// nuwright
◔ building · in validation

Helping-skills competency scorer

Ticks helping behaviours in session transcripts and quotes the words; code sets the level.

job
helping-skills competency scoring
domain
care · helping-skills competency
runs
workstation
class
judgment model
scale
27-billion-parameter reader; 120 behaviours across 15 items
since
2026

One call

  1. 01tidyCode puts six speaker-label styles into one form
  2. 02readLocal model ticks each behaviour and quotes the words
  3. 03checkCode confirms each quote is real practitioner speech
  4. 04rulebookCode turns surviving ticks into a level per item
  5. 05flagHarmful behaviour shown to a supervisor with the words

In · out

inputs
  • session transcript
  • published competency form
outputs
  • ticked behaviours with verbatim quotes
  • level per item, computed in code
  • harm flag for supervisor review, with the words

Where the person is

A quote that fails the check is dropped and the reason recorded. Harmful behaviour goes to a supervisor, with the practitioner's own words.

code decides● yes · the final say is deterministic

What is measured

Against: Expert-coded sessions from a public motivational-interviewing corpus, in a pre-registered test.

Agreement with expert-coded sessions: AC1 0.708 (95% interval 0.569 to 0.827)pre-registered
Sensitivity 19 of 20 and specificity 0.79 on 117 public motivational-interviewing demonstrationspre-registered
No behaviour invented in 126 rescores with the supporting evidence removedmeasured
9 of 9 helpful and unhelpful public role-play pairs ranked the right way round in every runmeasured (direction, not accuracy)
Word error rate 6.1% for on-device transcription, on 19 public videosmeasured

Stack · links

PythonQwen3.6-27B, run locallymlx-whisperpyannote
last updated 2026-10-03 · from source, not marketingbuilding · no result is claimed