BURKI LLM FALLBACK SETUP WORKSHEET Free planning resource | prepared October 23, 2026 Original guide: https://burki.dev/blog/burki-llm-fallback-provider-setup This is an unperformed setup and evaluation plan. No keys, paid requests, induced failures or saved settings were used to create it. Free download does not mean free production inference or calls. 1. TASK AND CURRENT VERSION Owning workspace / assistant: _______________________________ Task: _______________________________________________________ Conversation engine: ________________________________________ Required for this guide: Standard pipeline. GPT Live does not apply native fallback providers. Browser or exact telephone route: ___________________________ Current saved version / published state: _____________________ Legacy live settings warning? ________________________________ Approved test budget / authority / owner: _____________________ 2. KEEP ONE ANSWER CONTRACT Fictional Alder Reception approved fact: Reception is open Monday to Friday 09:00–17:00 in the business timezone, and closed on public holidays. Question: Can I visit reception at noon on a public holiday? Expected meaning: Reception is closed on public holidays. No exception or appointment is invented. Your approved facts and source/version: ______________________ Same question used for each candidate: _______________________ Expected answer meaning / unacceptable additions: ____________ _____________________________________________________________ 3. CANDIDATE INVENTORY, NO SECRETS Configuration path: Assistants > saved assistant > Voice > Model and provider choices > Fallback providers. Conditional form: enable switch, Add fallback provider, Fallback #1 > Provider / Model / API Key; Base URL for Custom Provider. Do not paste a provider router's models/ttft_timeout JSON into this form. Role Exact provider/model Credential owner/reference Primary ________________________ __________________________ Fallback1 ________________________ __________________________ Optional2 ________________________ __________________________ Optional3 ________________________ __________________________ Do NOT put API key values here. Record a secret reference/owner instead. For each candidate record account access, billing mode and approved rate source: Primary: ____________________________________________________ Fallback1: __________________________________________________ Optional candidates: _________________________________________ Provider change reviewed for model/key compatibility? ________ Actual source of credentials verified? _______________________ An explicit fallback key uses its own external credential path. A missing BYO key cannot silently borrow platform credentials. Managed use needs its own eligible configuration. A listed model is not proof that an account accepts it. 4. SHARED ALLOWANCE AND ORDER Order reviewed: _____________________________________________ Remaining shared token allowance and deadline: ____________________ Platform/provider/telephone cost components reviewed: _________ Externally billed costs missing from platform total? _________ Candidates share a session allowance, not fresh per-candidate budgets. If an admitted backup is selected, later turns in that session keep that choice. The controller does not periodically probe the primary. A valid candidate can still lack remaining admitted capacity. 5. PROPOSED ACCEPTANCE CASES Use existing approved evidence or later separately authorized evaluation. Do not corrupt live keys, induce paid failures or run traffic merely to complete a worksheet. Leave unperformed cases UNTESTED. For each record expected / actual / exact attempt evidence / reviewer decision. A. Primary completes normally. Expected: no unnecessary replacement; public-holiday answer follows approved fact. Actual/evidence: _____________________________________________ Reviewer/correction: _______________________ Status: UNTESTED B. Supported definitive rejection before any response chunk. Expected: next admitted candidate may be tried, subject to budget and exact rejection/closure evidence. Do not infer this from a generic error label. Actual/evidence: _____________________________________________ Reviewer/correction: _______________________ Status: UNTESTED C. Later turn after backup was selected. Expected: same selected backup remains sticky; no background primary probe. Actual/evidence: _____________________________________________ Reviewer/correction: _______________________ Status: UNTESTED D. Primary emitted text, tool, name or usage information before failure. Expected: no automatic replay through a backup for that operation. Actual/evidence: _____________________________________________ Reviewer/correction: _______________________ Status: UNTESTED E. Generic timeout, network uncertainty or HTTP 5xx. Expected: not a qualifying definitive rejection; no automatic replay. Actual/evidence: _____________________________________________ Reviewer/correction: _______________________ Status: UNTESTED F. Shared budget exhausted before backup request. Expected: configured row does not authorize extra allowance. Actual/evidence: _____________________________________________ Reviewer/correction: _______________________ Status: UNTESTED G. Caller interrupts delivery. Expected: interruption/cancellation alone does not become a provider fallback attempt. Review subsequent new turn separately. Actual/evidence: _____________________________________________ Reviewer/correction: _______________________ Status: UNTESTED 6. EVIDENCE AND RELEASE RECORD Known rejection: provider / official endpoint / status / supported meaning / operation or request reference / clean closure: _____________________________________________________________ Never copy a raw secret, private prompt or unneeded customer data into this record. Uncertain output/usage and remaining question: ________________ Exact deployed voice-worker source, if actually verified: _____ Provider accepted the intended request? Evidence: _____________ Caller-heard result, if relevant and authorized: ______________ Saved draft version: _________________________________________ Publication action/version, if authorized: ___________________ Actual route acceptance: _____________________________________ Owner and next review date: __________________________________ Configuration/source tests do not prove a real deployed failover or audible caller outcome. No universal outage-recovery or zero-cost promise. Official error context: https://platform.claude.com/docs/en/api/errors Cost components: https://burki.dev/pricing