How the style match score works.
Computed. Never self-reported.
Most AI writing tools that show a “human score” got it one of two ways: they asked the model to grade its own output, or they ran a third-party AI detector and displayed its guess. The first is circular — models are confidently wrong about themselves. The second inherits every false positive detectors are notorious for.
Shlokah does neither. Your writing samples are reduced to concrete, reproducible numbers in code, and every draft is measured against those numbers after the model has finished. The model never grades itself — it gets graded.
What gets measured
Sentence rhythm
Average sentence length and its variation (burstiness). Humans mix short punches with long, winding sentences; language models flatline near one length. The strongest single human/AI discriminator we measure.
Contraction rate
The share of “don't / it's / can't” versus “do not / it is / cannot” across your samples. A formal writer rewritten with contractions is as wrong as a casual writer rewritten without them.
Punctuation habits
Em-dashes, semicolons, exclamation marks, ellipses, parentheses and commas per 100 words — plus whether you use them at all. If you've never typed an em-dash, one in the output is a violation.
Casing and casual markers
Lowercase sentence starts (a real texting fingerprint), emoji frequency, and the filler words you actually use (“honestly”, “tbh”, “anyway”). Sanitizing a casual writer is a style failure, not an improvement.
AI-tell vocabulary
A curated list of words and stock phrases over-represented in LLM output (“delve”, “robust”, “it's important to note”…). Only penalized when they don't appear in your own writing — if you genuinely say “moreover”, you keep it.
Fabrication checks
Dates, times, weekdays, and commitments that appear in a draft but never in your request are flagged as inventions and force a revision. A model can't be trusted to police its own fabrications.
From measurements to a score
Every draft starts at 100 and loses points for each measured deviation from your baseline — a rhythm mismatch, a contraction-rate gap, punctuation you never use, AI-tell vocabulary foreign to your writing. The penalties are weighted by how strongly each signal separates one writer from another, with sentence rhythm weighted heaviest.
When a draft scores below threshold — or contains a fabricated specific — it isn’t shown to you. The exact violations (“average sentence length is 19 words; this person averages 11”) go back to the model as revision instructions, and the revised draft is measured again. You see the result only after it survives measurement, along with the score it earned.
The same rules power the free AI-tell checker — so what the checker flags and what the rewriter fixes are always the same definition of “sounds like AI.”
What the score is not
It is not a detector-bypass guarantee, and we don’t sell it as one. Detector verdicts swing with every model update on both sides. The style score answers a different, more durable question: does this text measurably match how you write? The reader who knows you is the real judge — the score exists so you can predict their verdict before you hit send.
Your samples stay private to your account, are never used to train models, and can be deleted anytime — the full policy is in our terms.
See your own number
Paste one AI draft and one real email into the demo — the score it returns is computed exactly as described above.
Run the free demo