Skip to content

Signals, evidence, and evals

A domain can react to events as well as to goals people enter. In Gabriel’s operating model, a data feed acquires an observation, a signal evaluates whether a declared condition is met, and a playbook does the proposed work under its usual limits and approvals.

For a recruiting persona, a feed could read a hiring announcement and map it to a company observation in the ontology. A signal can flag that observation for review. A playbook can then qualify the company and draft outreach. The observation, signal reason, and accepted result let a reviewer trace why the work happened.

The feed catalog links each feed definition and source workflow. Its ontology mapping connects source fields to the domain model. Signal and schedule rules are planned. In Gabriel, feed observations, enabled schedules, approvals, and run receipts live in the workspace.

A domain specialist supplies the judgment a goal or playbook cannot capture in advance. The harness brings the evidence to the specialist and waits at the affected decision. Each approval provides evidence about what was acceptable in that run. Corrections and rejections help the team find where its source mapping, rule, or playbook needs to change.

For the two Northwind records, the expert can explain why the shared website domain was insufficient and identify the stronger source. Save the proposed match, evidence, decision, reason if supplied, and definition revision in the workspace. Use that record to propose a new eval that expects the agent to ask for review when two prospects share a domain. The team can update the mapping or playbook, run the eval, and have the specialist review the change before an authorized owner adopts a new revision.

Gabriel can capture a fully accepted goal run as a candidate procedure for the owner to review. Future runs use the adopted revision with their own inputs, grants, evidence, and effect approvals. See the proposed review and improvement rules and the team manager skill.

Evals connect a requirement to the definition that implements it and the cases that verify it. A test can require a draft and verify that the workflow has no external side effect. A production gate can require a suite to pass before release.

Outcome evidence records what happened in an actual run. Gabriel’s ROI tools can use accepted outcomes and costs to assess impact. A generated draft is evidence of preparation. A delivery receipt or an accepted record change is evidence of a different outcome.

See eval requirements and cases and the persona data feed, signals, framework, and ROI skills.

A feed reports that an account is hiring. The source date is old, and the specialist rejects the proposed outreach. Keep the source date and the reviewer’s reason with the run. The domain owner then proposes a freshness rule and tests it on the rejected example and a recent positive example.

Candidate signal → dated source evidence → specialist decision
→ reviewed freshness rule and eval cases
→ published playbook revision → new team runs

Test that old evidence routes to review or is excluded according to the new rule, and that recent evidence still qualifies. Compare the revised run’s outputs and decision points. Each new run still uses the workspace’s current grants and approvals.

See feeds, signals, and eval cases.