Skip to main content
A reasoning prompt passing through a Kronos audit checkpoint and repair loop
Kronos audits the reasoning prompt SERV generates from your system prompt. If the audit finds a semantic problem, SERV repairs the generated prompt and checks it again before sending the request to the model.

​
Enable Kronos

Add -serv-kronos to a supported base model ID. The suffix is a model feature, not an application tool, so you do not add a function definition or implement a handler for it.
SERV removes the feature suffix before routing to the upstream provider. Use a real base model from the model catalog, then append -serv-kronos exactly once.

​
When to use Kronos

Use Kronos for system prompts with procedural rules, multiple conditions, or consequences that must remain consistent. It is especially useful when a small ambiguity in the instructions could change the model’s decision. Kronos does not replace application authorization, tool-argument validation, or testing. Keep those checks in your application.

​
Cost and behavior

Kronos adds audit work and may add repair work to the SERV prompt-generation phase. This increases cost and latency when that work runs. Prompt caching can reuse the audited result for matching requests; the upstream model still runs for every request. See Monitor usage and billing. You can combine Kronos with Multipath by using -serv-kronos-multipath. See Use Multipath for branching instructions.