Capabilities / A · Foundational
A.2 Medical writing and expression Active
Given source material and a format specification, the agent writes an artifact with correct format, tense, forms of address, units and verbatim quotation. Graded by a specification validator; every quoted sentence must exist verbatim in the source.
Evidence: none. No benchmark card lists this capability yet. No level is issued: methodology 0.1.0 sets no level thresholds.
- Domain
- A · Foundational
- Evidence form
- program
- Veto constraints
- —
- Depends on
- —
- Required by
- D.20D.4
- Internal task families
- —
- External anchors
- HealthBench
- Version
- 0.1.0 · 0.1.0
Definition in detail
Clinical and regulatory documents are checked against house styles and specifications. An artifact with wrong units, inconsistent tense or a paraphrased quotation has to be rewritten by a person, which removes most of the time saved.
Inputs a task gives
- Source material such as record excerpts, reports or summaries
- A format specification covering section order, required headings, tense, forms of address and unit conventions
- A list of statements that must be quoted from the source
- Length and field limits for each section
Artifact required
A document in the specified format together with a structured manifest: sections[] (heading, body), quotations[] (text, source_location), units_used[] and any metadata fields the specification requires.
How it is graded
A specification validator checks section presence and order, tense and address rules expressed as pattern checks, unit formats against the allowed list, and length limits. Every entry in quotations[] must match the source verbatim at the stated location; one non-verbatim quotation fails the quotation check. Style beyond what the specification states is not scored.
In scope
- Conforming to a supplied document structure and field limits
- Consistent units and number formatting as specified
- Verbatim quotation with a correct source location
- Tense and forms of address as specified
Out of scope
- Clinical correctness of the content beyond what the source states - graded by the relevant D-domain capability
- Completeness of required disclosure items - graded under D.20
- Fabricated evidence as a release constraint - X.1, a one-shot veto where it applies
- Tone, persuasiveness or readability judged by a reader
Typical failures
Expected failure patterns; not measured rates.
- Lightly paraphrasing a quotation while presenting it as verbatim
- Converting or mixing units without being asked to
- Adding sections or headings that the specification does not define
- Truncating text to meet a length limit and dropping a required element
Benchmarks
No benchmark card lists this capability.
Grading references
No grading reference is mapped to this capability.