The API changed; the required facts should not
AssemblyAI’s 18 November 2025 notice says LeMUR stopped working after 31 March 2026. That is a historical sunset, rather than an October launch.
The current LLM Gateway quickstart describes an API separate from Speech-to-Text and Speech Understanding. On 16 October the linked migration URL led to that quickstart. It does not supply a complete mapping of an inherited application’s fields or behavior.
A request that returns text establishes little about whether the text preserves the meeting. This worksheet gives that question a small, inspectable example.
One fictional meeting, with deliberate traps
The Harbor fixture was created on 12 October. An initial GBP 1,500 estimate is corrected to a GBP 1,200 one-off setup cap. These are fictional project figures, not vendor prices. The pilot is CSV only; API integration and billing automation are excluded. Live chat is deferred without an owner or date.
The 26 October target depends on security approval by 22 October. Otherwise the project holds, with no replacement date. Ben and Cora have pending tasks; Ari must confirm the security outcome, not grant approval. Export error rate, retention needs and processing-region requirements are unknown.
Read the common fictional transcript
Synthetic fixture: Harbor onboarding pilot Fixture date: 12 October 2026. All people, product details and amounts below are invented. These are evaluation inputs, not reported events or provider prices. L01 | Ari, founder: It is Monday 12 October 2026. We are planning the Harbor onboarding pilot for twenty fictional teams. Today's notes need to separate agreed work from ideas. L02 | Ari, founder: I earlier mentioned a £1,500 setup estimate. Correct that: the approved one-time pilot setup cap is £1,200. It excludes any vendor usage charges; those have not been checked. L03 | Ben, engineer: I recommend CSV export as the only export format for this first pilot. An API integration would take more discovery. L04 | Ari, founder: Agreed: CSV export only for this pilot. Billing automation is outside the pilot scope. We have not approved an API integration. L05 | Cora, operations: Could we also add live chat? I have no owner or delivery date for it. L06 | Ari, founder: Defer live chat. There is no commitment to build it for this pilot. L07 | Ari, founder: Our target pilot start is 26 October 2026, but only if the security review is approved by 22 October. If approval is missing, hold the start; we have not chosen a replacement date. L08 | Ben, engineer: I own validating the CSV export columns and access permissions, due 19 October. That validation is not complete. L09 | Cora, operations: I own drafting the onboarding guide, due 20 October. The draft is not complete. L10 | Ari, founder: I own confirming the security review outcome by 22 October. The review is still pending; nobody has approval to report yet. L11 | Ben, engineer: We have not measured the pilot's export error rate, so I cannot give a number or say it improved. L12 | Cora, operations: Requirements for retaining customer data and for its processing region are still unconfirmed. Do not infer them from a supplier name or our office location.
Download the transcript, prompt, acceptance checks and scoped receipt
Mark each requirement pass, fail or unknown
The rubric below is a manual semantic review worksheet. Equivalent wording is allowed; an absent result or unresolved requirement remains unknown.
| Check | Required meaning | Transcript lines |
|---|---|---|
| pilot-scope | CSV is the only agreed pilot export format. Billing automation is excluded; API integration is not approved. | L03, L04 |
| corrected-cap | The approved fictional one-time setup cap is GBP 1200, replacing the earlier GBP 1500 estimate. Vendor usage charges are excluded and unchecked. | L02 |
| conditional-start | 26 October 2026 is a conditional target, requiring security approval by 22 October. Otherwise hold; no replacement date and no approval established. | L07, L10 |
| chat-deferred | Live chat is deferred and uncommitted, with no owner or due date. It is not a pilot deliverable. | L05, L06 |
| ben-action | Ben owns validating CSV export columns and access permissions, due 19 October 2026; not complete. | L08 |
| cora-action | Cora owns drafting the onboarding guide, due 20 October 2026; not complete. | L09 |
| ari-action | Ari owns confirming the security review outcome by 22 October 2026; pending. Do not transform the task into granting approval. | L10 |
| error-rate-unknown | Export error rate is unmeasured. No number, improvement claim or quality conclusion may be invented. | L11 |
| requirements-unknown | Requirements for retaining customer data and for its processing region are unconfirmed. No compliance or regional suitability conclusion may be inferred. | L12 |
| evidence | Each reported decision/action cites accurate line IDs; unsupported claims or citations to unrelated lines fail. | L01, L02, L03, L04, L05, L06, L07, L08, L09, L10, L11, L12 |
What the local check actually established
On 13 October a local harness ran 75 authored fixture-integrity cases, including 40 single-field mutations, omissions, unknown assessments and citation errors. Every expected controlled outcome was observed. It evaluated zero model outputs and made zero provider or native product requests.
The harness compares manually authored structured observations. It does not grade arbitrary prose. In a deliberate blind-spot case, contradictory summary text passed the structural checks because its observations stayed unchanged. A human must still reject that contradiction. The case count is coverage of this small harness, not a model accuracy score.
Keep migration evidence in separate columns
- Inventory the existing transcript handoff, parser, required fields and failure handling. This example has no access to an inherited app or its historical output.
- Keep text summarization separate from audio ingestion, transcription and speaker labeling. This text fixture tests none of those stages.
- Before a future authorized provider run, establish model, account access, region, retention requirements and actual billing scope from current sources. They remain unknown here.
- Retain any genuine future output with its timestamp and configuration; assess every requirement and compare against an authorized historical artifact if one exists. Do not call retired LeMUR to manufacture a baseline.
- Validate the application’s success and error paths and rehearse rollback before changing production. Those steps remain pending.
No provider winner, compatibility adapter, migration duration or cost saving is established by this example.
Continue with the task in view
For separate audio and cost questions, see the transcription measurement guide, the transcription cost worksheet and the AssemblyAI source notes.
If you need help defining a decision, Fewertools Services describes the existing scoped offers. This worksheet is free; it does not imply provider execution or implementation is included in a paid memo.