How it works

Anyone can claim an AI agent saves time. This proves it, or it doesn't.

Back to proof

1 · What we compare

Every run gives three readings.

Owner says a person needs01

30 minA forecast, typed once at registration. Nobody checks it.

Agent reports02

236Runs, cost, tools. What it did — not what it was worth.

The work is worth03

9.2 minWe read the finished work and price it ourselves.

Readings 1 and 3 are the same question about the same unit: human minutes. Neither is how fast the agent runs.

2 · How we get the third number

The same four steps, every time.

We read the output

Not the prompt. The work the run produced.

We price the work

How long would a person need for this?

We take the middle

The median, so one odd result cannot move it.

We compare

Proven if it reaches 60% of the claim.

▶Show exactly how the work is priced, and why it can be trusted

A fixed rubric, run locally on the
finished output. No model, no network.

  0.5 min        to pick the task up
  + words ÷ 180  drafting
  + 0.25 min     per list item
  + 0.3 min      per heading
  + 0.4 min      per table row
  + 1.5 min      per code block
  + 0.2 min      per figure, max 12

Held between 0.1 and 480 minutes.

The owner's claim is not an input — the
rubric never sees it, so it cannot copy
a number it has not been shown. It reads
the finished work, never the request.

The owner’s claim is not an input. It cannot copy a number it has never been shown.
It reads the finished work, not the request. A big ask is not a big result.
We use the median of many outputs, not one.
The same output always prices the same. Nothing is sampled or guessed.
Every measurement is kept. Open any agent and see them all.

3 · The three answers

Only a proven claim counts as value.

Proven

The work is worth what the owner said a person's time was worth.

Unproven

Not enough priced outputs to say either way.

Disproven

The work is worth much less than the baseline claimed.

▶Show the rule behind each answer, and a worked example

Proven: priced >= 10 outputs AND measured minutes / claimed minutes >= 0.6 AND nothing contradicts it
Unproven: fewer than 10 priced outputs
Disproven: the priced work, or a tool used, contradicts the baseline

The same arithmetic on two of your agents, both at $62/hr loaded. The bar is measurement against claim; the tick is the 60% threshold.

Pipeline SummariserProven

Owner claimed: 15 min per run
The work is worth: 10.5 min, across 32 outputs
Value at the claim: $2,573
Value at the measurement: $1,801

70%

32 outputs are worth 10.5 min of a person's time, against a 15 min baseline — 70% of it. The claim holds up.

Candidate ScreenerDisproven

Owner claimed: 30 min per run
The work is worth: 8.8 min, across 28 outputs
Value at the claim: $7,688
Value at the measurement: $2,255

29%

Reaching crm_write, which it was never approved for.

4 · What is real here

Nothing measured is invented.

Reading the output and pricing the work in it	Real, live model call
Scoring the prompt against the declared job	Real, live model call
PII and unapproved-tool scan	Real
Proof, verdict and money arithmetic	Real
90 days of runs, costs and measurements	Sample
The minutes claims	Sample — self-reported, as in production