Prompt and parameter suggestions

Turn reviewed evidence into changes you can inspect and accept.

Relifold uses reviewer observations and curated reasoning to propose prompt or public-parameter changes. Suggestions are advisory and remain attached to the task they improve.

One pass at a time

Finishing a curation deck starts one improvement pass. Evidence arriving while that pass runs waits for the next pass and may be combined with other newly finished decks. Viewing a report does not start background work.

Each background improvement pass has a finite hosted cost ceiling. Its cost appears with the result when the pass completes or fails. Improvement work does not extend a completed session's wall time or prevent you from authoring and running a successor.

Review before accepting

1. Open the task report and read the counterexamples behind the suggestion.
2. Compare the proposed prompt or parameter with the current binding.
3. Create an editable successor version.
4. Accept the suggestions you want and dismiss those you reject.
5. Declare and run the successor, then compare its quality and cost.

A suggestion never changes a declared version or historical session. Shared prompt advice is produced once for one prompt identity and applies to every task using that prompt. Dismissal only clears that proposal from the version's advice inbox; it does not alter data or learning state.

A prompt advisor receives the exact judge-facing values, produced and trusted decisions, and available rationales for directly reviewed mistakes. It proposes only a causally implicated prompt-rule or example change. Candidate-coverage loss, incomplete execution, missing evidence, and deterministic-gate errors remain in their owning optimization channels and cannot be disguised as prompt advice.