ideation

name	ideation
description	LLM guidance for ResearchOS T4 idea generation.

T4 Ideation Guidance

Use this guidance for the Research Idea Formation & Evolution workflow. T4 is not a one-shot hypothesis writer: it forms a traceable Candidate Population, evolves it, and then asks the researcher to decide what should advance.

Evidence Responsibility

• Read the workspace-derived Evidence Index before proposing an Opportunity Map. Core and Bridge paper notes from full/partial reading and abstract-level reading are all available for recall.
• Respect Evidence Permission. Abstract-only material may broaden recall, identify taxonomy or bridge leads, and trigger a reading upgrade. It must not establish a mechanism, validate a design rationale, or support a final claim.
• Preserve source paths, section locators, reading level, uncertainty, and upgrade requirements in every Candidate Genome. Do not turn a missing source into an invented citation.
• Treat missing_areas.md as retrieval-coverage telemetry. It can motivate a Gap Exploration route only after synthesis and paper notes support a research question.
• When a promising route has a precise, searchable evidence gap, run at most one bounded targeted_literature_supplement for that gap instead of abandoning the route or asking the user to restart T2. New abstract notes broaden recall only. The central mechanism and design rationale still require claim-usable reading evidence.
• Combine the retrieved literature, project facts, user constraints, scientific reasoning, analogies, and model knowledge in one coherent analysis. Do not expose artificial evidence-versus-model channels or label every ordinary inference. Preserve traceability for the material external premises that a Candidate depends on. Verify a paper attribution, established mechanism, empirical result, numeric fact, resource claim, or external novelty claim before presenting it as known.

Role Boundary

• IdeaGeneratorAgent plans Opportunities and generates route-scoped Idea Seeds. It never scores, ranks, selects, archives, or declares novelty.
• IdeaScoringAgent independently scores Candidates. It does not generate, rewrite, merge, or delete Ideas. Its input hides route and parent/child identity.
• IdeaEvolverAgent creates only plan-bounded Mutation Child or Compatibility-gated Crossover Child Candidates. It cannot alter Parent selection, Gene Donor Maps, or Survival Selection.
• Runtime code owns fingerprints, Evidence Permission checks, families, lineage, contracts, atomic artifacts, and rollback. It must not author scientific content.

P0 Formation

• Generate the smallest P0 that still exposes materially different opportunity families. Treat Literature, Informed Brainstorm, Mechanism Challenge, Reverse Operation, Subgroup Failure, Gap Exploration, and Cross-domain or Bridge as lenses, not quotas. Skip a lens that adds no distinct decision value, and spend additional Candidates only where evidence or scientific reasoning reveals genuinely different mechanisms, settings, or falsification paths.
• A supplementary or Bridge route may be unsupported. Record the reason and preserve the route result rather than fabricating an Idea.
• If a workspace has a confirmed bridge plan, return one LLM-authored Bridge review for every bridge. A supported bridge Candidate is visible at Gate1. A missing viable Candidate requires an explicit no_candidate_available escape hatch with reason, kill criterion, and revisit condition.
• Every mature Candidate requires a concise title, one-line thesis, complete Idea Genome, the smallest non-duplicative set of one to four Contributions and draft hypotheses, a falsification path, risks, Evidence Composition, and an LLM-authored Gate1 presentation.

Evolution And Selection

• Standard mode performs one bounded P0 -> P1 improvement round when critique reveals a material, repairable weakness. Mutation and crossover counts are ceilings, not required output. Do not create children, repeat scoring, or run a crossover merely to fill a population shape. Preserve independent scoring and lineage when an actual evolutionary operation occurs.
• Keep Parents and archived Candidates. Do not silently remove an Idea because it was not selected for the Portfolio.
• A Candidate can be selected as a complete direction. Multiple complete Candidates are parallel by default; never merge them merely because the user named more than one.
• A cross-Candidate Hypothesis, Contribution, or Gene request creates a Human-composed Candidate only after Compatibility Check, independent scoring, and explicit confirmation. Never concatenate hypothesis strings into a final file.
• A complete Candidate selection first creates ideation/hypothesis_brief.yaml, ideation/selected/hypothesis_lineage.json, and ideation/selected/t45_search_targets.json. These are Pre-Novelty drafts, not formal experimental claims.