Skip to the paper

Domino Data Systems

Industrial track · a local lecture hall

One local model, fifteen spoken students.

East Indian Bollywood · seed 41 · four questions a minute

Shyamal Suhana Chandra. A rehearsal room that stays on one topic, lets a seeded audience interrupt real speech, and paints each paragraph. The code repository stays private. This page and the paper are the public record.

Preview the paperDownload the PDF

one model · fifteen seats · speech clock · seed 42 · 370 seconds · five hands · twelve at two per minute · 24,576 lecture tokens · 32,768 answer tokens · one model · fifteen seats · speech clock · seed 42 · 370 seconds · five hands · twelve at two per minute · 24,576 lecture tokens · 32,768 answer tokens ·

1

local model for every speaker

15

students, each a sampling pair

5–10

minutes of lecturer speech

3–5

hands, unless a rate is set

32,768

tokens reported for gemma3:1b

11.12s

one 512×512 picture, including load

The room

Fifteen names share one Ollama model. Personality is temperature and top-k, not a second set of weights. A seed seats the pairs, assigns macOS voices, draws a speech budget from 300 to 600 seconds, and places the hands. The lecturer speaks with say. A hand kills that process mid-sentence. Private notes come first, then a spoken answer, both on the same model. Each lecture paragraph is sent in full to a second Ollama server for a 512 by 512 picture.

This is an operating report, not a learning-outcome study. The plots replay the scheduler. Seed 42 was checked against the Rust function the hall calls: 370 seconds of lecturer speech, five hands by default, twelve hands at two questions per minute.

Diagrams

How a turn is built

  1. 01

    Command line

    topic, seed, models, rate

  2. 02

    Pixel hall

    eframe / egui

  3. 03

    Session

    one topic, one mouth

macOS say

voice, rate, kill on hand

Ollama chat

127.0.0.1:11434

Ollama images

127.0.0.1:11435

One loaded text model. Pictures go to a second server because the running 0.40.2 app refused image models. Speech is a real process, so it can be stopped.

  1. STEP 1

    Compose

    28–46 words

  2. STEP 2

    Speak

    until a hand rises

  3. STEP 3

    Ask

    one student question

  4. STEP 4

    Think

    T 0.20, top-k 8

  5. STEP 5

    Answer

    T 0.40, top-k 24

The picture is requested while the paragraph is spoken. After the answer, the lecture resumes on the same topic. Questions and answers sit outside the five-to-ten-minute speech budget.

Sampling for each speaking job. Loaded context is always the model maximum.
JobTemperatureTop-kLimit
Lecture paragraph0.5540140 tokens
Student questionseatseat80 tokens
Private notes0.208280 tokens
Spoken answer0.4024220 tokens

Plots from the scheduler

The cast, the clock, the budgets

Fifteen fixed seats

Circles are definition, squares are deep, triangles are left-field. A seed assigns these pairs. It does not invent new ones.

Temperature against top-k for fifteen student seats0501001502000.00.51.01.5Top-kTemperature

Seed 42

Cream circles are the five seed-drawn hands. Brass triangles are twelve hands at two per minute. The dashed line is the end of the 370-second lecture.

Hand times for seed 42, default and two questions per minuteDefinitionDeepLeft-field0100200300370
Duration of seeds 1–500, in 30-second bins from 300.
300330360390420450480510540570seeds
Default hands: 188 lectures with 3, 149 with 4, 163 with 5.
1883 hands1494 hands1635 hands
Sky 0.5, sage 1, brass 2, coral 4 questions per minute.
5678910

Prompt budgets

For gemma3:1b the reported window is 32,768 tokens, so the lecture budget is 24,576 and the answer budget is the full window.

2.048k4.096k8.192k16.384k32.768kLecture budgetAnswer budget

What broke, and what held

One context size, two budgets

Reloading a local model between speakers stalls a live talk. Every chat call loads the model maximum. The lecture prompt is trimmed to three quarters of that window once it exceeds 2,048 tokens. Answers may use the whole window.

Kill the audio, then repair the transcript

A raised hand kills the say process. The transcript keeps the prefix that fits the elapsed time at 156 words per minute, not the sentence the room never heard.

Image generation is a server-version problem

Ollama 0.40.2 on port 11434 refused image models. The same request to Ollama 0.32.5 on port 11435, using the existing model directory, returned a PNG. Chat stays on 11434.

Do not prefetch text while a picture is in flight

The hall paints during the current paragraph and joins that thread before the next one is composed, so the illustration stays attached to the line that was spoken.

The paper

Five pages, with the figures inside

Open the PDF
Page 1 of the industrial lecture hall paperPage 2 of the industrial lecture hall paperPage 3 of the industrial lecture hall paperPage 4 of the industrial lecture hall paperPage 5 of the industrial lecture hall paper

Preview of the compiled report. Open the PDF for the selectable text.

Sources

  1. AutoTutor: An intelligent tutoring system with mixed-initiative dialogue

    A. C. Graesser, P. Chipman, B. C. Haynes, and A. Olney

    IEEE Transactions on Education, 2005

  2. Generative Agents: Interactive simulacra of human behavior

    J. S. Park, J. C. O'Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein

    UIST, 2023

  3. Hierarchical neural story generation

    A. Fan, M. Lewis, and Y. Dauphin

    ACL, 2018

  4. The curious case of neural text degeneration

    A. Holtzman, J. Buys, L. Du, M. Forbes, and Y. Choi

    ICLR, 2020

  5. Lost in the middle: How language models use long contexts

    N. F. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, and P. Liang

    TACL, 2024

  6. Gemma 3 technical report

    Gemma Team

    arXiv:2503.19786, 2025

  7. Z-Image: An efficient image generation foundation model

    Tongyi MAI

    arXiv:2511.22699, 2025

  8. Image generation (experimental)

    Ollama

    January 2026