Keep the evidence beside the score

A useful assessment should help a person inspect a decision, including the reasons to question it.

One number hides several questions

A role may resemble someone’s background while an eligibility constraint remains unresolved. A promising match may rely on an incomplete job description. A requirement may share keywords with a resume statement without being supported by it.

Job Copilot keeps role fit, eligibility, assessment confidence and job-description quality separate. Each signal answers a different question.

Show the source statement

The evidence view pairs a requirement with the strongest retrieved resume statement and its support status. Direct, partial and missing support remain visible so a reader can review what the system is claiming.

This also sets a boundary for generated application materials. A cover letter should be grounded in the existing resume, with the user reviewing it before use. A retrieval result is not permission to invent experience.

Evaluate the narrow task

The repository’s ML work studies evidence support and ranking. Its documented synthetic-teacher evaluations are measurements of agreement with that evaluation contract. They are not predictions of hiring outcomes.

Keeping evaluation claims precise makes the work easier to improve. A missing-support error suggests a different change from a ranking error, and neither can be diagnosed from a polished summary score alone.