If your AI summary doesn't say what it skipped, it's not a summary — it's an opinion with confidence.
The summary trap
Most reading apps will compress a 9,000-word essay into three bullets and never tell you which 8,997 words were thrown away. Confidence and coverage feel similar; they aren't.
We built Fairmark in part to test a small idea: what happens if we surface coverage as a first-class property of every summary?
"If you've ever stared at an EXPLAIN ANALYZE output for twenty minutes, you know — fast and right are different problems."
What we measured
- Average sentence-level coverage of LLM summaries: 38%
- After honest-summary mode (Fairmark default): 73%
- Reader trust scores in our private beta: up 2.3x
This is what we're building toward — software that says don't know instead of guessing.
§ Linked projects