hypothesis-generation-pipeline

hypothesis-generation-pipeline

Goal:

• Dispatch one enabled generation strategy for a given ResearchPlan.

Inputs:

• ResearchPlan
• active state/STRATEGY_PLAN.json
• one enabled generation strategy
• optional resume context from state/PIPELINE_STATE.json

Outputs:

• hypotheses/<id>/HYPOTHESIS.json
• hypotheses/<id>/HYPOTHESIS.md
• hypotheses/<id>/ORIGIN.json
• literature/queries/<query_id>/* when hypothesis-generate-literature is selected
• hypotheses/<id>/REVIEW/*.json
• updated meta/INSIGHTS_FROM_REVIEWS.json when the hypothesis is viable
• updated proximity receipt/status artifacts, plus state/PROXIMITY_GRAPH.json when the embedding bridge succeeds
• updated tournaments/*.json
• updated islands/ISLANDS.json
• updated state/PIPELINE_STATE.json

Sub-skills:

• hypothesis-generate-literature
• literature-search
• hypothesis-generate-debate
• hypothesis-generate-assumptions
• hypothesis-review-pipeline
• insights-from-reviews
• hypothesis-proximity-update
• hypothesis-ranking-pipeline

Context Loading:

• Open skills/shared-references/schema-index.md.
• Read packages/agent_contracts/research_plan.py before consuming research_plan/RESEARCH_PLAN.json as the canonical generation brief.
• Read packages/agent_contracts/strategy_plan.py before consuming state/STRATEGY_PLAN.json as the active generation-stage routing input.
• Read packages/agent_contracts/hypothesis.py before dispatching any atomic generation skill that will write hypotheses/<id>/HYPOTHESIS.json.
• Read packages/agent_contracts/literature.py and skills/shared-references/literature-search-contract.md before dispatching hypothesis-generate-literature.
• Read packages/agent_contracts/state.py before assigning or updating islands/ISLANDS.json.
• Read packages/agent_contracts/pipeline_runtime.py before updating state/PIPELINE_STATE.json.
• Read state/STRATEGY_PLAN.json and confirm the current round permits the selected generation strategy.
• Read RUN_POLICY.yaml when review rigor or downstream optional review behavior depends on the effective run policy.

Execution Contract:

• This pipeline skill does not own hidden prompt templates.
• state/STRATEGY_PLAN.json is a routing input to this skill, not a generation-stage artifact that this skill may rewrite ad hoc.
• research_plan/RESEARCH_PLAN.json is a required canonical input. If it is missing or invalid, stop immediately and return control to the top-level workflow or configuration stage instead of attempting generation.
• Append-only routing audit artifacts such as state/STRATEGY_DECISIONS.jsonl remain owned by the top-level orchestration layer and the canonical router surface in python -m tools.policy.plan_strategy <run_dir>.
• It coordinates downstream generation, review, insights, proximity, and ranking skills; those downstream skills remain responsible for the field-level canonical shapes of the artifacts they write.
• The selected generation strategy must be consistent with state/STRATEGY_PLAN.json.
• When the selected generation strategy is literature_exploration_generation, the dispatched hypothesis-generate-literature skill must call tools.search_literature(run_dir, request) and consume a non-blocked EvidenceBundleContract before writing a literature-grounded hypothesis.
• The generation pipeline must not accept prompt-invented literature evidence in place of literature/queries/<query_id>/EVIDENCE_BUNDLE.json.
• On a fresh run, execute one generated hypothesis per selected generation strategy so the initial frontier mirrors the full seed set instead of collapsing to a single synthetic output.
• On a regeneration pass triggered from evolution, execute one generated hypothesis per selected generation strategy and then refresh the evolution plan before continuing.
• After each generated hypothesis is written, immediately run:
  • hypothesis-review-pipeline
  • insights-from-reviews when the hypothesis is viable
  • hypothesis-proximity-update for each viable hypothesis
  • hypothesis-ranking-pipeline
• The generation pipeline must not skip proximity because no embedding vector is already present. The bridge owns provider invocation and records a receipt/status when the provider is disabled, unavailable, invalid, or failed.
• Do not generate, infer, or fabricate embeddings in prompt output. Proceed to ranking through the documented receipt-gated fallback path only after hypothesis-proximity-update has recorded a skipped or failed receipt/status.
• For initial frontier seeding, assign one non-empty island_id to each viable generated hypothesis, then call tools.ensure_run_islands_for_hypotheses(run_dir) before the round is considered complete.
• Newly created seed islands are initialization records only: they must keep decayed_reward = 0.0, decayed_visits = 0.0, and visit_count = 0.
• Initial frontier island assignment is distinct from the later single-island reward / decay mechanics in hypothesis-evolution-loop; do not apply tools.update_single_island_reward(...), tools.update_run_single_island_reward(...), or any manual reward/visit increment during seeding.
• Persist only canonical island fields. Do not add dashboard-derived or router-derived fields such as hypothesis_ids, ucb_score, or strategy_label to islands/ISLANDS.json.