Keep assistance useful without confusing it with authority.
A human-in-the-loop system becomes trustworthy only when it can distinguish a suggestion from a decision. Taught records both, then applies project-scoped review and promotion rules to decide what may advance.
- Independent review
- Reviewers do not inherit live vote direction while making their own judgment.
- Attributable actions
- Human decisions remain connected to reviewer, source, task, and revision.
- Visible disagreement
- Conflict is evidence to adjudicate, not noise to erase.
- Frozen gates
- Dataset membership and model promotion depend on explicit, reproducible rules.
A governed continual-learning loop
- 01Define authority
Set roles, task rules, quorum, and eligibility before review begins.
- 02Collect blind judgment
Capture independent human actions around source content and optional proposals.
- 03Register eligible evidence
Advance agreement and route conflicts into an explicit adjudication path.
- 04Challenge the incumbent
Train a candidate, evaluate it on frozen evidence, and promote only after its gates pass.
Trust boundaries the workflow preserves
- Model suggestions never count as human votes.
- Points and rewards never increase review authority or evidentiary weight.
- A candidate model does not replace the incumbent merely because training completed.
Questions teams ask before they begin
Direct answers based on the current Taught product and deployment boundaries.
What makes this human-in-the-loop rather than human-approved automation?
The system records independent human actions as their own evidence and keeps model proposals separately attributable.
How does Taught handle reviewer disagreement?
Disagreement remains visible and can be routed to project-defined adjudication instead of being silently flattened into a label.
Can rewards change whose decision counts more?
No. Project points do not change vote weight, review authority, or the evidentiary value of a decision.