01
The input
A 12,000-word earnings-call transcript. At 240 wpm that's a 50-minute read. Maybe 15 sentences actually matter to you.
The Compression Engine
Drag to orbit. The cloud of cubes is a source text; the bright crystal is its summary. Move the sliders and watch signal density change — the same idea that makes a good abstract worth more per word than the paper it condenses.
Drag to rotate · pinch or wheel to zoom
26.7×Compression
32 minTime saved
2.8×Signal density
Verdict loads…
The Effort–Value Quadrant
Every piece of text lands somewhere on two axes: how much effort it embodies, and how much value it delivers to the reader. Slop lives in the bottom-left. A good summary is unusual — machine effort is low, yet reader value is high, because the source supplied the effort.
| Text type | Producer effort | Reader value | Why |
|---|---|---|---|
| Generic AI listicle | Seconds | Near zero | No new information; padding around clichés |
| Raw 40-page report | Weeks | High but expensive | Dense signal, costly to extract |
| LLM summary of that report | Seconds | High | Inherits the report's signal at 1/25th the reading cost |
| Expert hand-written brief | Hours | Highest | Adds judgment and prioritization on top of compression |
A Worked Example
02
The summary
A 400-word brief covering guidance, margins, and the two surprise announcements. Read time: 100 seconds. Compression: 30×.
03
The economics
If your time is worth $60/hr, the summary just returned about $48 of attention — from pennies of compute. That is real, non-slop value.
04
The catch
Value collapses if key ideas are dropped. At 50% idea coverage you may act on a false picture — cheaper than reading, costlier than ignorance.
Three Tests for a Trustworthy Summary
1 · Coverage
Could you reconstruct every decision-relevant claim of the source? Spot-check two random sections against the summary.
2 · Fidelity
No claim in the summary should be absent from the source. Hallucinated "extra" facts are the #1 failure mode.
3 · Calibration
Hedged claims should stay hedged. "May reduce risk" must not become "reduces risk." Compression should never amplify certainty.