Scoring

How the 0–100 human-likeness score is computed, and which findings charge it.

Every result carries a human-likeness score from 0 to 100. Higher is more human.

What charges the score

Only findings from detection-intent rules — the subset corpus-validated to distinguish AI-generated from authentic human legal prose.

Style-intent findings are reported and fixed like any other, but charge nothing. A document full of semicolons and Oxford commas can still score 100.

The formula

Each charging diagnostic contributes a base penalty by severity, multiplied by its own weight (default 1):

Severity	Base penalty
`error`	5
`warning`	3
`suggestion` (and legacy `info`)	1

Those penalties are summed, normalized per 1,000 words, and decayed:

penalty  = Σ (base_penalty × weight)
density  = penalty ÷ word_count × 1000
score    = round(100 × e^(−density ÷ 100))   clamped to 0…100

Two consequences follow from the shape of that curve.

Length does not help. Normalizing per 1,000 words means the score measures issue density. Padding a document with clean paragraphs to dilute its findings does not work.

The decay is smooth. There is no cliff, and no threshold to game — the score falls off continuously, so removing any single finding always moves it in the right direction.

Density rules — no-hedging, no-em-dash-overuse — charge in proportion to how far past their threshold the text runs, rather than a flat penalty per match. A document slightly over the hedging threshold is charged slightly; one far over is charged far more.

Worked example

A 100-word passage with one warning-severity detection finding of weight 2:

penalty  = 3 × 2            = 6
density  = 6 ÷ 100 × 1000   = 60
score    = round(100 × e^(−0.6)) = 55

Reading it honestly

The score is a density measurement of specific, enumerated patterns. It is not a classifier, and it does not claim to detect AI authorship.

The useful reading is comparative: the same author’s before and after, or one draft against another.