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 wire-compatible with the TypeSafe /v1/systemone specification, ships as von-sdk for Python and TypeScript, and runs on CPU (OpenVINO), CUDA, ROCm and Apple MPS.

Results

JevBench v1.4 (the public System One benchmark; scores from results/v1.4/jevbench-v1.4-results.json in the jevbench repo). I/C/S/K are the Intelligence, Calibration, Speed and Cost axes; the composite is their equal-weight geometric mean with JevBench's generalization gate applied. Public-tier accuracies are chance-uncorrected fractions. Latency is JevBench's adjusted p50 (self-hosted rows: raw ×2 + 0.15 s). Cost is JevBench's tariff estimate per 1,000 decisions.

system	params	I	C	S	K	composite	easy	standard	hard	sealed	p50 latency	$/1k	endpoint
Jev 1.13 (TypeSafe, closed)	—	53.1	76.3	83.3	52.0	63.3	1.000	0.990	0.741	0.367	0.65 s	0.040	api
hopper (Qwen3.5-4B + LoRA)	4B	48.0	79.1	86.8	58.7	59.4	1.000	0.969	0.650	0.341	0.41 s	0.024	gpu
Qwen3.5-4B frozen (semif)	4B	44.4	66.8	83.7	59.5	47.7	1.000	0.979	0.595	0.263	0.55 s	0.022	gpu
jeff (GLiFormer-large)	~400M	36.8	67.9	63.5	76.6	30.6	1.000	0.760	0.377	0.331	2.03 s	0.006	cpu
Laya (ModernBERT-large + option marker)	421M	36.1	63.7	71.1	86.2	30.3	0.944	0.729	0.341	0.308	1.72 s	0.003	cpu
Von 1.2 (board row)	395M	34.5	75.7	70.5¹	77.8	27.5	0.931	0.688	0.373	0.279	0.34 s¹	0.006	cpu
Von 1.3 (this repo, same weights + chains)²	395M	—	—	88.4¹	—	—	1.000	0.556	0.441	—	0.36 s¹	0.005³	cpu
GLiNER2-large	~300M	31.1	24.8	61.7	73.3	15.1	0.986	0.625	0.364	0.286	2.34 s	0.008	cpu

¹ The board's Speed 70.5 for Von is a v1.3 carry-over with no recorded hardware. Remeasured 2026-09-27/28 under the JevBench protocol: Von 1.2 raw p50 0.096 s on a 4-vCPU Xeon 8488C (OpenVINO), adjusted 0.34 s, Speed 89.0; Von 1.3 with chains on 0.104 s / 0.36 s / 88.4; 0.023 s raw on an A10G, Speed 94.2 (both versions). All inside the Jev-class latency line (≤ 1.30 s adjusted). Details and the submission payload: results/speed_remeasure.md.

² Von 1.3 is an inference-engine release over the unchanged von-1.2 weights: chain-of-options on by default. Public-tier accuracies are the local 231-item run (benchmarks/data/gate_chains_oos.json); the board's Von 1.2 public numbers were measured by JevBench on the same items (0.931/0.688/0.373) and Von's own local run of 1.2 gives 1.000/0.556/0.378, so compare hard 0.378 → 0.441 (+6.3 pp, 2 vs 9 discordant, McNemar p = 0.065, MDE 12.9 pp) within the local protocol. Zero discordant pairs on easy, standard and jabr v2 (989 items). Sealed and I/C/K wait for the board's own run.

³ Real tokenizer count reported in usage.input_tokens (was chars/4): 63 tokens/decision on easy+standard, 1,103 on hard; blended $0.0049/1k at the encoder tariff, Cost 79.3.

Older, longer tables (jabr v2 49-task suite, ViZDoom, training-data coverage) live in docs/benchmarks.md.

Von vs Laya

Same trunk class (ModernBERT-large encoder + option-marker head), same endpoint kind. Every number is from the JevBench v1.4 results file unless marked.

metric	Von 1.2	Laya	source
parameters	395M	421M	model cards
Intelligence	34.5	36.1	v1.4 axes
Calibration	75.7	63.7	v1.4 axes
Speed (board)	70.5 (unmeasured carry-over)	71.1	v1.4 axes
Speed (remeasured, CPU)	89.0	—	results/speed_remeasure.md
Cost	77.8	86.2	v1.4 axes
Composite	27.5	30.3	v1.4
public easy / standard / judge / hard	0.931 / 0.688 / 0.685 / 0.373	0.944 / 0.729 / 0.692 / 0.341	v1.4 tiers
sealed accuracy (308 items)	0.279	0.308	v1.4 sealed
sealed ECE	0.107	0.172	v1.4 sealed
adjusted p50 latency	0.34 s (remeasured)	1.72 s	v1.4 / remeasure
Jev-class latency line (≤ 1.30 s)	inside (remeasured)	outside	v1.4 rule
$ per 1,000 decisions (estimate)	0.0055	0.0029	v1.4 cost
order-invariant option scoring	yes (independent-options masking)	not stated	this repo

Where Laya leads: standard tier (+4.1 pp), Cost, sealed accuracy. Where Von leads: Calibration (+12.0), hard tier (+3.2 pp), sealed ECE, and, once the remeasurement is on the board, Speed and Jev-class eligibility.

Quickstart

pip install von-sdk          # or: uv add von-sdk
bun add von-sdk              # TypeScript / Node

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?")
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) and can also talk to any /v1/systemone server via VON_BASE_URL.

Server

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-1.2.0",
  "state": {"error": "Disk volume /var/log at 98% capacity."},
  "questions": {"needs_action": {"type": "noul", "instructions": "Does this require operational intervention?"}}
}'

Container

A CPU and a CUDA image are in PR #12; once it lands: docker run -p 8000:8000 ghcr.io/wfzyx/von:cpu. Until then build from the Dockerfile on that branch.

Configuration

flag / env	default	effect
--device / VON_DEVICE	auto	cuda, mps, openvino:gpu, openvino:cpu, cpu. Auto prefers OpenVINO CPU over plain torch CPU.
--max-state-tokens N / VON_MAX_STATE_TOKENS	8192	States longer than N tokens are middle-truncated (60 % head, 40 % tail) so question and options always fit the 8192 window. Truncated responses carry a truncation field and X-Von-Truncated / Warning headers.
--on-overflow truncate|refuse / VON_ON_OVERFLOW	truncate	refuse answers an oversize state with HTTP 422 ("exceeds the … context window") instead of truncating. Required under no-truncation rules such as the Decision Index.
--chains DIR / VON_CHAINS_DIR	bundled library	Chain-of-options library (below). --no-chains / VON_CHAINS_DIR=off disables it.
VON_CHAINS_MAX_CALLS	16	Encoder sub-decisions a chained item may spend.
VON_CHAINS_MAX_STATE_TOKENS	4096	Chains stand down on longer states (each sub-decision re-encodes the state).
VON_API_KEY	unset	Bearer token required by the server when set (client falls back to TYPESAFE_API_KEY).
VON_MODEL_ID	wfzyx/von	Hugging Face repo or local checkpoint directory.

Chain-of-options

Deterministic multi-step computation for temporal/numeric items, driven by Von's own Choice decisions and zero generated tokens. Nothing in it reads the question's wording; it fires on computable structure in the state (two dates, a date and a duration, or two amounts).

1. a regex proposer lists typed spans (dates, durations, time zones, amounts, percents, tables) — it never decides;
2. 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, …);
3. computed datetimes become spans for the next round, so chains compose (warranty end → days to claim) without composed definitions; bounded by rounds, facts and a sub-decision budget;
4. a computed value that lands on exactly one option answers through a strict matcher; several grounded candidates go to Von as a Choice between the computed facts; otherwise Von reads the original state plus every computed fact, with provenance.

Chains are TOML files in src/von/chains/library/ (deadline+timezone, month-window, proration, cumulative-vs-limit, elapsed-window, days-between, weekday, leap-year). Degenerate bindings (same instant twice, zero duration) are rejected before execution; a lone date fires nothing. Gates (benchmarks/gate_standard.py --cand-chains, paired McNemar): public hard 37.8 → 44.1 % (2 vs 9 discordant, p = 0.065), zero discordant pairs on easy, standard and jabr v2. 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.

Training and calibration

Von 1.2 is ModernBERT-large with an option-marker head, trained with a listwise softmax cross-entropy + Brier objective on ~63k operational decision items plus synthetic two-hop and numeric sets, with independent-options attention masking so option order cannot change the answer. Calibration is an input-conditioned temperature map fitted post hoc (checkpoints/von-1.2/marker_calibration.json). Every accuracy claim in this repo goes through benchmarks/stat_gate.py (paired exact McNemar + minimum detectable effect); results below the MDE are reported as UNRESOLVABLE, never as wins.

License

Apache-2.0.