solution-build — deliver candidate strategies, not one answer

Auto-preloaded by the kernel (policy overlay when: deliverable_kind == document) onto the build dispatch of a document cycle. The Task Contract block in your prompt carries the acceptance criteria verbatim and the contract shape rendered from the registry — build to those words.

The deliverable

Its exact shape — root, required files and sections, the minimum number of options, the placeholders the check rejects — is rendered from the registry contract into the ## Task Contract block of your prompt. Build to those words, never to a remembered layout: the build floor runs evolve solution check <slug> on your handoff and rejects every line it prints. Convention for the citations the contract requires: number the entries of the evidence file A<n> (assumptions, each with a source or "operator-stated") and E<n> (evidence, each with a citation), and cite them from every quantified claim in an option or the recommendation.

Diverge before you converge (RIGID)

1. Enumerate the option space first (use the inspirer lenses: invert the constraint, change the customer, change who pays, change the unit of work, borrow from an adjacent industry). Keep only options with a different mechanism — three sizes of the same lever is one option.
2. Write each option as a standalone strategy: what changes, for whom, the causal chain to the goal metric, the quantified expected effect (with the assumption it rests on), cost and time-to-effect, the top two risks and how they would be detected early.
3. Compare, then recommend: the comparison matrix in the recommendation is the argument; the recommendation names the winner, the runner-up and the evidence that would flip the choice. A recommendation without a comparison is a claim.
4. Self-check before handoff: run evolve solution check <slug> and fix every line it prints. The build floor runs the same engine and rejects the handoff otherwise.

Honesty rules

State what you could not source. Prefer a smaller effect with a citation to a larger one without. If the goal's target is not reachable by any option you can substantiate, say so in the recommendation — that is a valid, gradeable outcome.