work / care / matriculate← rare impact fund toolkitHelping-skills competency scorer →
● delivered · first part delivered May 2026
matriculate
Delivers a response rubric and a 500-record training corpus for a near-peer advising bot.
- job
- codify near-peer advising
- domain
- education · advising
- runs
- hosted · API
- class
- tool
- scale
- 500 records
- since
- 2026
One call
- 01PromptA scenario from a ten-category prompt library
- 02GenerateA reply written in a near-peer adviser's voice
- 03ScoreScored against the response rubric
- 04GateFailures excluded; crisis failures go to a person
- 05TemplateNames replaced with placeholders filled at deployment
- 06SplitTraining, validation and test sets
In · out
inputs
- advising-fellow survey responses
- advising code and policy documents
- prompt library (10 categories)
outputs
- response rubric (12 dimensions)
- 500-record training corpus
- crisis and distress playbook (draft)
- college-process answer key (draft)
Where the person is
Any reply that fails the safeguarding floor is excluded; failures in crisis categories go to a person for review.
code decides◌ not stated
What is measured
Against: A 12-dimension response rubric on a calibration sample; expert rating not yet done.
500 qualified training records delivereddelivered
463 of 499 generated replies passed the quality gate (92.8%)measured
A fast five-test scorer agreed with the full 12-dimension rubric 90% of the time on the calibration samplemeasured (model against model)
With
MatriculateProject Evident
Stack · links
Claude Opus (generation)Claude Sonnet (scoring)PythonJSONL