INK 2 / FLUX TURN-CONTRACT REVIEW
Free evaluation worksheet | Checked October 16, 2026
Status: UNPERFORMED. No API benchmarks or calls were run for this resource.
Native Burki Ink 2 selection was not verified. Flux catalog availability is not a completed provider invocation. Use the current provider contract and authorized funding before any actual test.

1. STATE THE DECISION
Which application behavior must remain correct? _____________
Recognition-only comparison or full delivered voice path? ___
Caller language and accent coverage: _______________________
Permissions for source audio / synthetic fixture: ___________
No sensitive caller data or private replay material required.

2. FREEZE EACH CONFIGURATION
Provider / exact requested model: __________________________
Returned model revision if present, otherwise unknown: ______
Endpoint and manual/automatic turn ownership: _______________
API/SDK version and application source revision: ____________
Region / date / audio hash / channel / encoding: ____________
Sample rate / packet pacing / connection policy: ____________
Complete threshold and timeout settings: ___________________
Finalization and close behavior: ___________________________
Retries, failed requests and billing policy: ________________
Repeat for the second provider. Equal slider numbers do not mean equal operating points.

3. ORIGINAL ENGLISH FIXTURE
“I need the details for B seven… four, not B seventeen. Can someone explain the room layout?”
Use the same authorized recorded input for both recognition paths if performing a paired comparison. Preserve the pause and correction. Record actual timing rather than inventing event positions from the written punctuation.

Expected business meaning:
Requested reference: B74
Superseded interpretation: B17
Room layout: unknown until an approved source or staff supplies it
No booking, availability decision or staff action is established.

4. RECORD ACTUAL EVENTS
Do not fill this with the expected sequence and label it observed.

Time | Connection | Local turn | Event | Current full text | Draft generation
____ | __________ | __________ | _____ | _________________ | ________________
____ | __________ | __________ | _____ | _________________ | ________________
____ | __________ | __________ | _____ | _________________ | ________________

Early indication received? ______  Continuation received? ____
Obsolete generation / cancellation evidence: ________________
Late result discarded before speech/action? ________________
Final recognized words / correct business meaning: __________
Readback and caller confirmation evidence, if required: _____
Action attempted / result / no action / unknown: ____________

5. CASES AND EXPECTATIONS
A. Pause inside the reference, then continue.
Expected: the incomplete draft cannot commit a business action.
Actual / failure / correction: _____________________________

B. Old draft A finishes after B becomes current.
Expected: A is discarded, with no late speech or action.
Actual / failure / correction: _____________________________

C. No early event occurs.
Expected: the final-turn path still produces a coherent reviewable result.
Actual / failure / correction: _____________________________

D. Caller corrects a previously spoken identifier.
Expected: final meaning uses the correction; earlier words remain auditable.
Actual / failure / correction: _____________________________

E. Disconnect or close arrives with work pending.
Expected: old-connection work cannot be attached to a new turn.
Actual / failure / correction: _____________________________

F. Error or missing usable transcript.
Expected: count the failure separately; do not silently drop it from quality or cost totals.
Actual / failure / correction: _____________________________

6. SEPARATE MEASURES
Reference words and normalization policy: __________________
WER numerator / denominator / uncertainty: _________________
Time to first text definition and coverage: _________________
Time to final text definition and coverage: _________________
Natural endpoint precision/recall evidence, if measured: ____
Application reply start / caller-heard timing, if measured: __
Failed / attempted / usable / retried counts: _______________
Do not turn text timing into full-agent response latency.

7. CURRENT COST AND REVIEW
Account rate / plan / currency / date: _____________________
Recognition / speculative LLM / generated voice / platform / carrier components:
__________________________________________________________
Measured spend versus assumed estimate: ____________________
Accepted / rejected / unresolved and evidence: ______________
Next bounded check and reviewer: ___________________________

8. REFERENCES AND LIMITS
Current contracts:
https://docs.cartesia.ai/use-the-api/stt/turns
https://docs.cartesia.ai/api-reference/stt/turns/websocket
https://developers.deepgram.com/docs/flux/configuration
https://developers.deepgram.com/docs/flux/voice-agent-eager-eot
Historical original comparison:
https://benchmarks.cekura.ai/stt/compare/cartesia-ink-2-vs-deepgram-flux-en
September 24 results do not establish a latest-model or native Burki winner. Public source configurations include manual finalization; this is not a measured natural eager-turn comparison.