Product health

Generated documents · US and EU

Product health

Cost and speed of generated documents, by instance, for Nov 10–17, 2026. US DavidAI against the 7 days before. EU DavidAI: its first full week. Meeting minutes: first readings from Nov 10–16.

Week of Nov 10–17 No readings need attention

Thermometer

Each reading is colored against its own previous 7 days. Dashed cards hold first readings that aren't colored yet. Click any card for details.

Within 15% of previous week 15–40% worse More than 40% worse First readings, not colored yet

Quality over time

Glasslake’s quality score out of 100 for its meeting minutes, one reading per meeting size per week. How it’s measured is explained below the table.

How quality is measured

Each reading uses the same set of invented test meetings, small, medium and large, in Portuguese, English and French. Glasslake writes minutes for every meeting twice. A reviewer model then grades each document against the meeting transcript, without being told how it was produced.

Attendees
Who was in the meeting
Decisions
What was decided
Open points
What is still unresolved
Actions
Who does what next
Header
The title and details at the top
Faithfulness
Matches what was said, nothing invented
Usefulness
Works for someone who missed the meeting

Small meetings are also checked against a written answer key. On Nov 10, 82% of those checks passed. Medium and Large rest on far fewer documents, so read them as indicative.