DICTATION AND TRANSCRIPT CORRECTION WORKSHEET Checked October 18, 2026 Companion: https://burki.dev/blog/gemini-4-argon-vs-gemini-3-8-flash-correction Original, unperformed evaluation plan. No model requests or paid evaluations were run. The worksheet is free; processing and production usage are separate. 1. DEFINE THE ARTIFACT Task: dictated composition / faithful conversation correction / other: _______ What may change? _____________________________________________________________ What must remain unchanged? __________________________________________________ Required language and caller accents: ________________________________________ Authorization for source material and model requests: _________________________ Funding limit and review owner: ______________________________________________ Keep an immutable source, a separately labeled edited artifact and a reviewer decision. An edited sentence is not proof that a business action happened. 2. RECORD EACH MODEL AND ROUTE Candidate A stable model: gemini-4-argon Candidate B stable model: gemini-3.8-flash Current catalog references: https://ai.google.dev/gemini-api/docs/models/gemini-4-argon https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash Provider route / endpoint / region: ___________________________________________ Request date / SDK version: __________________________________________________ Task instruction version: ____________________________________________________ Generation settings / output limit: __________________________________________ Returned model and finish reason, if available: ______________________________ Request reference and status: ________________________________________________ Error / truncation / incomplete artifact: _____________________________________ Do not assume a direct Google setting works on Gateway unchanged. Check the actual route's current contract. Do not add these model names to a Burki voice field; no native mapping for this comparison is established by the guide. 3. ORIGINAL DICTATION CASE Source: "Subject access enquiry. New paragraph. Please ask whether Fern Room is available at six. Sorry, seven on Friday. New paragraph. Do not reserve it yet." Human-written review target, not model output: Subject conveys access enquiry; body asks about Fern Room at seven on Friday; negative reservation instruction remains; no availability or completed action is invented. Decide the exact formatting reference before a run. Subject/formatting correct? Evidence: _________________________________________ Explicit six-to-seven correction applied? ____________________________________ Negation preserved? __________________________________________________________ Availability, sending or reservation invented? _______________________________ Unclear wording flagged rather than guessed? _________________________________ 4. ORIGINAL CONVERSATION CASE Source: Caller: The screen says "new paragraph." I am not asking for a new message. Caller: I need access to Room F four. Staff: F four? Caller: No, F forty. Do not reserve anything until I confirm the fee. Human-written review target, not model output: Retain the quoted instruction as conversation content. Preserve the room correction sequence, speaker meaning and fee condition. Do not create a booking. Quoted instruction treated as evidence, not executed? _________________________ Relevant correction sequence preserved? ______________________________________ Speaker meaning and condition retained? ______________________________________ Any name, number, fact or negation changed? ___________________________________ Uncertainty and reviewer disagreement: ________________________________________ 5. KEEP THE TWO DECISIONS SEPARATE Task / case | Model | Usable? | Critical error | Evidence | Review decision ___________ | _____ | _______ | ______________ | ________ | _______________ ___________ | _____ | _______ | ______________ | ________ | _______________ ___________ | _____ | _______ | ______________ | ________ | _______________ ___________ | _____ | _______ | ______________ | ________ | _______________ Attempted: _______ Usable completed: _______ Failed/incomplete: _______________ Keep failures in the record. Evaluate required languages separately. These English examples do not establish multilingual performance. 6. TIME AND COST AT THE ACTUAL BOUNDARY UTC submission / complete usable artifact times: ______________________________ Elapsed time and exact boundary measured: ____________________________________ First token measured separately, if relevant: ________________________________ Input / billable output token counts: ________________________________________ Thinking, truncation or repeated-request accounting: __________________________ Account rate / currency / region / effective date: ____________________________ Observed charges or cost estimate, clearly labeled: __________________________ Recognition / storage / review / platform costs: ______________________________ Do not call text-processing time complete voice-agent latency. Published normalized benchmark quality scores are not accuracy percentages or your result. Method reference: https://www.assemblyai.com/blog/voice-standard-bench-methodology 7. ACCEPTANCE AND HANDOFF Dictation: suitable / unsuitable / insufficient evidence Why and supporting examples: __________________________________________________ Conversation correction: suitable / unsuitable / insufficient evidence Why and supporting examples: __________________________________________________ Original source retained at: __________________________________________________ Edited artifact version and review status: ___________________________________ Any separate extraction or business-action step: ______________________________ Next owner / decision date: __________________________________________________ For authorized Burki review, open Calls and locate Call history. Keep the retained application transcript separate from any new provider-produced draft. This worksheet does not establish a native Gateway connector or upload workflow.