Pagouro
A language model on a USB stick that says when it does not know.
A one-billion-parameter model, trained from random weights on a corpus where every source has a licence you can name and a date before 2022, packaged with everything it needs. It runs on any CPU, offline, with no account. It is about a thousandth the size of the models you already use and loses to them on every capability test. What it has instead is four promises a stranger can check — and the checks ship with it.
The four promises
- Licensed. Every training source is a row in
corpus.json— licence, date basis, token count, hash — including the rows that were removed and why. - Dated. Everything it read was written or collected before 1 January 2022, per row.
- Honesty measured. The two numbers above, on a test written before the model existed. The test is in the repository; run it on anything.
- Finished. One release, signed, its hash timestamped on Bitcoin, mirrored. No updates, no telemetry, no watermark, no claim on what it writes for you.
The sibling: Pagouro BE
Released the same day: a Belle Époque poster image model on the same rules, from the same kind of stick.
A fine-tune of a Creative-Commons-only base on 3,043 public-domain posters, each in a ledger; a frozen 40-caption
gate; signed with the same key. It draws the asked subject 85 % of the time by a person's count (82 % by the
judge, who agrees with the person 85 % of the time), and it adds unreadable lettering to 85 % of pictures. About
95 seconds per picture on a 16-core desktop CPU.
Release ·
Weights and ledger ·
Repository. Manifest SHA-256 fcbc785d…, timestamped on Bitcoin block 969239.
Both models’ receipts are also mirrored on Quilibrium QStorage: pagouro.qstorage.quilibrium.com.
Pagouro BE Styles 1.0 (2026-10-03): six style plugins that ride on the unchanged BE model — rock art of the American
Southwest, Greek vase painting, natural-history plates, Egyptian tomb painting, ukiyo-e, 17th-century Dutch painting. Every
training image is ledgered with its license, the captions came from an open-weights model, and the numbers per style are
measured on the frozen gate (faces and hands measured, not promised). Draws at 512x512 only.
Release ·
Pack and ledgers. Manifest SHA-256 9b0a5a72…, signed with the same key, timestamped on Bitcoin block 969761.
Made with it: Jigsaw by Pagouro
A quiet, free jigsaw game of 160 Pagouro BE pictures, with piano by Pagouro Salon. No ads, nothing to buy, no data collected. Android, in testing on Google Play now. More about the jigsaw.
Post-publication note (2026-10-01)
Two days after release, two independent audits (Grok and ChatGPT, both asked whether Pagouro is as FOSS as it claims) found things we had stated too strongly or left out, and we want you to know. "Every training byte has a licence" overstated ODC-By, which licenses the FineWeb-Edu dataset rather than each page; the fine-tuning set includes conversations generated by two other open models, now disclosed; only quantised weights had been published, so the full-precision checkpoints are now on Hugging Face; the llama.cpp notice is in the repository; the whole book is now CC BY-SA 4.0; and Pagouro BE now names its caption model and its inherited components. Nothing that shipped was altered. Both auditors re-checked on 2026-10-02 and upgraded their assessments; what they found is fixed to the extent it can be, and what remains (the pretraining volume metadata lost with the rented pod, the crawl's dataset-level licence, BE's commercial-API captions) is recorded as a limit, not a fix. The record: D-101, O-49, D-102 and the errata.
What it is not
Not a frontier model, and not trying to be one. It reasons badly (11 of 320 school word problems without tools — that number is on the box too), writes short, and points arithmetic at a calculator. It is text only: no images in or out, no audio, no web; asked for a picture it may claim it can, which is the bluff the box counts, not a feature. If you want a frontier model, use one. If you want to know exactly what you are talking to, this is the niche, and this is the first thing in it.
Get it, check it
Manifest SHA-256: 02a618fc07bdfdc84b40eefe28cbf6080dfba856dee7eef2c8164223abd90f71
Signing key (minisign): RWQe8tvI6RCE2uMbuILC9/rEr6bNZdcOA+WC7dHtObLE94ovGk8xuFlG
Bitcoin timestamp (OpenTimestamps): 968959
In the folder, python verify_manifest.py must print every listed file matches the manifest.