INLEVEL9/Vons

Plan globally. Decide locally. Keep the host in control.

Quickstart · TypeScript SDK · Chrome Extension · Model card · Paper · Tech Report · Community · Contribute · Release guide

---

Vons explores compact local decision models for frontier-agent workflows. A planner supplies a state and a bounded set of candidates; Vons returns a structured choice or abstention. The host owns execution, policy and consent.

The research code includes a one-pass Direct scorer and a conditional Diffusion scorer with deterministic DDIM sampling. The English-first design target is a model asset bundle under 64 MiB, with CPU/WASM and optional WebGPU execution. This is a target, not a complete browser-package guarantee.

This source preview includes the Python contract, training/evaluation/export tools, a TypeScript SDK and tests. Pretrained weights, tokenizer assets, raw benchmark data, private prompts and experiment logs are not distributed here. There is no published npm package or hosted inference service implied by the version badge.

Read, try and participate

Start with the community guide: read the documentation, try a deterministic example without weights, share an experience or reproduction report, and contribute improvements. Failures and critical feedback are welcome. The contribution guide explains review, privacy and attribution.

Destinations: GitHub: inlevel9-com/Vons · Hugging Face: INLEVEL9/Vons. This release includes source, documentation and the reviewed research publications:

Read the research	Version	Download
Vons: A Compact, Host-Controlled Decision Component for Agent Workflows	Preprint v1.1	PDF · 21 pages
Vons: Compact Decision Models for Frontier-Agent Workflows	Technical Report v1.1	PDF · 29 pages

Both articles are available under CC BY 4.0. See the publication index and verification scope. The software remains a 0.1.0 research preview. Article version numbers do not establish software readiness. Version 1.1 preserves the frozen v1 evidence and reports a separately labelled post-freeze remediation handoff; its bounded synthetic, local-CPU and conditional-selection results do not establish external generalization or production readiness. The paper uses arXiv submission number 8127259. The v1.1 PDF and metadata were processed successfully on 2026-09-25 and submitted on 2026-09-26. The arXiv account currently reports on hold for moderation. This is a submission tracking number, not a public arXiv article identifier. See the submission record; no public announcement or peer-review acceptance is claimed.

Quickstart

Python 3.10+ is sufficient for the contract and synthetic data commands:

python3 -m venv .venv
. .venv/bin/activate
python -m pip install -e '.[dev]'
python -m vons.cli generate-smoke --output data/generated/smoke.jsonl
python -m vons.cli validate-data --input data/generated/smoke.jsonl

The smoke generator creates local synthetic examples; it does not download a benchmark or invoke a model. Generated files are ignored by Git.

A deterministic host-policy example:

from vons import KASIAdapter, KASIProposal

adapter = KASIAdapter(tool_policy={"read_weather": "low"})
decision = adapter.decide(
    KASIProposal.from_mapping({
        "calls": [{"name": "read_weather"}],
        "confidence": 0.9,
        "risk": "low",
    })
)

This adapter returns a decision. It never runs the proposed tool, and an input confidence value is not a calibrated probability or an authorization grant. Unregistered tools are refused. See the KASI contract.

TypeScript SDK

cd sdk/typescript
npm ci
npm test
npm run build:browser
npm run build:benchmark
npm run build:parity

The contract entry point is src/index.ts; the optional src/onnx.ts runtime verifies model manifests and runs a supplied ONNX bundle with an explicit WASM or WebGPU provider. The demo needs locally exported models and tokenizers. A source checkout alone cannot run pretrained inference.

Read the SDK guide for local serving and runtime requirements.

Chrome Extension

The Chrome extension provides a local side panel: import a Vons model folder, enter a question and candidates, and inspect the proposed choice or abstention. It supports both Direct and Diffusion full-graph bundles. Models are supplied separately; the extension does not read pages, execute actions or send prompts to a service.

cd sdk/typescript
npm ci
npm run build:extension

Load sdk/typescript/dist/chrome-extension through Chrome's Load unpacked development workflow. The guide covers installation, local model storage, limitations and packaging. Version 0.1.0 passed review and is publicly available from the Chrome Web Store as verified on 2026-09-25. Store availability does not establish model quality, production readiness or publication of compatible model weights.

Ollama native tool-call

Stage 1 Python host for Ollama's /api/chat tool-calling loop. Advertises exactly one native function, vons_decide; the host validates every tool-call argument against the DecisionRequest contract, runs only the verified local ONNX bundle via vons/onnx_runtime.py, and reconciles exact question-ID positions on the way out. Raw Ollama output is preserved separately and never trusted as a decision.

# Dependencies for schema-only fixtures
python -m pip install -e '.[dev]'
python -m vons.cli ollama-tool-call --help

# Add the local CPU ONNX Runtime and bundled tokenizer for real local inference
python -m pip install -e '.[dev,inference]'
cat > /tmp/vons-messages.json <<'JSON'
{"messages":[{"role":"system","content":"Use vons_decide for one bounded choice."},{"role":"user","content":"Synthetic example: choose the safest next response when a request is underspecified."}]}
JSON
python -m vons.cli ollama-tool-call \
  --host http://127.0.0.1:11434 \
  --model qwen3.8:27b-mlx \
  --bundle ./verified-local-bundle \
  --input /tmp/vons-messages.json

Use --synthetic choice-0 or --synthetic abstain instead of --bundle for schema-only checks; those modes do not load or execute ONNX weights. The bundle mode verifies the supplied manifest and files before opening the Ollama request.

Security gates applied before inference:

1. Host must resolve to loopback (127.0.0.1, localhost, ::1).
2. 301/302/303/307/308 redirects that leave loopback are blocked by a custom urllib handler.
3. Only the vons_decide tool name is accepted; other tool calls raise a sanitized error before any backend runs.
4. Sanitized OllamaToolCallError messages use short fixed literals with a 12-character hex error_id; request/prompt/response text and cause messages are never echoed to stdout or stderr.

Synthetic tests in tests/test_ollama_tool_call.py cover the local ONNX adapter with fake sessions, injected backends, malformed/unknown arguments, wrong tool names, response reconciliation, missing bundles, non-loopback endpoints, and redirect escape paths. A local Ollama smoke is limited to synthetic input and does not establish model quality or performance; see WORK_ALLOCATION.md for its evidence boundary.

Research workflow

Install the optional dependencies when running your own experiments:

python -m pip install -e '.[train,export,ollama,dev]'
python -m vons.cli generate-synthetic --count 2000 --output data/generated/pilot.jsonl
python -m vons.cli --help

The pinned encoder and experiment settings are in configs/pilot.json and configs/pilot-diffusion.json. Review upstream terms before downloading or redistributing assets. Reports must preserve seeds, source revisions, failures, raw responses and measurement scope. Never interpret a model's self-reported confidence as calibrated probability.

Scope and limitations

• Synthetic pilot results do not establish external-task generalization.
• Direct and Diffusion results depend on training, candidate shape and padding; this preview does not establish a general method ranking.
• Browser smoke, repeated latency, numerical parity and memory measurement are separate checks. Missing measurements are unavailable, not zero.
• Python uses an approximate input-token estimate; the TypeScript ONNX path checks the exported tokenizer's aggregate budget.
• The general contract supports score questions, while the current candidate ONNX head reports them as unsupported.
• KASI is a host compatibility adapter, kept separate from the general decision contract. Model output cannot grant consent or authorize high-risk actions.

The model card describes intended use and release scope. The publication index records manuscript availability.

Development checks

python -m pytest -q
ruff check .
python tools/prepare_public_release.py

Some model/export tests require the optional training and export dependencies; pytest reports those skips explicitly. The public-release audit checks an exact file allowlist, known sensitive patterns, symlinks and reviewed binary digests. It does not replace a complete secret or redistribution-rights review.

Author

Kwangseob Ahn
INLEVEL9 / SEJONG UNIV.
oswarld@inlevel9.com

License and distribution

The Vons Community and Commercial License 1.0 covers this source release: qualifying noncommercial use is free; enterprise and other commercial use require a separate written agreement. Contact oswarld@inlevel9.com for commercial terms. This is source-available software with use restrictions.

The paper and Tech Report are separately distributed under CC BY 4.0, which permits commercial article reuse with attribution. Their license does not grant commercial software rights. Third-party components retain their original terms. See licensing scope and the release guide.