wfzyx/von

An open-source, non-autoregressive System One decision model. Calibrated Choice / Noul / Score inference from a 395M ModernBERT encoder, one forward pass, CPU-served.

_{Von playing Doom. Every action is one forward pass scoring six option descriptions against a text rendering of the depth buffer. Same shipped weights that answer routing questions; no policy network, no RL.}

Von answers three question types over any JSON or text state without generating tokens: Choice (pick one of K described options, with a probability over all of them), Noul (probability that a condition holds) and Score (calibrated position on an ordinal scale). It is a drop-in server and client for the TypeSafe /v1/systemone wire protocol, ships as von-sdk for Python and TypeScript, and runs on CPU (OpenVINO), CUDA, ROCm and Apple MPS.

Weights, training data, calibration and the full benchmark record live on the model card. Training on your own labels, or on another encoder (German, smaller): docs/finetune.md. This README is about using it.

pip install von-sdk          # or: uv add von-sdk
bun add von-sdk              # TypeScript / Node
docker run --rm -p 8000:8000 -v von-hf:/data/huggingface ghcr.io/wfzyx/von:cpu

Weights (~3 GB) are fetched from the Hub on first use into HF_HOME.

import von

r = von.decide(
    state="Database replication lag on cluster us-west-2 exceeded 45 seconds.",
    choices={
        "infrastructure": "Database, hardware, network, or server failures",
        "billing": "Invoices, payments, refunds, subscription queries",
        "feature_request": "Requests for new platform capabilities",
    },
    instructions="Classify the root cause domain of this incident.",
)
r.choice, r.confidence, r.probabilities   # 'infrastructure', 0.84, {...}

p = von.judge(state="Connection pool exhausted on port 5432.",
              instructions="Is this blocking customers?",
              criteria={"true": "Customer-facing requests fail or stall", "false": "Internal only"})
s = von.rate(state="Memory at 98%, OOM killer firing.", criteria=["Nominal", "Degraded", "Critical"],
             instructions="Assess degradation level.")

Several questions over one state cost one forward pass:

resp = von.system_one(
    state={"ticket": "INC-4091", "message": "Payment gateway timeouts on charge authorizations. Urgent."},
    questions={
        "intent": von.choice(instructions="Nature of the ticket?", criteria={"payment_failure": "...", "access_issue": "..."}),
        "is_urgent": von.noul(instructions="Needs immediate SLA intervention?"),
        "severity": von.score(instructions="Rate severity.", criteria=["Low", "Medium", "High", "Critical"]),
    },
)
resp.answers["intent"].choice, resp.answers["is_urgent"].noul, resp.answers["severity"].score

TypeScript mirrors the same API (decide, judge, rate, systemOne).

Two rules that save you a bad afternoon: judge without criteria is the weakest path (generic "holds / is false" descriptions, surface cues win) — pass criteria or phrase it as a described 3-way Choice. And Von is English only; other languages get token matching with confidence it has not earned.

Acting on confidence. von.patterns.confidence_gate(state, questions, threshold=0.80) splits answers into automatic / escalate. The 0.80 default is measured, not guessed (benchmarks/sweep_threshold.py, held-out jabr v2 + coding-agent probes, Von 1.2 weights): Choice at ≥0.80 keeps 25% of items at 92.4% accuracy (90% lower bound 89.4%, n=702; the rest sit at 59%); Noul keeps 16% at 83% (gated on noul_raw, the pre-band probability — the committed noul sits in a fixed band and carries no gate signal). Lowest cutoff that clears 90% kept-accuracy with 90% confidence: 0.82. In a cascade with Qwen3.5-4B on JevBench public, Von-first matches the 4B alone from 0.75 up (cost lever, not an accuracy lever). Out of domain the ranking itself breaks (judge pairs, agent-shadow probes: kept-accuracy flat at any cutoff) — a threshold cannot rescue that, only labels and a refit can. Full curves: results/threshold_sweep_*.json.

von serve exposes /v1/systemone, byte-compatible with the TypeSafe specification. Anything that talks to Jev talks to Von by changing the base URL; the SDKs do the same via VON_BASE_URL.

von serve --host 0.0.0.0 --port 8000        # auto-selects cuda / mps / openvino:gpu / openvino:cpu / cpu
curl -X POST localhost:8000/v1/systemone -H 'Content-Type: application/json' -d '{
  "model": "von-1.3.0",
  "state": {"error": "Disk volume /var/log at 98% capacity."},
  "questions": {
    "needs_action": {"type": "noul", "instructions": "Does this require operational intervention?"},
    "team": {"type": "choice", "instructions": "Who owns this?",
             "criteria": {"sre": "Infrastructure and capacity", "app": "Application code"}}
  }
}'

Response extras beyond the spec, all optional for clients: usage.input_tokens is the real tokenizer count over every encoder pass; a middle-truncated state carries a truncation field plus X-Von-Truncated / Warning headers; with --on-overflow refuse an oversize state gets HTTP 422 mentioning the context window instead (the Decision Index no-truncation rule). Criteria values may be strings or structured JSON. VON_API_KEY turns on bearer auth (clients fall back to TYPESAFE_API_KEY).

Container tags: cpu / latest (OpenVINO, linux/amd64), <version>-cpu, a UTC calver. Flags after the image name go to von serve. A CUDA image builds from the same Dockerfile with --build-arg TORCH_BACKEND=default (not published; the wheel set is several GB).

Deterministic multi-step computation for temporal/numeric states, driven by Von's own Choice decisions and zero generated tokens. It fires on computable structure in the state (two dates, a date and a duration, two amounts), never on the question's wording: a regex proposer lists typed spans; every chain whose typed slots can be filled is bound (Von picks among candidate spans when a slot is ambiguous) and executed by a fixed operator library (add_duration, in_zone, elapsed_hours, prorate, cumsum, is_leap_year, …); computed datetimes feed a bounded second round so chains compose; a computed value landing on exactly one option answers through a strict matcher, otherwise Von reads the state plus every computed fact with provenance. Chains are TOML files in src/von/chains/library/ — add your own. Latency cost lands on the hard tail (hard-tier p50 4.2 s on 4 vCPU, 0.45 s on an A10G); serve chain-heavy loads on GPU or lower VON_CHAINS_MAX_CALLS.

JevBench v1.4, the last board revision before Von's Noul band rule (v1.5.1 numbers, sealed-tier breakdowns, Decision Index and jabr v2 are on the model card). I/C/S/K = Intelligence, Calibration, Speed, Cost; composite is their geometric mean with the generalization gate applied.

¹ Remeasured under the JevBench protocol: raw p50 0.096 s on a 4-vCPU Xeon 8488C (OpenVINO), 0.023 s on an A10G; both inside the Jev-class line. results/speed_remeasure.md. ² Chains on vs off on the 111 public hard items: 2 vs 9 discordant, McNemar p = 0.065; zero discordant on easy, standard and jabr v2. Every accuracy claim in this repo goes through benchmarks/stat_gate.py (paired exact McNemar + minimum detectable effect); below the MDE it is reported as UNRESOLVABLE, not as a win. Older tables: docs/benchmarks.md.

Apache-2.0.