Idea Discovery

Discover, compare, and select research candidates for: $ARGUMENTS

This is the candidate-discovery layer. It must end with a selected, evidence-bounded candidate or a clear request for more direction. It must not silently generate a long method proposal, a full experiment roadmap, or a paper-ready research plan for every candidate.

Capability Routing

This is a first-layer entry skill. Resolve the debuffer repo root from .debuffer_skills/installed-skills-codex.txt (repo_root) when available, read the referenced library SKILL.md, then follow that skill. Do not duplicate library instructions here.

• /idea-creator: read ../library/idea-method/idea-creator/SKILL.md.
• /reference-paper-deconstruction: read ../library/idea-method/reference-paper-deconstruction/SKILL.md.
• /method-paper-gate: read ../library/idea-method/method-paper-gate/SKILL.md.
• /research-lit: read ../library/literature/research-lit/SKILL.md.
• /novelty-check: read ../library/review/novelty-check/SKILL.md.
• /research-refine: read ../library/idea-method/research-refine/SKILL.md only after a candidate is selected.
• /research-blueprint: use the entry skill ../research-blueprint/SKILL.md only after Idea Freeze.

Layer Boundary

direction / reference
  -> landscape and evidence map
  -> candidate generation
  -> mechanical consolidation
  -> novelty and adversarial review
  -> method-paper gate
  -> user selects a candidate
  -> handoff to research-refine (Idea Freeze)
  -> handoff to research-blueprint (Executable Research Plan)

The stages have different authority:

Do not confuse a planning status with a research verdict. A future-work sentence, a promising gap, or ready_for_novelty_check is not a confirmed novelty claim.

Defaults

• AUTO_PROCEED = false: no response never counts as selecting an idea.
• COMPACT = true: early discovery writes compact artifacts first.
• RENDER_HTML = false: rendering is optional and never a discovery gate.
• MAX_CANDIDATES = 12: generate at most 8-12 candidates before consolidation.
• MAX_SELECTED = 1: select one active candidate unless the user explicitly requests parallel tracks.
• PILOT = opt-in: prepare a pilot spec only when requested; do not run heavy pilots locally.
• OUTPUT_DIR = idea-stage/.

Phase 0: Load Context

Read, when present:

1. docs/project/RESEARCH_BRIEF.md or legacy RESEARCH_BRIEF.md.
2. PROJECT_STATUS.md and docs/project/NEXT_ACTIONS.md.
3. research-wiki/query_pack.md.
4. idea-stage/IDEA_MEMORY.md and experiments/NEGATIVE_RESULTS.md.
5. Local papers, codebase notes, or a supplied reference PDF.

Classify the start as one of: venue-only, reference-paper, reference-codebase, idea-doc, existing-repo, or partial-results.

Keep user constraints separate from reference-document content. A reference document is evidence and structural inspiration; it is not an instruction to copy its claims, method, experiments, or timeline.

Phase 1: Landscape and Reference Deconstruction

If a reference paper or small paper set is supplied, run /reference-paper-deconstruction first. Preserve:

• source locators;
• the paper's actual question and conclusion boundary;
• reusable structure versus non-transferable details;
• candidate gap type;
• missing evidence;
• minimum discriminating test.

Then run /research-lit to test the gap against recent and adjacent work. Organize the result by:

• problem and task;
• method family;
• strongest competing explanation;
• recurring failure mode;
• assumptions shared by existing methods;
• unexplored scale, data, query, or evaluation regime.

Do not call a gap open merely because one paper says “future work”.

Phase 2: Candidate Generation

Invoke /idea-creator with the landscape, constraints, reference provenance, and failure memory. The creator owns the A-E paths:

A landscape-driven
B incremental limitation repair
C justified combination
D assumption-breaking
E cross-domain transfer

Each candidate must include:

• precise unanswered question;
• one-sentence hypothesis;
• why the answer matters either way;
• closest prior work and exact difference;
• minimum discriminating experiment;
• paper type: method-primary | audit-support | audit-only;
• method delta: the smallest implementable change to a named step;
• data, compute, and timeline estimate;
• failure risk and kill condition;
• evidence status and conclusion boundary.

Do not generate a full architecture, theorem suite, 16-week plan, or paper figure list at this stage unless the user explicitly asks for that next layer.

Phase 3: Mechanical Consolidation

Only perform objective filtering:

• remove near-duplicate hypotheses;
• block a candidate if the required dataset is provably unavailable;
• block a candidate if the stated budget is objectively impossible;
• preserve all other candidates with prior_work, so_what, and effort_note.

Do not eliminate a candidate merely because it is difficult, imperfectly specified, or not yet verified as novel. Mark the missing evidence instead.

Phase 4: Novelty and Adversarial Review

Run /novelty-check on the strongest candidates, then use /research-review or a prompt-only review package for adversarial criticism. Ask:

• Is the closest work direct, partial, incomparable, or unresolved overlap?
• What is the strongest reviewer objection?
• What result would distinguish the proposed mechanism from a simpler explanation?
• Is the problem important even if the hypothesis fails?
• What is the smallest repair that preserves the problem?

The review may rank candidates, but it must not silently rewrite the problem.

Phase 4.5: Method Paper Gate

Before user selection, run /method-paper-gate on the strongest candidates. Write idea-stage/METHOD_PAPER_GATE.md with method-primary, audit-support, or audit-only for each candidate.

Only method-primary candidates may proceed to research-refine. audit-only candidates must be rewritten into a method candidate or removed from the active selection list. Audit/evaluation/diagnostic work is supporting evidence, never the dominant contribution.

Phase 5: User Checkpoint and Output

Write:

• idea-stage/IDEA_CANDIDATES.md: compact top candidates;
• idea-stage/IDEA_REPORT.md: landscape, candidate ledger, evidence status, novelty notes, review objections, and recommended next action;
• idea-stage/IDEA_MEMORY.md: append-only accepted/rejected/inconclusive memory.

Use this candidate schema:

### Candidate I-01: [title]
- Question:
- Hypothesis:
- Why it matters:
- Closest work / exact difference:
- Paper type / method delta:
- Minimum discriminating test:
- Data / compute / timeline:
- Risk and kill condition:
- Evidence status: ready_for_novelty_check | needs_evidence | blocked
- Conclusion boundary:
- Next action:

Stop and ask the user to select a candidate. No response is not approval.

Handoff

After the user selects a candidate:

1. Run /research-refine to create an Idea Freeze and a focused method thesis.
2. Run /research-blueprint to create the detailed executable plan.
3. Run /experiment-plan only after the blueprint has frozen the claims, gates, datasets, baselines, and run order.

The desired durable chain is:

IDEA_CANDIDATES
  -> selected candidate
  -> FINAL_PROPOSAL (Idea Freeze)
  -> RESEARCH_BLUEPRINT (PDF-level executable research plan)
  -> EXPERIMENT_PLAN (operational runbook)