How it works
Your skills describe how you want code written. Bearing turns preferences that need judgment and context into checks that run at the end of an agent turn—alongside your traditional linters.
After each agent turn, Bearing sends the code diff and your rules to Jev. Jev evaluates each rule and returns a structured answer with probabilities. Bearing uses your chosen threshold to decide whether to flag a finding, request review, or let the agent carry on. Findings include the guidance you wrote in the rule.
Define what good looks like
Ask your agent to propose rules from your skills and project conventions. Review the question, answer criteria, and threshold for each rule, along with the message the agent should receive when it fires. Rules live in .bearing/rules/.
Check the change
Agent hooks capture changes made during the turn. Bearing selects the rules that match the changed files and sends those rules and diffs to Jev. Rules can also request the surrounding file when they need more context.
Feed the result back
When a result meets your rule’s threshold, Bearing reports the finding with the guidance you wrote. Uncertain results are flagged for review. The agent can use that feedback to correct its work; Bearing does not rewrite the code itself.
You can also run bearing check yourself to check the whole working change. Work that could not be checked is reported separately, so it is never mistaken for a clean result.
More about Jev
Jev is a model from TypeSafe that evaluates questions against the context you provide. In Bearing, that context is your code change, and the questions come from your rules.
For structured decisions, Jev can be much faster and cheaper than a general-purpose LLM. TypeSafe reports around 194× faster responses and 445× lower cost than the LLMs compared in its workflow benchmarks. That speed and cost make frequent checks practical inside an agent’s workflow.
Jev returns structured answers and probabilities. Bearing uses those answers to decide whether to flag a finding, request review, or let the agent carry on. You choose the criteria, threshold, and feedback.