Crumb AxiomThe Institute for Toast Alignment
Intelligence,
to taste.
We taught a machine to make perfect toast.
Then we asked a human what “perfect” meant.
Open experiment / 001
Does your toaster know you?
More heat is easy.
Knowing when to stop is the research.
Tabletop series
Blind comparison / 1 of 3
Which would you eat?
Choose by appearance. The recipe levels stay hidden until you finish.
Three choices. No universal right answer.
Your preference profile
A taste for balance.
Illustrative match = 100 − 10 × distance from your inferred level. No actual AI or cooking advice.
Method note A small, deterministic preference experiment. Your choices stay in this browser. This is a demonstration of alignment, not a trained model.
Your lab notebook
| Run | Preference | Recipe | Match | Baseline |
|---|---|---|---|---|
| No saved runs yet. Your next breakfast could be evidence. | ||||
Saved on this device. Up to 20 runs.
Research, lightly browned
A better score.
For whose breakfast?
Optimization needs a target. Human preference gives it one. Our tiny laboratory makes that difference visible, one slice at a time.
001 The more-is-better fallacy ↗
A fixed baseline always serves level 8. If you prefer level 3, extra browning moves it further from your target. In our toy metric, its match is 50%. Bigger is not a synonym for better.
002 A preference is not a universal law ↗
Three choices narrow a range from 1–10. We round the midpoint of the remaining interval. This assumes a stable, single preferred level; three comparisons cannot establish a complete taste profile. Try again if neither sample feels right.
003 Show your working ↗
Every export includes your three comparisons, selected levels, inferred target, recipe and score formula. Someone else can reproduce the numbers without believing a leaderboard.
A commons for better questions
Proof of taste.
Not proof of genius.
Crumb Axiom explores community-owned preference benchmarks. The planned CRUMB token would support proposal bonds and votes on which comparisons the community studies next.
The lab works today. Token issuance, a chain, governance contracts and proof submission have not been implemented. No rewards or financial returns are offered.
Download the research kit ↓A seven-second hypothesis
Watch the experiment.
One machine. Two outcomes. A human decides.
Download film ↓