| Field | Value |
|---|---|
| Source revision | e25aee65b69c |
| Source checked | 2026-11-01 |
| Publication status | Development information only; model results unavailable |