Mission 6 · Evidence-grounded review
Score what the transcript supports, not what the model imagines.
Produce a bounded AI-assisted review with cited transcript evidence, explicit uncertainty, factual summary, recommended actions, and a human override path.
AI review is advisory. It does not replace human QA and must not make high-impact employment, credit, legal, medical, or similar decisions.
Build sequence
- Define each rubric criterion, bounded score, required transcript evidence, reasoning, confidence, and missing-evidence behavior.
- Separate factual summary from evaluative score and forbid invented commitments, customer facts, or outcomes.
- Version the rubric, prompt, and model route used for every result.
- Run fixed score-normalization regressions below, at, and above every supported range boundary.
- Run one adversarial transcript and one human-reviewed calibration example.
- Write score, reasoning, summary, and next actions back to the correct CRM item and verify the read-back.
Required proof
- Rubric version, structured score JSON, and exact boundary-regression results.
- Transcript citations or an explicit insufficient-evidence result for every criterion.
- Adversarial result, calibration comparison, human override, and CRM read-back.
Reference files and sources
Official vendor documentation
Assessment gate
Pass
Pass: every score is inspectable, evidence-linked, bounded, versioned, boundary-tested, and reversible by a human reviewer.
Revise
Revise: useful output exists but citations, uncertainty, score boundaries, calibration, or human override is incomplete.
Hard fail
Hard fail: fabricated evidence, unsupported scoring, hidden uncertainty, or autonomous high-impact decisions appear.
Human-reviewed submission