Templates That Learn (Without Silently Mutating Production)
Draft, deploy-gated. The full narrative publishes with the Lab launch; the sections below are content-free by construction.
A template that improves from feedback is useful. A template that silently rewrites itself in production is dangerous. Family Manager’s templates learn, but every change goes through candidates and a reviewer gate, never a quiet in-place mutation of what is live.
Feedback creates candidates, not edits
When a reader, household member, or reviewer signals that an extraction was wrong, that signal does not edit the running template. It creates a candidate: a proposed variant that has to earn its way into production. Production stays exactly as it was until a candidate is evaluated and explicitly promoted.
This is the same propose-then-approve discipline the rest of the product uses, applied to the templates themselves.
Role-aware weighting
Not every signal carries the same weight. Feedback is attributed to a coarse role, household member, admin, beta tester, staff QA, or template reviewer, and weighted accordingly. A template reviewer’s correction counts differently from a single end-user thumbs-down. Roles are a closed set of labels; the underlying identities and any free-text notes never leave the platform.
One signal, end to end
The story the Lab tells follows a single signal through the loop:
- Feedback, a coarse signal arrives, attributed to a role.
- Candidate, a variant is created; production is untouched.
- Eval, the candidate is measured against a baseline on an objective metric.
- Reviewer gate, a human decides promote or discard.
- Promotion or demotion, only on promotion does production change, and a rollback path exists if a promoted variant later regresses.
At no point is raw private feedback text exposed. The Lab shows counts, roles, coarse evaluation bands, and promotion state, never the words a family typed.
Why the gate matters
Self-modifying systems fail in the worst way: quietly, in production, with no audit trail. Putting a reviewer and an explicit promotion step between feedback and production keeps the learning loop honest. Improvement is real, but it is always a decision someone made, not a drift no one noticed.