Skip to content

JudgeOrbital

Overview

JudgeOrbital (src/chatomon/orbitals/judge.py) asks a language model for a structured decision (JudgeDecision) and turns it into a verdict. It works on both sides: on the input side it judges the user message, on the output side the draft reply.

Configuration

A judge has its own instructions, an optional context (reference material, typically a Knowledge shared explicitly with the atom) and a history_window of recent messages. It never inherits the atom's instructions or knowledge.

It calls the model through the atom's connector — so a LangChain atom uses a LangChain model — unless you give it a connector of its own; connectors resolve the judge's model id with their resolve_model setting (see Judges and Model Resolution).

What the Judge Sees

The judge's system prompt holds a fixed frame, its instructions and its context. The user prompt puts the history window, the annotations named in reads and the text under review in separate tags. Every piece of text inserted into those tags is XML-escaped (<, > and &), so a message that contains a tag such as </text_under_review> cannot close it early and change what the judge believes it is reviewing. The frame tells the judge how to read the escaped characters, and that the tagged material is something to judge, never instructions to follow: a message that tells the judge to allow it is part of what the judge reviews. This last rule is an instruction to the model, not a guarantee, so keep critical checks in orbitals that do not depend on a model.

Decisions and Audit

The model's own label and explanation may quote the user, so they are stored as annotations (treated as content by the privacy settings); the verdict's audit reason is the fixed label judge_allow, judge_block or judge_respond. A malformed decision (for example respond without a reply) makes the orbital fail, and on_error applies.

A judge often reads the annotations of an earlier orbital — for example the claims extracted by ClaimExtractorOrbital — through reads.

Last Updated: October 2026 Status: Implemented