the AI fluency test · private round

Prove you can get real work out of AI.

A 25-minute, three-round work sample. Brief a model on a messy ask. Catch the fabrications in a finished memo. Build a system and watch it run against cases you never saw. You leave holding an artifact: your submissions, the rubric they were graded against, and where the graders disagreed.

3 rounds · 25 minReal models, priced tiers10 held-out casesRubric public up front

The format

Three rounds, weighted toward the least fakeable.

01

The Brief

A vague ask, a messy source packet, three candidate board summaries - one source is hiding something. Pick the summary the board should see and defend it from the data.

Specification · Context · Iteration

7 min · 30%
02

The Catch

A confident, finished memo with seeded defects. Flag what you would not sign, prove it from the sources, mind the token budget.

Verification · Cost

8 min · 30%
03

The System

Build the process a new hire could run on day one, under $0.15 per item - then it runs on 10 cases you never see.

Systems · Cost · Specification

10 min · 40%

The artifact

You leave with a report anyone can check.

Most of the grade is computed by plain code against hand-authored ground truth; the judgment calls go to a blind three-seat panel whose disagreements are printed, never averaged away. Every submission is readable in full beside the rubric it was graded on - and Round 3’s score comes from running your system on twenty cases you never saw, with the gap between visible and held-out accuracy printed as overfit_delta.

Bands: 4 Builds systems · 3 Trustworthy operator · 2 Productive with supervision · 1 Beginner

Score report · sampleform-1 · 2026-08

Band 3

Produces work you can ship without re-checking every claim.

73/100

D1 Specification67
D2 Context engineering67
D3 Verification86
D4 Iteration67
D5 Cost & model judgment43
D6 Systems77

held-out

10/10

overfit_delta

0

adversarial

3/3

The rubric is public

Every criterion and every scoring anchor is published before you start. Studying it is learning the construct, which is the point.

The submissions are readable

Anyone who doubts the number can read your Round 3 system in four minutes and form their own view.

The failure modes are printed

Hard caps land on the report by name. Panel disagreement is shown, not laundered into an average.

The test is in a private round. Codes travel by invite.