BURKI STANDARD-PIPELINE INTERRUPTION STYLE WORKSHEET
Prepared October 18, 2026. Free resource; real voice/evaluation usage can cost money.

Purpose: record a specific preset change and review one conversation reliably.
No provider call, listening evaluation or settings save was performed here.

1. RECORD THE BASELINE BEFORE CHANGING ANYTHING
Authorized operator and assistant:
Current saved/published state and release process:
Conversation engine:
Recognition provider/model and language:
Language model/provider:
Voice/model:
Browser or telephone route, device and relevant environment:
Current Interruption style:
Current Fine-tune interruptions values and any deliberate overrides:
Specific symptom, with actual evidence if available:

Example useful symptom: a whole acknowledgment repeatedly stops an explanation.
Avoid a broad target such as make it natural without a concrete review question.

2. FIND THE APPLICABLE CONTROL
Voice -> Model and provider choices -> Conversation engine.
Current engine labels:
  Standard pipeline · advanced provider choices
  GPT Live · default for new assistants

The presets in this worksheet belong to the Standard pipeline.
GPT Live handles its conversational interruptions natively and hides cascade
controls; preserved cascade settings are not native GPT Live tuning.
Do not migrate engines solely to see a control without reviewing cost,
credentials, funding, model and supported-action requirements.

Standard-pipeline path: Workflow -> Call behavior -> Interruption style.
Options: Fast barge-in, Balanced, Patient listener.
Selecting a style applies grouped preset values, replacing several prior
fine-tuned values. Record intentional overrides before selecting another style.
Some separately saved overrides can still take priority when configuration is
resolved. For imported/API-managed settings, review the full saved configuration
through the authorized process; a label does not prove every effective default.

Fine-tune interruptions includes Minimum speaking time (sec), Cooldown between
interruptions (sec) and Filter filler sounds. Do not turn their numbers into
measured audible-stop latency. Timing, confidence and filtering checks interact.
Minimum speaking time concerns assistant speech age, not caller-speech duration.

Active listening is a separate form field labelled for assistant acknowledgments.
Filtering caller filler/acknowledgments is a different question. Do not change
unrelated settings during a single-variable comparison.

3. WRITE THE PROPOSED CHANGE
Proposed style:
Why this style is relevant to the observed symptom:
Preset values after selection:
Overrides to retain deliberately, and why:
Unrelated settings kept unchanged:
Required current publication/release review:

Style review questions:
Fast barge-in: does it handle genuine corrections readily without unwanted stops?
Balanced: do whole acknowledgments avoid needless restarts while corrections work?
Patient listener: does more guarded handling avoid false stops without ignoring
real corrections?

These are questions, not measured results or recommendations by demographic.
Explicit stop/correction phrases can take a different path from acknowledgments.
Current local phrase rules include English examples; no multilingual guarantee.
Recognition providers differ in confidence information. Do not infer accuracy
or caller-heard behavior from a preset name, helper text or a saved setting.

4. FICTIONAL HOURS CONVERSATION PLAN
Alder Service Desk is invented. Approved fact: Saturday hours 9 AM to 1 PM.
Assistant gives a short weekday-hours explanation.
Proposed starting style for this example: Balanced, not a tested winner.

Case A: caller says right as a whole acknowledgment during the explanation.
Expected review question: was there an unnecessary stop/restart?
Case B: caller says Actually, I mean Saturday.
Expected answer uses Saturday facts, without continuing the abandoned weekday answer.
Case C: caller clearly asks the assistant to stop during speech.
Expected review question: how did it yield and continue under this actual route?
Case D: caller pauses in a longer unfinished thought.
Expected review question: is this turn-completion trouble rather than barge-in?

Illustrative answer for Case B: Saturday is 9 AM to 1 PM.
No such output or timing was measured in a call for this worksheet.

5. COMPLETE ONLY AFTER A SEPARATELY AUTHORIZED EVALUATION
Case and baseline/proposed configuration:
Date/time and actual provider/model versions:
Prompt/facts/device/route held constant:
Expected behavior:
Actual heard behavior and supporting evidence:
Recognized caller text, if available and permitted:
Corrected fact retained in answer:
Stale speech, unnecessary restart or ignored correction:
Timing definition and actual measurement, if legitimately measured:
What remains unverified:
Decision and responsible reviewer:

Do not score a fast stop as full success when the answer uses the old fact.
Do not use transcript timestamps as audio-stop measurements. Compare the same
route and conditions; browser and telephone observations are not interchangeable.

6. BEFORE A REAL RELEASE
Review current Publish status. Saved and published telephone versions differ
where versioned publication is in use; legacy live settings can affect calls
differently. Follow the actual authorized release process and retain the prior
configuration plus review reason. A saved preset is not listening acceptance.