TypeSafe Jev 1.13

A decision model that assigns probabilities to named outcomes using supplied context. Jev complements numerical forecasters with a different kind of answer; it does not generate a sequence of future values.

Training background

TypeSafe describes Jev as a semantic decision model trained with Reinforcement Learning from Continuous Decisions. The public documentation does not disclose a complete training corpus or establish specialist training on financial or weather time series.

Available through Pretick

Available through POST /v1/outcomes/forecast and MCP forecast_outcomes. Supply a question, 2–12 outcome descriptions, and context or timestamped observations. Pretick calculates numerical summaries and validates the probability distribution. One successful call costs $0.02. Jev is experimental, separate from numerical consensus, and does not fetch live evidence. Requests are processed by TypeSafe through OpenRouter.

{
  "model": "typesafe/jev-1.13",
  "question": "At the next daily observation, will the price be below 102, between 102 and 104 inclusive, or above 104?",
  "context": "Daily price observations in USD. This is illustrative example data, not an actual market.",
  "series": {
    "data": [
      100,
      100.5,
      101,
      100.8,
      101.4,
      102,
      101.8,
      102.4,
      103
    ],
    "horizon": 1
  },
  "outcomes": [
    {
      "id": "below",
      "description": "Price below 102"
    },
    {
      "id": "within",
      "description": "Price between 102 and 104 inclusive"
    },
    {
      "id": "above",
      "description": "Price above 104"
    }
  ]
}

Use POST https://pretick.ai/v1/outcomes/forecast or MCP forecast_outcomes. Request format and limits. Each success costs $0.02.

Using this model for price forecasting

Reading the evidence

Initial retrospective test: 18 held-out days of gasoline-price observations, September 10–27, 2026. At noon America/New_York, estimate whether the last observed price before midnight falls more than one cent, stays within one cent, or rises more than one cent. Only that day’s observed prefix and 14 prior days were supplied, including a smoothed historical-frequency baseline. Jev chose the correct range on 9/18 days. Mean multiclass Brier score (sum across three outcomes; lower is better): Jev 0.691, historical-frequency baseline 0.667, unchanged-price baseline 1.333. Mean provider latency was 254 ms. This small, single-series test does not demonstrate an improvement in probability quality; calibration and broader forecasting accuracy remain unvalidated. The provider version was released within the evaluation period, so this is not a prospective or proven training-independent benchmark. Numeric precision, counting and date comparisons are documented model limitations.

Sources and version

• Official model documentation
• Choice primitive
• Jev 1.13 limitations
• Hosted provider

Hosted checkpoint: typesafe/jev-1.13
Revision: Provider-managed; tested typesafe/jev-1.13-20260917 on October 1, 2026
License/access: Hosted service; provider terms apply.

Responses identify the exact model version. These pages describe Pretick's integration; the original model may support additional features.