Back to Projects

Scout

An operational workflow for agent evaluation

Scout finds conversations relevant to a project's work and drafts replies for human review. Human grades, feedback snapshots, and offline replay support experiments on relevance, classification, and policy.

agentsevaluationhuman review

Scout is an operational agent workflow for finding relevant conversations and drafting replies for human review. It also serves as a concrete environment for improving classifiers, evaluators, and the policies that turn scores into actions.

From conversations to review

Scout monitors Discord, Farcaster, and Bluesky, evaluates posts against project dossiers, and drafts engagement comments. Platform adapters gather messages and parent context; keyword routing filters clearly irrelevant material before a typed JIG agent evaluates relevance, drafts, critiques, and optionally revises a reply. Results are persisted and collected into a digest. Outbound content creation and publishing live in a separate application.

Human grades feed immutable feedback snapshots. Offline replay compares candidate models and prompts without changing live state. Classification scores describe features of a post; policy turns those scores into decisions such as respond, review, or drop; human judgments provide a separate basis for evaluating those decisions. A score is not itself an authorization to act.

What to inspect

The three-minute PAA evidence tour follows task declarations, evaluator evidence, autonomy state, promotion and demotion, and deployment limits. The grading and feedback guide and evaluation documentation explain the improvement path.

The related rubric study tests separate feature scores and fixed policy against blind human grades. It measures agreement with a reviewer’s labels rather than objective ground-truth accuracy.

Implementation status

Scout integrates PAA contracts and an event-sourced autonomy control plane. The checked-in inbound_reply_surfacing task is in shadow deployment; canonical_promotion is disabled. Neither task is wired to active runtime enforcement in the inspected source baseline, and no Scout task is deployed at autonomous status.

Reference evidence in the repository is fixture-derived. Production-derived grading and feedback remain in the originating deployment and are exported only in redacted, publication-safe form. Control-plane demonstrations do not establish autonomous production operation.

Fix the rubric

After testing rubric repeatability, Scout separated feature scoring from policy. JEV v5 plus fixed policy improved keep/drop agreement from 64% to 77% on a blind 100-post holdout, with ongoing production holdouts set up for further validation.