acceptance-criteriadraft

Acceptance Criteria: Trustworthy analysis → editable send (US-2 + US-3 + US-4)

Story Context

Covers the core loop: a consented, depersonalised session becomes a reproducible, evidence-quoted analysis and an editable client summary the coach explicitly sends. Assumption: ASR/diarization and the analysis pipeline are integrated; reproducibility threshold is provisional (F1 ≥ 0.70).

Happy Path

AC-1: Analysis with evidence

Given a depersonalised, speaker-separated transcript, When the coach runs analysis, Then the system returns topic, insights, commitments, and next steps, each with ≥1 linked transcript quote.

AC-2: Edit and send

Given a generated draft summary, When the coach edits it and presses Send, Then the summary is delivered to the chosen channel and marked sent + logged.

Edge Cases

AC-3: Uncertain diarization

Given overlapping or low-quality audio, When speaker attribution confidence is low, Then affected turns/conclusions are flagged low-confidence (not silently asserted).

AC-4: Re-run reproducibility

Given the same session, When analysis is run again, Then extraction sets agree within the defined noise band; divergence beyond it is surfaced as a quality warning.

Error States

AC-5: Depersonalisation guarantee / failure

Given a transcript about to be analysed, When any external LLM call is made, Then identifiers are replaced with placeholders first; if depersonalisation cannot be verified, processing halts and the coach is told — no raw identifiers leave the protected perimeter (PII-leak audit = 0).

Given a session without recorded consent, When the coach attempts analysis, Then the system refuses, explains the consent requirement, and offers to capture consent.

AC-7: Send failure

Given a Send action, When the delivery channel errors, Then the summary is NOT marked sent, the draft is preserved unchanged, and the coach can retry.

Non-Functional Criteria

AC-8: Reproducibility threshold

Given the pilot gold sample, When F1 is computed on extractions, Then F1 ≥ 0.70 (provisional client-safe gate) — below it, output is not promoted to "sendable" status.

AC-9: Human-in-the-loop invariant

Given any client-facing output, When it is produced, Then it is never delivered without an explicit coach Send (no automatic client delivery anywhere in the system).

Notes

  • AC-8 threshold is inherited from the plan page and must be re-validated with coaches — flagged Unknown.
  • "Light edits" as a trust proxy is tracked separately in 10-metric-design-experimentation.md.