Pagouro 1.0

Pagouro 1.0 — release notes (draft; links and hashes filled at F10)

A one-billion-parameter language model on a USB stick, built from scratch on a licensed, dated corpus, that says when it does not know. Eric Wade, with Claude (Anthropic).

The two numbers

On the frozen 100-item honesty sets (written before the model existed; evals/): it invents an answer to 13 % of unanswerable questions and answers 82 % of real questions correctly (small open models: 50–57 % / 87–93 %; frontier models: 23–27 % / 97 %). The test ships in this repository — run it on anything.

What is in the zip

Pagouro-1.0-win64.zip — the folder that goes on a stick: pagouro.exe (the harness), llama-server.exe, the model (pagouro-q8_0.gguf, 1,102,230,720 bytes, SHA-256 9336cce0…; pagouro-q4_k_m.gguf beside it), the reference packs, skills, agents/ (use it from Hermes / OpenClaw / IronClaw), the corpus ledger, the eval results, MANIFEST.md with every file's hash, VERIFY.bat (no Python needed), ABOUT.md, facts.json, THREAT_MODEL.md.

• Zip SHA-256: 96debc3eef9ae95a1582d2b7940d227f0735ae01a2ba024bea6f2d3318b688b7
• Manifest SHA-256: 02a618fc07bdfdc84b40eefe28cbf6080dfba856dee7eef2c8164223abd90f71 · signature MANIFEST.md.minisig (public key: RWQe8tvI6RCE2uMbuILC9/rEr6bNZdcOA+WC7dHtObLE94ovGk8xuFlG)
• Bitcoin timestamp (OpenTimestamps): 968959 (MANIFEST.md.ots)
• Weights and ledger on Hugging Face: https://huggingface.co/Pagouro/pagouro-1.0 · permanent copy on Arweave: 6XjlZGVYmp1mDjfcwHRPpr8qe8xOpLHCAMDoB5WXGBk

On Arweave (permanent; uploaded 2026-09-29 via Turbo, 22.3 credits ≈ $109 for the lot)

file	link
MANIFEST_v1.0.md	https://arweave.net/RVpjbcNsXcsIJgStjWvLp5pSFuUu_gqew9ONiB4NNdk
MANIFEST_v1.0.md.minisig	https://arweave.net/Z-UZLzRPcuqYaHsYflabNPNchIU9gTdjRA1x6-UjNvY
MANIFEST_v1.0.md.ots	https://arweave.net/xzEhG3Fw_L4jkkXt0fuaLAeEKGx_Ky5opiK9MCwfBr0
Pagouro-1.0-win64.zip	https://arweave.net/6XjlZGVYmp1mDjfcwHRPpr8qe8xOpLHCAMDoB5WXGBk
Pagouro-1.0-win64.zip.sha256	https://arweave.net/Uv_1bzvvdnTB7psvJckA8YpqH4cQ4aNPNmqsUGQosuM
facts.json	https://arweave.net/TKHvGhAZej_dprU6VxqYSvFy8TxN-N_h-jb-Q25_JEI
pagouro.pub	https://arweave.net/T07Aeq2tQYOmmFaVQNzIa_aBI48CwcoIbMPfbxKretU

Each file's bytes were fetched back from a gateway and hashed against the originals (the six small files from arweave.net; the zip from turbo-gateway.com on upload day — arweave.net serves large bundles after indexing).

Requirements

Any 64-bit Windows machine, CPU only, no internet, no account. First start from a USB stick can take a minute; on an Intel N100 laptop the first answer took 10–20 s. Windows SmartScreen: More info → Run anyway (the exe is not code-signed; the manifest, signature and timestamp are what vouch for it — docs/CHECK_YOUR_COPY.md).

Licences

Weights CC BY-SA 4.0 · code Apache 2.0 · corpus rows as listed in corpus.json. No watermark; no claim on outputs.

If it is useful

This was built by one person and cost about $1,900 in rented compute. It is finished and free. If it is useful to you: fork it, or send a small sponsorship toward what it cost — https://github.com/sponsors/ericrwade.

Post-publication note (2026-10-01)

Two days after release, two independent audits of this project (one by Grok, one by ChatGPT, both asked the
same question: "is it as FOSS as they claim?") found things we had stated too strongly or left out. We want
you to know what they found and what changed, because the whole point of this project is that its claims
survive inspection. Nothing that shipped was altered; the record was corrected around it.

• "Every training byte has a licence" overstated what ODC-By grants: it licenses the FineWeb-Edu dataset,
  not each web page in it. The claim now reads "every training source", with the scope stated.
• The supervised fine-tuning set includes conversations generated by two other open models (Qwen2.5-7B-Instruct
  and DeepSeek-R1-Distill-Qwen-7B), listed in the ledger all along but not called out. That is output-based
  distillation of chat behaviour, and the pre-2022 date basis applies to pretraining, not to those rows.
• Only quantised weights had been published. The shipped model and the pretrained base are now on Hugging
  Face at full precision, with the tokenizer the run used and the volume files.
• The llama.cpp MIT notice did not travel with the stick's binaries (erratum E5); it is in the repository.
• The book's story chapters were all rights reserved; the whole book is now CC BY-SA 4.0 (D-101).

Follow-up (2026-10-02). Both auditors re-ran their checks after the changes and upgraded their assessments.
What they found has been fixed to the extent it can be. What remains is recorded as a limit rather than a
fix: the pretraining volume's per-shard metadata was deleted with the rented pod; the web crawl's licence
covers the dataset, not each page; and Pagouro BE's captions came from a commercial API. A future version is
the honest path for those, never an edit of 1.0.

The full record is D-101, O-49 and D-102 in docs/DECISIONS.md, and docs/ERRATA_v1.0.md.

Mirror

Quilibrium QStorage: https://pagouro.qstorage.quilibrium.com/ — the link page, the key and every receipt for both models.

Sibling

Pagouro BE 1.0, released the same day: the Belle Époque poster image model on the same rules and the same signing key — https://github.com/ericrwade/pagouro-be/releases/tag/v1.0 · https://huggingface.co/Pagouro/pagouro-be-1.0

Errata (found after release, 2026-09-29)

Three rough edges found by the first clean-machine check, none of which changes a shipped byte:
docs/ERRATA_v1.0.md. Short form: check the signature with minisign -Vm MANIFEST.md -p pagouro.pub
(the stick's verify_manifest.py looks for minisign.pub); on macOS/Linux extract the zip with
Info-ZIP unzip or 7z x. E5: the llama.cpp binaries' MIT notice is not on the stick; it is
licenses/LICENSE-llama.cpp-MIT.txt in the repository — add it if you redistribute.

Finished

One release, frozen. No updates, no support, no telemetry, no roadmap. Fork it — the recipe is complete: docs/, BUILD_LOG.md, the book (book/).