Zeno Divergent v0.68

A small, open model that checks a panel of forecasts before you trust it.

Give it one question, many forecasts and each forecaster's track record. It returns:

1. Forecast: what to use, and when to keep the panel's own answer.
2. Failure risk: the chance the panel is confidently wrong together.

Under 0.5M parameters per seed (471,495) · 1.6 MB ONNX · runs on any CPU · 3 seeds. Live now: web demo · API (version: "v0.68") · MCP tool zeno_check.

Model at a glance

Why it is smaller than v0.57 (9.16M): v0.68 uses the shared kernel introduced in v0.67. It is one small network for every source, built to read who disagrees and who fails together rather than to memorise each source. It has no deep panel transformer. Size was not what held either version back: v0.57's 9.16M parameters also kept the reference on most sources.

Try it online (no install)

• Web demo: https://huggingface.co/spaces/mbarbosa1/zeno-try-it (v0.68 section)
• HTTP: POST https://zenodivergent.dev/api/public/v1/check with {"version": "v0.68", "question": {...}}
• MCP: https://zenodivergent.dev/api/public/v1/mcp, tool zeno_check, argument version: "v0.68"

All three run this exact revision through the zeno068 package on CPU, so they give the same answer as a local download.

Quickstart (CPU, about 1 minute)

pip install huggingface_hub
huggingface-cli download ZenoDivergent/zeno-divergent-v068 --local-dir zeno-v068
cd zeno-v068
pip install -r requirements.txt          # numpy, onnxruntime, safetensors
python parity_check.py                   # runs every example; prints "ok" for each

import json, zeno068

z = zeno068.load(".")                                   # ONNX Runtime; zeno068.load(".", "torch") also works
out = z.predict(json.load(open("examples/crypto-request.json")))
print(out["served"])          # forecast + 90% range you can use (the trained reference)
print(out["failure_risk"])    # calibrated chance the panel is confidently wrong together
print(out["support"])         # release status, and why the model could not run (if it couldn't)

z.predict(request, experimental=True)["experimental"]   # the raw Zeno correction: NOT released, untested

What every answer contains

Input formats

1. Plain request. This is the same shape as the public API POST /api/public/v1/check. See examples/crypto-request.json. The numbers in that file are made up, to show the format.

{"source": "crypto", "decision_at": "2026-01-10T12:00:00Z", "lead_hours": 1,
 "task": {"forecast_type": "quantile", "variable": "btc-usd", "horizon": "1h"},
 "inputs": {"panel": [
   {"forecaster_id": "a", "available_at": "2026-01-10T12:00:00Z",
    "quantiles": {"0.05": 90120, "0.5": 90650, "0.95": 91180},
    "history_sequence": [{"scaled_error": 0.4, "missed": false, "label_available_at": "2026-01-10T09:00:00Z"}]}]}}

• Each member gives one of: quantiles, probability (yes/no questions), or point with error_scale.
• Forecasts available after decision_at are dropped.
• history_sequence holds the forecaster's past resolved outcomes. scaled_error is (forecast − outcome) divided by that outcome's panel spread. Outcomes known after the decision are dropped. Use zeno068.history_sequence(...) to build it.
• If task or every history_sequence is missing, the answer is a labelled pass-through.

2. Prepared panel case. This is the exact training format, used for the two real examples/forecastbench-*.json cases from the untouched window. Each one's .expected.json also records what actually happened.

Sources this package accepts: crypto, ecb_spf, forecastbench, health_covid, health_flu, health_rsv, markets, philfed_spf, philfed_spf_prob, sea. Any other source is refused, not guessed.

Using the ONNX files directly

onnx/kernel-s{11,12,13}.onnx hold the shared network. onnx/fragility-s{seed}.onnx hold the failure-risk checkpoint.

• Forecast heads: these are small linear layers stored in model-s{seed}.safetensors (heads.<source>.weight/bias, strength). zeno068/__init__.py (_forecast) shows how they are applied on top of each source's reference.
• Failure risk: this is sigmoid(joint_miss_logit) mapped through calibration-s{seed}.json (isotonic), then averaged over the three seeds.
• Match check: research/export-check.json compares ONNX against the original PyTorch training code on 200 real cases across 3 seeds. The largest difference is 3e-6.

What was measured (confirmation period, 3 seeds of the main version)

Status in one line: every source serves its trained reference. The run stopped after 4 of 12 trainings, so the matched controls the release rule needs are incomplete. The failure-risk output is labelled underpowered.

v0.57 stays the default model on zenodivergent.dev: it still improves crypto and some weather forecasts, and v0.68 does not beat it yet. Ask for v0.68 by name with version: "v0.68".

no_history (1 seed) gained only +0.35% on crypto. That suggests forecaster track records drive the crypto gain, but one seed is not a test.

Failure-risk test (untouched ForecastBench window, rescored 14 Oct with score_sealed/1.1.0)

• 2,585 held predictions on 735 questions. 577 had no eligible label and were excluded, not counted as "no failure".
• Weekly review of the riskiest 10%: 13 confident shared failures in total. Zeno caught 1–2; confidence alone caught 0.
• Verdict: underpowered. This window was also seen during an earlier rehearsal, so it is not fully prospective.

v0.57 vs v0.68

Known limits

• The failure-risk part reads a frozen copy of the shared network, so it can't shape what the network learns.
• Two matched controls (no_dependence, shuffled_dependence) never ran, and no_history ran 1 of 3 seeds. So no source can be released.
• Timing for two controls was estimated, not measured.
• Range repair needs per-case confirmation predictions, which this run didn't save. Ranges are withheld rather than widened after the fact.
• Same-answer checks between this download and the live API apply only if v0.68 is ever switched live.

Files

Licence: CC BY-SA 4.0. Training data includes ForecastBench (CC BY-SA 4.0).