A project built in one house. The code, the measurements and the open problems are public.
Source on GitHub ↗A home system, built for a hackathon
Lanternina invents an afternoon for an adolescent, has a parent approve it, and then runs it in a real room — on paper, on e-paper displays, and on physical buttons. It reads what comes back in handwriting, decides what happens next from that, and stops at an hour the parent set.
A mini-PC in a house drives two e-paper displays, an inkjet and a flatbed scanner. Every language and vision call goes to Microsoft Foundry in the EU. No model runs on the device, and there is no offline mode. The house holds the clock and the sealing keys, and it is the only thing that starts work.
There is no layout engine. An image model is given every string that has already passed the safety gate, told to letter exactly those characters, and returns a finished A4 page. The design work is the prompt, and the page format lives in one Markdown file rather than in code.
No QR code, no ArUco markers, no declared grid — nothing printed on the paper that is there for a machine. Reading is handing a model two images, the blank and what came back off the glass, and asking what is different. An earlier version had all three, and they were deleted.
The dashboard is inert by construction: approving writes a row and does nothing else. The house asks, on its own clock, and the answer arrives inside the reply. There is no inbound path to a device in a room, which is also why there is nothing on the home network to attack.
An agent has nowhere to record that its own output is acceptable, becauseProposal has no status field. Only the safety gate can mint the type a user-facing payload requires. Each boundary test was mutation-checked: the violation was injected and the test watched to fail.
The page below was drawn whole by an image model, from words that had already been screened. Somebody writes on it and puts it back on the scanner glass; the reading decides the next moments, which are written while the sheet is still on the table.
The reader's vocabulary is deliberately poor: a page came back with marks orblank, and nothing else. Not a count and not a fraction, because a count of somebody's marks is one step from a score. The reading is then gone — the type refuses to be pickled, copied or cached.
Dates are attached because prompts change and the numbers move with them. Everything here was measured against the real services, not estimated.
| Devising a whole afternoon | 68–157 s | n=22, Sep 2026 |
|---|---|---|
| Rereading it with a second model | 14.4–22.8 s | Sep 2026 |
| Drawing one page | 19–33 s | Aug 2026 |
| Ink on a drawn page | 0.5–2.7 % | Aug 2026 |
| Reading a returned page | 4.4–5.5 s | Aug 2026 |
| Button press to reading on the display | ≈ 35 s | 19 Aug 2026 |
| Printed ruler, measured with a ruler | 50 mm asked, 50 mm got | 4 Aug 2026 |
The last row is the one that is not about speed. The print chain does not rescale, and that was checked by hand — a chain that silently rescales would make every physical dimension on every page wrong in a way no test would notice.
Raspberry Pi Compute Module 5 on a PoE base board, Raspberry Pi OS. Two 7.5" e-paper displays on XIAO ESP32-S3, running a patched TRMNL firmware against our own BYOS server. Epson ET-2870 over IPP and eSCL, driverless. Frames 3D printed, units assembled by hand.
Microsoft Foundry in the EU for every language, vision and image call, behind a single router module that is the only place allowed to import a model SDK. Azure AI Content Safety on everything generated and on every inbound image. Container Apps for the API, Cosmos DB, Blob Storage, Static Web Apps. Infrastructure as Bicep.
A push to main publishes the API through a user-assigned managed identity with a federated credential, so there is no password anywhere in the pipeline — measured end to end at 1 m 35 s. The image is built on the runner rather than with az acr build, because the latter needs rights that would let the identity reconfigure the registry.
Engagement optimisation is easy to add to a system like this and would do harm, so it is a written rule rather than a matter of judgement. tests/test_boundaries.py refuses the vocabulary in the stored shape, in the panel, in the browser and in the prompts.
Why each boundary exists, and what it costs · every prompt, whole, as the model receives it
This site sets no cookies, collects no statistics, and loads nothing from anybody else's server.