PENITENT.EXE — the confessor

The model behind PENITENT.EXE, an artwork: you type a confession, a machine weighs it against the seven deadly sins, and a weighted coin decides whether you are absolved.

It runs entirely in the visitor's browser (onnxruntime-web). The confession is never sent anywhere; only these files are downloaded.

This is a performance piece, not a moral authority. Its judgements are opinions a machine learned to imitate, and it will be wrong.

What it does

For one confession it answers nine questions, one per sin plus despair and remorse, each as a probability over rubric levels 0–3 ("gravity"):

pride  greed  lust  envy  gluttony  wrath  sloth  despair  remorse

Architecture: an Ettin-68m encoder with a Laya/open-jev decision head (the question and its rubric levels are in the input, one [MASK] per level; the head scores each marker).

Training

Short confessions from Reddit's confession communities, each scored by a large language model against the same written rubric, with the full distribution over levels kept as the training target. On 1,000 held-out confessions its top sin agrees with the teacher 77% of the time.

Files

file	what
model.onnx	the graph; int8 weights, computed in fp32; any input length
tokenizer.json, tokenizer_config.json	the encoder's tokenizer
penitent.json	the nine questions and the constants the page needs