The loop learns, with eval gates and reversible rollbacks
Shared-source matches and role-weighted feedback turn into learned candidates. Each is gated by an eval, promoted only when it passes, and rolled back when it regresses — so learning never silently degrades the loop. You see counts, coarse weights, and closed-enum states, never a name or a comment.
- Candidates in flight
- 9
- Eval passed
- 6
- Eval failed
- 3
- Evals run
- 9
- Feedback lanes
- 5
- Rollbacks
- 1
Who weighs in, and how much it counts
Feedback is grouped by role lane, each with a coarse learning weight. The lane, its count, and its weight are all that surface, never an identity, never a comment.
- Family user64 signals · weight 1.0×
- Household admin22 signals · weight 1.5×
- Beta tester18 signals · weight 1.8×
- Staff QA11 signals · weight 2.5×
- Template reviewer7 signals · weight 3.0×
Passed and failed
A candidate is promoted only after it passes an eval. Failures are first-class and counted, not hidden.
- Eval passed6
- Eval failed3
Candidates by state
Where each learned candidate sits in its lifecycle, from candidate through promoted, demoted, or rolled back.
- Candidate9
- Under review4
- Promoted3
- Demoted1
- Rolled back1
Demonstrated rollbacks
Learning is reversible. When a promoted candidate regresses, it is rolled back, and the count of those rollbacks is shown, never hidden.
Where candidates come from
See how a single forwarded document runs through the governed pipeline and becomes the reviewable approval cards that feed this learning loop.