research-director

name	research-director
description	Rank candidate experiments, decide what not to run next, maintain the factory queue, and choose one decisive next research goal before autonomous loops spend compute.

Research Director

Use when choosing the next research direction, starting a new /goal loop, reviewing a branch after surprising evidence, or asking whether the research workspace is still moving toward the right objective.

This skill does not replace research-experiment; it decides which experiment deserves to be created or run.

Read First

1. RESEARCH.md
2. programs/index.yml and the active program cards
3. OPEN_QUESTIONS.md
4. CLAIMS.md
5. DECISIONS.md
6. experiments/index.yml
7. scorecard Status: lines in factory/scorecards/
8. the latest relevant synthesis note under research-log/

Output Shape

Produce one ranked shortlist, then choose exactly one next move.

1. candidate
   expected information gain:
   cost:
   most likely way it fools us:
   what it would prove:
   what it would not prove:

Selected next move:
Not next:
Stop/pivot gate:

Factory State

When a candidate is selected, create or update its scorecard under factory/scorecards/F###_short_name.md with its Program: RP###, mark it Status: selected, and point the dashboard Goal: line at it. Keep at most one selected scorecard at a time; the scorecard Status: lines are the queue.

Ranking normally happens inside the highest-priority active program. Changing program priorities, pausing a program, or starting a new one is a portfolio move: record it in the RESEARCH.md router and programs/index.yml, and for a new program follow the Program Layer checklist in AGENTS.md (card, index row, first Q###, D### with revisit condition) before writing its first scorecard.

If the next move is ready for autonomous execution, use plan-research-goal after this skill and write the prompt into the scorecard's Goal Prompt section. Program-level LOOP_/BATCH_ prompts still live in factory/goals/.

Taste

• Prefer one decisive experiment over a broad sweep.
• Prefer tests that can falsify the current favorite idea.
• Penalize experiments that only rename a failed mechanism.
• Do not promote a result until it survives the stress test implied by its own failure mode.
• Name what should not be run next; this is part of the output, not a footnote.

Anti-Bloat Rule

If RESEARCH.md, CLAIMS.md, or DECISIONS.md have become long ledgers, choose compress before more experiments as the next move. The factory should not spend compute while the belief layer is unreadable.