EXPERIMENT 002 / OPENAI REALTIME / 14 OCT 2026

The receptionist
is alive.

One voice. Two native tools. A very real answer
from a very fictional hotel.

LATEST APPROVED RECORDINGPartial playout

A room with a voice

GPT-Realtime 2.1 · Marin · Caller + agent

48 kHz
MONO / PCM

0:000:27

Loading the specimen…

The conversation

FOLLOW THE AUDIO

CSynthetic caller

Original caller WAV in main playback · text is source prompt, not provider ASR

⌁MarinCAPTURED OUTPUT

Caller + agent use original recordings with approximate alignment. Segment highlighting from offline ASR. No word timestamps were retained; transcription errors are preserved.

◧

The tape ends here. The session cap cut the closing explanation. No booking was made.

THE VERDICT

Good answer.
Unfinished sentence.

The model checked real deterministic tools and spoke the correct availability and price. The provider accepted hangup and local media closed. Full spoken completion remains incomplete.

✓ Grounded price✓ Native tool receipts◧ Closing tail cut

UNDER THE MICROSCOPE

Every claim has a paper trail.

Read-only · no provider sessions

TURN → BUFFER ONSET383 ms

Browser observation, not audible latency

ESTIMATED USAGE COST$0.1364 USD

Actual bill unknown · unreconciled

NATIVE TOOLS / BOOKINGS2 / 0

Availability + quote · no approval

Experiment configuration Model, voice & limits

Event journal Provider ≠ application

Evidence & integrity Original hashes and downloads

These are approved fictional demonstration excerpts. Original private journals remain outside this deployment. Hashes allow independent comparison; they do not prove every claim by themselves.

Source acceptance report ↗ Private GitHub access required Download publication manifest ↓Download sanitised replay evidence ↓

What this experiment cannot tell us Known limitations