Rehearse ChatGPT Voice with the evidence on screen
Prepared on 20 September 2026 for the Monday meeting and the 30 September seminar. This is a rehearsal sheet, not a record of a successful Voice or room-audio test. Use public sample questions; do not enter claimant details.
What already works
The Monday handover contains the fixed care-home question and its evidence replay link. The public service has passed actual external-client and browser delivery checks. Those checks do not establish that the presenter's ChatGPT Voice session can call the service.
There are three separate things to observe:
- Voice: the app hears the question and speaks a response.
- Evidence access: that session actually calls Ask OKF and receives the source version, context identity, records and limitations.
- Room presentation: the audience can hear the response and read its sources.
A plausible spoken answer, screen reading or a successful text-mode call does not establish all three. Remote MCP and browser WebMCP are also different access routes; record which one was actually used.
Five-minute desk rehearsal
-
Open the current care-home demonstration. Recreate the evidence and keep its source text, gaps and context ID visible.
-
In the ChatGPT desktop app, open the intended task and use Start voice chat if available. First-use microphone setup requires the presenter. The current official Voice guide, checked on 20 September, explains availability, microphone setup and the separate dictation feature. Do not change accounts or permissions merely to continue an unattended test.
-
If Ask OKF is already connected, use this instruction in the voice session:
Use Ask OKF for the care-home question shown in the handover. Identify the exact source version and context ID, read the evidence and its gaps, then explain only what that evidence supports. Keep the no-partner qualification visible. If you cannot call the tool, say so rather than answering from memory.
-
Inspect an actual tool call and its returned evidence. Confirm the version, question and budget before comparing it with the browser package. A different budget can legitimately produce a different context. Follow the connection guide if setup is needed; connection and Voice acceptance still need separate checks.
-
Ask the AI to identify an unanswered part, then interrupt it and ask it to repeat the source limitation. Stop the voice session after the rehearsal.
Do not infer tool access from a voice product name, a subscription or an appshot. The exact call is the evidence. The service's insufficient status must remain visible in the explanation. A real call still needs claim-level review; it is not an entitlement decision.
Room rehearsal with the presenter and AV operator
Connect the actual projector and public-address system, then deliberately select the microphone input and sound output in the Mac's sound settings. Check the routes again after reconnecting cables. Agree a reachable mute control and an operator responsible for recovery. If a mixer feeds the AI microphone, keep the AI's own output out of that feed and test for feedback.
Read a source paragraph from the back of the room. Check the AI's volume, interruption, network delay and recovery. Keep a visible text equivalent of the spoken response. These observations require the real room and equipment; the agent has not changed audio settings or granted microphone access.
Record the result without personal data
| Check | Record after the rehearsal |
|---|---|
| Host | App version, desktop host and whether the intended task offered Voice |
| Connection | Remote MCP or WebMCP; actual tool names and visible invocation |
| Evidence | Question, source commit, budget and returned context ID |
| Answer | Claims checked against headings, conditions, exceptions and cited records |
| Boundaries | Insufficiency, ambiguity and truncation stated; unsupported follow-up handled |
| Audio and projection | Device routes, readable sources, feedback/mute/interruption outcome |
| Result | Passed, failed or not run for each check; date and observed limitations |
Do not publish account identifiers, recordings of bystanders or raw private logs. This blank table is not a completed acceptance receipt.
Reliable fallback for Monday
Use the verified public evidence replay and narrate the selected passage, provenance and gaps. If the network fails, use the retained screenshots and context package linked in that handover, clearly labelled as a recorded observation. Describe a text-only or recorded route as such; do not present it as a live Voice tool demonstration.
The original seminar integration plan retains the wider sequence towards the 24 September content freeze and 30 September seminar. Track the actual rehearsal under DWP-BL-016 in the work-package ledger.