Write a character, see what it reads
A digit, a capital or a small letter. A small network reads it with weights that were typed by hand, one reason per line. Nothing was trained.
Write here
What it thinks
Inside the network
Everything behind the answer for the character on the pad: the feature maps the network builds, the handwriting-style rule that won, and the referees between look-alikes. Write on the Try it tab and come back.
What the network sees (28 × 28)
Your strokes are redrawn with a pen 12% of the character's height, the bounding box is fitted into 20 × 20, and the centre of mass is moved to the centre, as MNIST was made. Each map below is one named feature, lit where it fires.
Why
The winning style rule
Referees that spoke
A referee is a hand-written judge for one pair of look-alike classes. It speaks only while both are within 1.5 points of the top score.
How far to trust it
Measured on training images set aside and never looked at while writing the weights. The percentages are a softmax of the class scores with temperature 0.5, typed by hand and never fitted; the second table shows how often an answer shown at each level was right.
Scores
Are the percentages honest?
How it's made
A convolutional network whose every nonzero parameter comes from one line of a plain-text edit list, each with its reason. Inspired by WPLM, which programmed a language model the same way.
How the weights were written
How the page works
The page runs the plain-text edit list and architecture directly: it starts every parameter at zero and applies the edits in the browser. Your drawing never leaves this page.
Drawing pipeline: strokes are redrawn with a pen 12% of the character's height, the bounding box is fitted into 20 × 20, and the centre of mass is moved to the centre, as MNIST was made.
"Digits only" and "Letters only" read outputs that hear only the referees within that kind, so a letter-vs-digit look-alike cannot pull the answer.
Files and the log
- README.mdWhat it is, the results, how to run it.
- LOG.mdHow the weights were written: every decision, the numbers before and after.
- AGENTS.mdHouse rules and how to improve it.
- weights/charnet.editsThe network itself, one edit per line.
Weights and code written by Claude Opus 5.5 (Anthropic) with the project's author. Sister project: WP-Ductus, a calligraphy teacher built the same way. MIT licence.