| name | hypothesis-generate-literature |
| description | Generate exactly one literature-grounded hypothesis candidate for the active round. |
hypothesis-generate-literature
Goal:
- Generate exactly one literature-grounded hypothesis candidate for the active round.
Inputs:
research_plan/RESEARCH_PLAN.json
state/STRATEGY_PLAN.json
literature/queries/<query_id>/EVIDENCE_BUNDLE.json produced by tools.search_literature(...)
- optional parent hypothesis and review artifacts when the round is part of an evolution continuation
Outputs:
hypotheses/<id>/HYPOTHESIS.json
hypotheses/<id>/HYPOTHESIS.md
hypotheses/<id>/ORIGIN.json
literature/queries/<query_id>/* search bridge artifacts for the evidence query used by this generation
Context Loading:
- Open
skills/shared-references/schema-index.md.
- Open
skills/shared-references/literature-search-contract.md.
- Read
packages/agent_contracts/literature.py before building or consuming any search bridge request or evidence bundle.
- Read
research_plan/RESEARCH_PLAN.json.
- Treat
research_goal as the task anchor.
- Treat
preferences as the quality axes that the hypothesis should optimize for.
- Treat
constraints as hard boundaries that the hypothesis and experiment design must satisfy.
- Read
state/STRATEGY_PLAN.json and confirm that the current round allows literature_exploration_generation.
- If the round is parented, read the selected parent hypothesis and its latest review summary before generating a child. Improve the known weaknesses instead of paraphrasing the parent.
Execution Prompt Contract:
- System Intent:
- You are generating one scientifically grounded candidate hypothesis from literature exploration.
- Required Reasoning Focus:
- Before generating the candidate, call
tools.search_literature(run_dir, request) with consumer="hypothesis-generate-literature" and use the returned EvidenceBundleContract as the formal external evidence input.
- Use relevant prior work, gaps, contradictions, or unexplored connections from the evidence bundle to motivate the candidate.
- Produce a specific falsifiable claim with a mechanism and a concrete experiment path.
- Make the mechanism explicit enough that another reviewer can critique it step by step.
- When novelty depends on a conjectural link, keep that link explicit rather than hiding it in vague wording.
- Do Not Do:
- Do not merely restate established literature.
- Do not emit multiple final candidates in one round.
- Do not ignore explicit constraints from the research plan.
- Do not write an informal summary in place of the canonical hypothesis artifact.
- Do not invent papers, DOIs, arXiv IDs, venues, citation counts, abstracts, or literature claims not present in the evidence bundle.
- Do not treat model memory as a substitute for
tools.search_literature(...).
- Output Shape:
- Wrap the final result into the canonical shared
HypothesisContract.
origin.content.statement: 2-3 sentences maximum.
origin.content.mechanism: 2-3 sentences maximum.
origin.content.experimental_design: one concise multiline string with 3-6 numbered steps.
origin.content.experimental_design must remain one string field containing embedded line breaks; do not emit it as a list, array, or nested object.
origin.content.summary: one concise sentence.
origin.content.category: 1-5 words.
Execution Steps:
- Open
skills/shared-references/schema-index.md, skills/shared-references/literature-search-contract.md, then read packages/agent_contracts/hypothesis.py and packages/agent_contracts/literature.py before writing hypotheses/<id>/HYPOTHESIS.json or consuming search bridge artifacts.
- Read the required artifacts.
- Confirm that this round is allowed to use literature exploration.
- Build a focused
SearchRequestContract for the active research goal and call tools.search_literature(run_dir, request).
- Read the returned
EvidenceBundleContract. If retrieval_metadata.status is blocked, stop or return a degraded state; do not write a literature-grounded hypothesis.
- Identify one literature-grounded gap or underexplored mechanism from the evidence bundle that can answer the active research goal.
- Generate exactly one candidate hypothesis for this round.
- Wrap the result into the canonical
HypothesisContract. Any origin.retrieval_results entries must be derived from the evidence bundle rather than invented in prompt text.
- Write
hypotheses/<id>/HYPOTHESIS.json, hypotheses/<id>/HYPOTHESIS.md, and hypotheses/<id>/ORIGIN.json.
- Validate the emitted artifacts before declaring success.
Artifact Rules:
hypotheses/<id>/HYPOTHESIS.json is the authoritative artifact and must serialize the canonical shared contract.
ORIGIN.json may mirror the generation payload and grounding notes, but it does not replace the canonical hypothesis artifact.
- Formal external evidence must be traceable to
literature/queries/<query_id>/EVIDENCE_BUNDLE.json.
- Because
origin.strategy is literature_exploration_generation, the canonical hypothesis origin must include non-empty retrieval_results, evidence_bundle_ids, and literature_query_ids derived from the returned evidence bundle; validation fails if these bridge linkages are omitted or point to unknown artifacts.
- Any literature search or evidence-gathering notes that are not in the canonical evidence bundle should live in the trace, not be mixed into the contract fields.
Completion Rule:
- This skill is complete only when exactly one new canonical hypothesis artifact has been written from a non-blocked evidence bundle and the emitted files validate for downstream review.