Best of N and refinement

Ask CLIO for several alternatives when a task benefits from different approaches: a summary, a report introduction, an explanation, or a plan. Best of N makes independent attempts at the same task. Each candidate has its own tab, so you can compare complete answers without mixing their text together.

Development preview

This guide describes the reviewed development build. Its corrections for five individual candidates, separate reasoning, and evaluation criteria are not included in beta 5.2.

Ask for alternatives

1. State the task, the number of drafts, and what makes a good answer.

   Use Best of N to write five different introductions to this report.
   Each should be one paragraph for a general scientific audience.
   Compare clarity, fidelity to the supplied results, and concision.
   Let me choose the one to continue with.

2. Open the Alternative drafts block. Select each Try tab to read its candidate. Reasoning and tool activity can be expanded separately.

3. Open Evaluation criteria to check the instructions used for the comparison. When you are choosing, select Pick Try on the candidate you want.

The agent decides when to use the draft tool; asking for numbered alternatives in ordinary prose does not by itself establish a Best of N run. Check for the draft block and its individual tries. The default maximum is eight attempts; an operator can configure a lower limit.

Choose who judges

You choose

Include “let me choose” in the request. The run waits for your selection, and the conversation continues from the selected candidate. Other candidates remain available in the recorded run.

For example, five user-judged attempts in a real Codex/Luna check each produced one short poem. The five tabs represented five separate candidates, rather than five sets of poems.

The model chooses

Ask CLIO to select against specific criteria, such as “accurate to the source, plain language, under 150 words.” Each evaluated candidate receives a score from 0 to 1. Selected identifies the chosen candidate.

In a separate real Codex/Luna check, three short poems received scores of 0.85, 0.95, and 0.92 against the same criteria. The second candidate was selected. These are one model’s judgments for that example, not measured correctness or a promise that the selected answer is better for every reader.

Parallel drafts or sequential refinement?

Approach	How attempts relate	When to use it
Best of N	Independent candidates run in parallel, then a person or model selects.	Explore different approaches to one task.
Refine	Each attempt uses feedback from the previous one and runs sequentially.	Improve a candidate against a concrete criticism.

In a user-judged refinement run, add an optional comment to the candidate. Refine Try requests another attempt using that feedback; picking without a comment accepts the candidate. For example: “Keep the opening, remove the repetition, and define the acronym.”

Several attempts use more model work than one answer. Check their activity and the session’s usage when deciding how many to request. For factual work, verify the evidence behind the selected answer as well as its wording.

Keep the result useful

Tell CLIO what to do with the chosen result: revise a document, develop the selected plan, or use the selected explanation in a dashboard. Selection chooses an answer; it does not independently approve a protected file change or publish a result outside the workspace.

DashboardsDevelop the chosen analysis into a connected interactive report.

Memory and evidenceInspect the recorded work and sources behind an answer.