We bring the generator–evaluator pattern into the collaboration framework: a person sets intent and boundaries, a generator unfolds candidates, an evaluator checks them with rules, tools and models, and human judgement and authorisation of action are brought in as the task requires.
Model judges have biases that need calibration.[2] Review should therefore combine independent evidence, human judgement and feedback from real tasks, covering three levels:
| Level | Core question | Basis for review |
|---|---|---|
| Output | Is the result correct, relevant and within constraints? | Sources, tests and inspection of deliverables |
| Trajectory | Can key assumptions, operations and tool calls be verified? | Observable records, data sources and test results |
| Alignment | Does it answer the person's intent and real situation? | Human review, authorisation and task feedback |
Human participation runs through goals, exploration, review and revision. A system should preserve people's ability to step in and correct it, and keep responsibility in design, deployment and use clearly arranged.





