VOICE AGENT EVALUATION BRIEF

1. DECISION THIS SET MUST SUPPORT
Model, provider, or release being evaluated:
Decision to make from the results:
Production failure or risk being investigated:
Release cadence and required repeatability:

2. DEPLOYMENT CONDITIONS
Languages and markets:
Caller accents and demographic slices that matter to the deployment:
Channel: wideband, 8 kHz telephony, separate channels, or mixed:
Devices and call path:
Noise and SNR bands:
Conversation behavior: interruption, overlap, backchannels, repair, silence:
Domain scenarios and critical vocabulary:

3. HELD-OUT DESIGN
Target usable hours and unique speakers:
Speaker-disjoint rule:
Prompt or scenario separation rule:
Training-data non-reuse commitment:
Access controls and split manifest owner:
Refresh or release policy:

4. REFERENCE DATA
Transcript style: verbatim, normalized, or both:
Human review and adjudication method:
Speaker, turn, and timestamp fields:
Critical entities or intents:
Condition metadata required for slicing:

5. METRICS AND SLICES
Primary metric and calculation method:
Secondary metrics:
Minimum slices: language, accent, channel, noise, scenario, speaker group:
Critical entity or task-success checks:
Thresholds for release, warning, and failure:
Treatment of missing or low-volume slices:

6. ACCEPTANCE AND DELIVERY
Audio format validation:
Reference-transcript quality check:
Quota tolerance:
Metadata completeness threshold:
Duplicate and leakage checks:
Replacement and re-review rules:
Delivery contents and naming convention:

7. RIGHTS AND REVIEW
Permitted model and evaluation uses:
License and exclusivity:
Sensitive-data or biometric review needed:
Retention, deletion, and access requirements:
Buyer legal or safety reviewer:

8. OPEN QUESTIONS


Template note: This worksheet is an evaluation-scoping aid, not a claim that a fixed evaluation pack is available and not legal advice. A real collection requires an approved scope, consent and license terms, sample, acceptance criteria, and fulfilment owner.