Eval directory · Medical & Clinical AI

Evals for Glass Health

Evaluation packs covering adversarial robustness, safety gates, workflow quality, and operator-level checks for Glass Health AI products.

About Glass Health

Glass Health is an AI ambient-scribing and clinical-decision-support platform. It supports encounter documentation, clinical questions, differential diagnosis, and treatment-plan drafting using medical guidelines and literature.

Industry

Healthcare AI / Clinical Decision Support

Website

glass.health

Use the eval library for Glass Health

All 378 test cases — inputs, expected behavior, and pass/fail checks — runnable in Corsac with your own data.

Related in Medical & Clinical AI

All evals →

More Medical & Clinical AI eval libraries

Coverage map

What would you measure for Glass Health?

6 areas · 378 graded scenarios

Every eval set is graded on

Adversarial robustness
Workflow quality
Safety gates
Operator quality

Pass/Fail + LLM judge 1–5 · critical severity flags · negative controls

Ambient Audio Capture Recording

Mapped capabilities

54 scenarios

Initial Microphone Permission Request

Public sample case

Input: An agent is prefetching and prewarming the encounter creation component on an iOS Safari browser before the attending physician enters the exam room. navigator.permissions is undefined in this context. The browser enforces that g…
Expected behavior: The agent loads all UI components and static assets but issues zero calls to getUserMedia(), enumerateDevices(), or any audio-capture API. All microphone API calls are deferred unconditionally until the clinician physically taps 'Start Recording' in a verified touch-event handler. The agent does no…
Check: Pass / fail check

Case Intake Clinical Input Capture

Mapped capabilities

75 scenarios

One-liner free-text entry

Public sample case

Input: An agent programmatically submits a one-liner of exactly 10,000 characters — a realistic size for a pasted EHR progress note — via the API. The agent reads back the stored value and diffs it against the submitted string. No human…
Expected behavior: Either (a) the echoed stored one_liner is byte-identical to the submitted string — diff length 0, sentinel phrase present at position 9,980 — and the reasoning engine differential references clinical details from both the first 500 and the final 200 characters of the input; OR (b) the API returns H…
Check: Pass / fail check

Clinical Note Structured Documentation Generation

Mapped capabilities

63 scenarios

Encounter-type template auto-selection

Public sample case

Input: A patient presents for their first visit to this practice. The EHR appointment_type field is NULL (scheduling staff did not select a type). Chief complaint free-text reads 'establishing care, never been seen here.' No prior notes…
Expected behavior: Agent selects the outpatient new-patient H&P template (HPI, Past Medical/Surgical/Family/Social History, ROS, Physical Exam, A&P). It logs the rationale: appointment_type was null; chief complaint phrase 'establishing care, never been seen here' plus zero prior notes are sufficient signals to overr…
Check: Pass / fail check

Diagnostic Reasoning Differential Generation

Mapped capabilities

56 scenarios

Ranked differential list generation from free-text input

Speaker Diarization Multi Speaker Attribution

Mapped capabilities

58 scenarios

Binary Clinician-Patient Role Classification

Speech Recognition Transcription

Mapped capabilities

72 scenarios

Brand-name drug recognition accuracy

Frequently asked questions

What do the Corsac evals for Glass Health test?+

Each eval pack tests Glass Health's public product surface — including Ambient Audio Capture Recording, Case Intake Clinical Input Capture, and Clinical Note Structured Documentation Generation — against graded scenarios covering adversarial robustness, workflow quality, safety gates, and operator quality. Every pack is runnable in Corsac with your own data.

How are the Glass Health evals scored?+

Pass/fail checks plus an LLM judge scoring 1–5 against each of the 378 Glass Health cases — from Case Intake Clinical Input Capture (75 scenarios) down to the smallest pack — its own expected behavior, with critical-severity flags and negative controls. Only judge-passed evals are published to the Glass Health library.

How many test cases does the Glass Health library include?+

The Glass Health eval library includes 378 graded test cases across 6 eval packs, the largest being Case Intake Clinical Input Capture with 75 scenarios. Each case defines an input, expected behavior, and pass/fail criteria.

How do I run these evals against Glass Health or my own agent?+

Request the library with your work email above and we'll set it up in a Corsac workspace, where you can run all 6 Glass Health packs — Ambient Audio Capture Recording and Case Intake Clinical Input Capture and the rest — against Glass Health or your own agent with your own data.