Scale
0-100
Higher means closer to the provided writing region.
How Idiolect measures writing voice.
Voice Match is a relative score for how closely a draft reads like a person's own writing region. It is a measurement primitive for personalization, not an AI detector, not a clone claim, and not forensic authorship proof.
Best signal
Delta
Compare generic output against the in-voice draft.
Instrument
LUAR-MUD
Used when the validated scorer is configured.
The method
What happens when a draft is scored.
01
Start with a writing region
Voice Match compares a candidate draft against examples the person actually wrote. More representative samples make the region more useful.
02
Embed the region and candidate
When the validated scorer is configured, Idiolect uses the LUAR-MUD authorship embedding. If the validated scorer is unavailable, the product labels the estimate instead of pretending it is the same instrument.
03
Return a relative score
The 0-100 number is a similarity score for this draft against this writing region. The useful product signal is the delta versus a generic draft, not an absolute identity claim.
How to read the number
AUC is measurement power, not clone fidelity.
Idiolect reports AUC around 0.94 for the deployed Voice Match configuration. That describes author-separation power in a validation setting. It does not mean a generated draft is 94 percent identical to a person.
In product, the safer reading is comparative: did the rewritten draft move closer to the user's writing region than the generic draft did? That is why Idiolect leads with Voice Match deltas and keeps the limits visible.
Good inputs
- Several pieces written by the same person.
- Text in the same register as the target draft.
- Enough length to show rhythm, vocabulary, and recurring moves.
- A comparison against generic output so the delta is visible.
Limits we state up front
Voice Match is not an AI detector and does not classify text as human or machine.
It is not forensic authorship proof and should not be used to accuse or identify a person.
Short fragments, topic overlap, genre shift, and too few samples can distort the score.
Genericness Score is separate: it estimates default-AI resemblance, not voice similarity.
Build with it
Developers can call the Score API directly, or let an agent use the MCP connector to load a user's Voice Card and check a draft before sending.