Von 1.1

Von is an open-source, non-autoregressive System One decision model. It answers structured decisions — pick an option, judge a condition, rate a level — in a single bidirectional forward pass, with no token generation and no chain of thought.

• Version: 1.1 (von-1.1.0)
• Size: 395M (ModernBERT-large backbone + Option-Marker scoring head)
• Context: 8192 tokens
• Repo: https://github.com/wfzyx/von
• License: Apache-2.0

Previously published as wfzyx/von-1.0. The repo was renamed once model naming moved to version numbers; the old id still redirects, so existing installs keep working.

How it works

The premise and every candidate option are packed into one sequence. Each option gets a [MASK] marker that attends to the state and to the other options simultaneously, so options are scored jointly rather than one at a time. One forward pass yields a full probability distribution over the options.

This is why Von is fast: cost is a single encoder pass regardless of option count, instead of one generation per candidate.

Usage

pip install "von-sdk>=1.1.0"

import von

answer = von.decide(
    state="Order #123 was never delivered and the customer is asking for their money back.",
    choices={
        "refund": "Issue a refund",
        "track_order": "Help track the package",
        "escalate": "Escalate to a human agent",
    },
)
print(answer.choice, answer.confidence)

Serve the native TypeSafe-compatible /v1/systemone endpoint:

von serve --model von-1.1 --port 8000

A TypeScript/JavaScript SDK is also available: npm install von-sdk.

Calibration

Von returns calibrated probabilities. Confidence is produced by an input-conditioned temperature map — a bounded linear function of the option-distribution entropy, the state length, and the option count — stored in marker_calibration.json as calibration_map.

A single global temperature cannot serve both ends of a difficulty range: easy items should stay sharp, hard items must admit they are near chance. Temperature scaling is monotonic, so this never changes an answer — only how confident Von claims to be.

Measured on JevBench's public items using its own ece_top_label metric:

Tier	Accuracy	Mean confidence	ECE
easy	93.8%	88.1%	0.060
standard	66.7%	79.5%	0.133
hard	36.9%	46.0%	0.090

The map was fitted on the 231 public JevBench items, so treat these as in-sample. Split-half validation (fit on half, score the untouched half) puts the JevBench Calibration axis in the 68–82 range. No JevBench item was used for gradient training.

Evaluation

JevBench public splits (official harness):

Split	n	Accuracy
easy	48	0.938
original (standard)	72	0.653
hard (public half)	111	0.351

Von is strongest on short, well-posed decisions and weakest on long multi-clause policy documents requiring multi-hop composition. A marker-distance probe confirms the model reads its premise accurately out to ~2048 tokens, so the hard-tier gap is compositional depth rather than a retrieval or context-length limitation.

Files

File	Purpose
model.safetensors	ModernBERT-large backbone
option_marker.pt	Option-Marker scoring head
marker_calibration.json	Fitted calibration map + temperature
config.json, tokenizer*	Standard HF config and tokenizer

Limitations

• English only.
• Not a generative model: it selects and scores, it does not write text.
• Near-chance on long multi-hop legal/policy reasoning; do not use it unsupervised for high-stakes contract adjudication.
• Calibration was fitted on public benchmark data and may drift on very different domains. Refit with benchmarks/fit_calibration.py if you depend on the confidence values.

Citation

@software{von2026,
  title  = {Von: An Open-Source System One Decision Model},
  author = {Panisa, Victor Hugo},
  year   = {2026},
  url    = {https://github.com/wfzyx/von}
}