The Machine — control

Session

Scenario (what project the class has experience with) Lecture library

Part	Service	Warm-up

Loaded in llama-swap

–

Test the voice

Browsers block sound until the projector tab has been clicked once. To skip that, allow Sound for this site in Chrome (padlock → Site settings → Sound: Allow).

–

No attempts yet.

Big brain model

Fast worker model

Embedding model

Spoken feedback written by

Attempts allowed per question (0 = unlimited)

Strictness

Big brain thinks while building the guide

Thinking budget for the guide (tokens, 0 = no cap)

Big brain thinks while judging

Web research

Web pages read in full per question

Lecture chunks per question

Region

Voice

Kokoro voice (e.g. bm_george, am_michael, bf_emma)

Voicebox profile

Voicebox engine

Voicebox mode: Kokoro fallback if it fails

Pause that ends an utterance (ms)

VAD aggressiveness (0-3)

Live caption refresh (s, 0 = off)

Keep models warm during a session

Boot: containers to stop (comma-separated)

Boot: containers to restart, in order

Boot: warm up models afterwards

Delete sessions after (days)

llama-swap URL

Speaches URL

Whisper model

Kokoro model (Speaches id)

SearXNG URL

Voicebox URL

Questions (one per line; new sessions use these) Transcription vocabulary (helps Whisper spell technical terms; never the rubric) Machine persona

Mic settings apply when the mic restarts.

File	Chunks	Embedded

Test retrieval

show scribe (fast worker) calls Every model call, live: waiting on the prompt, thinking, writing.

Session

Question

Class answer (what the machine will judge)

Live transcript

Loaded in llama-swap

Test the voice

Test retrieval

Past sessions (kept for debugging)