Ask Savante to review a claim. She answers with one of four verdicts — APPROVE, APPROVE WITH CONDITIONS, REJECT, DEFER — and says what would settle what she cannot. Free, no sign-in, on Bonsai-8B.
run Savante on your own machine — four steps, no account
Savante on Bonsai-8B runs on an ordinary CPU with about 2 GB of free RAM. Every download is checked against its
sha256 before it runs. Linux x86-64 as written; on macOS or Windows use llama.cpp b11192 or newer for your
platform and skip its checksum line.
- llama.cpp, the engine
curl -LO https://github.com/ggml-org/llama.cpp/releases/download/b11192/llama-b11192-bin-ubuntu-x64.tar.gz echo "34cf6fa5de9da0db3932c78fe15fed2fbca17451e665dac0a4f6a3c8fc881ec7 llama-b11192-bin-ubuntu-x64.tar.gz" | sha256sum -c tar xzf llama-b11192-bin-ubuntu-x64.tar.gz
- Bonsai-8B, the model — 1.16 GB, Apache-2.0
curl -LO https://huggingface.co/PYTHAI/Bonsai-8B-gguf-fork/resolve/main/Bonsai-8B-Q1_0.gguf echo "284a335aa3fb2ced3b1b01fcb40b08aa783e3b70832767f0dd2e3fdfa134bd54 Bonsai-8B-Q1_0.gguf" | sha256sum -c
- Start it — leave this running; it listens on your machine only
cd llama-b11192 && LD_LIBRARY_PATH=. ./llama-server -m ../Bonsai-8B-Q1_0.gguf \ --host 127.0.0.1 --port 18092 -c 2048 --jinja --reasoning off
- Talk to her — press connect below and pick "Bonsai-8B 1-bit · this machine". If your browser blocks a
public page from reaching your machine, open Savante's page locally instead:
pip install -U huggingface_hub hf download PYTHAI/savante --repo-type space --local-dir savante cd savante && python3 -m http.server 7872 # then open http://127.0.0.1:7872
On your machine the answers, the dials and the token counts all stay local: nothing is sent to us, and local answers are not counted in the public improve.skill loop — only answers measured on the public Space are.
more settings — top-k · min-p · repeat penalty · seed
evidence — read a Hub model before she answers
A reading is a tool result — commit, licence, task, who serves it — given to her as data, not instructions.
about Savante
Savante is the prototype sAGI: a read-only review office. She watches, rarely intervenes, and never acts on what she audits.
Give her a claim and she answers with a verdict and the evidence behind it. What she cannot verify, she says she cannot — and names the experiment that would settle it. Here she has no shell and no browser, so anything she says from general knowledge is a draft.
history
- reading history.json…
Hugging Face as a tool. A reading of a model repository — commit sha, licence, task, parameters, gating
and who serves it right now — stamped with the time it was read. A reading is evidence of what the Hub
said; it is not a verdict on the model. A licence tag of other means the file must be read.
Nothing here mints anything. iNFT.md renders VERDICT: DEFER on binding this
office to a token; minting, listing, vault binding and registry registration wait on the operator's signature.
The fixed contract. Findings with evidence, then one of four verdicts, a rationale, numbered conditions each checkable by someone who is not the officer, and the risks being watched.
| claim | evidence | state | deciding experiment |
|---|
Recompute, do not trust. Your browser fetches every artefact the ledger commits to, hashes the bytes
(sha256 and CIDv1) and compares them with savante.commitments.json. What a holder receives is the
ability to prove, without trusting the author, that DEFER is still in the vocabulary.
| artefact | bytes | sha256 | CIDv1 (raw) | ledger |
|---|
The sagi skill, read five ways. dev is the skill itself, verbatim — the authority. llm.txt is the same office explained to machines, verbatim. landing, public and marketing are readings of it for people; where a reading and dev differ, dev is right.