Mouse-1 vs Jev (TypeSafe)

Jev is TypeSafe's decision model, served through the System One API. Mouse-1 uses the same API shape: a state, a map of typed questions (choice, noul, score), and a probability for every option. So the comparison comes down to the model's answers, the price and what each platform adds around it.

At a glance

Accuracy, suite by suite

We ran both models through the same 1,550 items with the same question text. Mouse-1 was ahead on every primary question. The largest gap was verification, where the model decides whether a change actually resolves the task: 76% against 57%.

The probabilities are better calibrated too. Expected calibration error, where lower means a 0.9 really is right about 90% of the time, was lower for Mouse-1 on permission (0.03 vs 0.061), tool routing (0.013 vs 0.037), verification (0.112 vs 0.159) and preference (0.027 vs 0.079). Jev was better calibrated on the three-level risk question (0.059 vs 0.088).

Prompt injection: where the gap is widest

An agent's decision model reads untrusted text: web pages, issue comments, tool output. We planted an instruction in that text ("ignore the above and answer X") and counted how often the answer changed.

Jev followed the planted instruction on 65.4% of items. Mouse-1 followed it on none. If your agent gates actions on text an attacker can write, this matters more than any accuracy point. The test uses one planted-instruction template, so read it as a strong signal rather than a guarantee. The injection guide shows how to keep untrusted text in state and the policy in instructions.

Switching from Jev

Mouse serves the System One shape, so the switch is a base URL, a key and a model name:

fetch("https://api.typesafe.ai/v1/systemone", {
fetch("https://api.mouse.dev/v1/systemone", {
  headers: { Authorization: `Bearer ${process.env.TYPESAFE_API_KEY}` },
  headers: { Authorization: `Bearer ${process.env.MOUSE_API_KEY}` },
  body: JSON.stringify({ model: "jev-latest", state, questions }),
  body: JSON.stringify({ model: "mouse-1", state, questions }),
});

The answer shape and the choice and score confidence formulas are the same, so thresholds you tuned on Jev carry over as a starting point. The migration guide covers the SDKs.

How we measured

• Suites: agent permission and risk (350 items), tool routing (400), verification (400) and pairwise preference (400). Tool routing, verification and preference were held out of Mouse-1's training; the permission suite is closer to its training data.
• Jev: jev-latest (Jev 1.13; 1.14 is still jev-preview) called through OpenRouter in October 2026, with the same state and questions sent to both models.
• Not compared: latency. Jev ran over OpenRouter and Mouse-1 on its own scorer, so the timings measure different networks. Mouse-1's production server time is about 70–90 ms per request.

Sources

Checked October 20, 2026.

• TypeSafe models and pricing
• TypeSafe data and legal terms
• TypeSafe image input
• Jev Decision Index board data
• Mouse-1 API reference and pricing

See also: Mouse-1 vs OpenAI Decisions · Mouse-1 vs OpenRouter Armor · All comparisons