Skip to content

Semantic rules: compare presentations of related nodes #170

Open

Description

@mindsocket

Summary

A semantic rule with `semantic.scope` sends related nodes to the model provider as a flat list, each with its relation to the evaluated node (`parent`, `child`, `ancestor`, ...). We don't know whether the presentation affects verdicts. A layout that shows the structure (nested text, a graph, a diagram) might judge coherence across the tree better, or make no difference.

Experiment

  • One scoped rule, one fixed node set, two presentations: the current flat list with relations, and one structural layout.
  • Run each several times per provider and compare verdict distributions and, where the provider returns them, probabilities.
  • The batching drift test in `tests/judge/live.test.ts` is a starting point for the harness.

Outcome

Keep the flat list unless a structural layout changes verdicts in a consistent direction. Record the result either way.

Relates

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions