Proposal policies — evidence into proposals

Module: agentdescent.policies · Contract: ProposalPolicy.propose(ctx: ProposalContext) -> Sequence[str]

The seam between a rollout's outcome and the proposal the strategy turns into a diff. The engine consumes one proposal per rollout; a policy returning several is refused rather than truncated — batched rollouts are what makes k > 1 usable, and silently keeping the first would report work that never ran.

Implemented

No shipped implementations yet — the field takes any object with the protocol's propose method. The MethodPolicy ports do not use this seam: their proposal logic is plain actor code (a propose callable), which is the right home when there is no decision to swap, only text.

Write one when the rule for proposing needs to vary independently of the prompts — e.g. propose only below a reward threshold, or route to different reflectors by task cluster.