Home / Experiments Read the example, then the limits. This selection contains integration examples and official demos. Jev Tracker has not reproduced these runs.
Routing & action checks Doom Wikiracing How to judge a demo Inside an agent workflow Two places to put a decision model, explained without the setup code.
A decision inside the agent loop LangChain shows Jev choosing a model and checking a proposed tool action before the application proceeds.
Read this as two decision points, not a replacement for the whole agent. Another model still produces the answer.
Source authors: Sydney Runkle and Hunter Lovell
What it does not establish The middleware is experimental. Only named tools are checked; a blocked call is not a request for human approval.
This is an integration example, not a Jev Tracker reproduction or a guarantee of safe actions.
A useful next test Does routing preserve answer quality, and which unsafe actions escape the check? Test both before trusting the workflow.
Integration Published Sep 17, 2026 Sep 19, 2026, 04:17 UTC
View source LangChain TypeSafe integration documentation
Integration Source checked Sep 19, 2026, 04:17 UTC
View source Explanatory flows based on the documentation. Not recorded runs.
Choose the model Input Latest user message and developer-defined model criteria.
Jev decision Jev selects among the configured models.
Application action The chosen model handles the run and generates the response.
Check the action Input A proposed call to a tool selected for checking.
Jev decision Jev assesses whether the call is risky or insufficiently authorized.
Application action The middleware refuses a risky call instead of executing it.
The model’s output is a decision. The application still owns execution, permissions and recovery.
Two demos, different decisions Provider demonstrations, not independent replications.
Game control
Playing Doom without seeing the screen TypeSafe’s game demo makes decisions from text-based state, not images.
Input Structured game state expressed as text. Decision Select a game action. Then Game code carries out the action. The useful distinction is between the model choosing and the game code acting.
What it does not establish Not evidence of visual gameplay. TypeSafe notes that a conventional non-AI bot could play better.
A useful next test What happens when the state omits something important? That is a better follow-up test than judging a highlight clip.
Official Published Sep 15, 2026 Sep 19, 2026, 04:17 UTC
View source Navigation
Choosing the next Wikipedia link The official Wikiracing demo moves toward a target page through repeated link choices.
Input A destination and candidate links. Decision Choose the next link. Then The program follows the link and repeats. A bounded choice can guide a longer task without generating a new URL.
What it does not establish The post describes a two-stage process for larger choice sets. Different reasoning settings limit the model comparison.
A useful next test Can the system recover from a poor link choice? Compare completion and path length under matched conditions.
Official Published Sep 15, 2026 Sep 19, 2026, 04:17 UTC
View source
What would make the evidence stronger? Questions to ask when evaluating a demo.
Can someone else repeat it? Look for the inputs, model version, surrounding code and a clear success condition. A video alone leaves too much of the setup unseen.
Are the comparisons like for like? Match the task, reasoning settings, allowed actions and failure handling. A faster decision is not necessarily a better completed task.
What happens when it is wrong? Ask for failure cases, not just highlights. Check whether a wrong choice is reversible and whether a person can intervene.
Read the benchmark conditions Decide whether it fits your task Start with one bounded decision and a way to check whether it is right.
For a first attempt, choose something reversible and use non-sensitive inputs. Keep a person responsible for any consequential action.
Sources behind this selection A source check means we read the publication, not that we tested the model. Source type describes who published it, not whether a conclusion is proven.
Building a Harness with Jev Integration Published Sep 17, 2026 Sep 19, 2026, 04:17 UTC
View source LangChain TypeSafe integration documentation Integration Source checked Sep 19, 2026, 04:17 UTC
View source Introducing System One Models & Jev Official Published Sep 15, 2026 Sep 19, 2026, 04:17 UTC
View source Read our correction principles