hypothesis-evolution-loop

hypothesis-evolution-loop

Goal:

• Run the iterative evolution loop from the current persisted run state until convergence.

Inputs:

• current persisted run artifacts
• active state/STRATEGY_PLAN.json
• optional resume flag

Outputs:

• updated hypotheses
• updated island statistics
• updated proximity receipt/status artifacts, plus state/PROXIMITY_GRAPH.json when the embedding bridge succeeds
• updated tournament artifacts
• updated convergence counters
• updated state/EVOLUTION_STATE.json
• appended state/EVOLUTION_ROUNDS.jsonl
• updated state/PIPELINE_STATE.json
• updated state/CURRENT_STAGE.json

Context Loading:

• Open skills/shared-references/schema-index.md.
• Read packages/agent_contracts/pipeline_control.py before updating state/EVOLUTION_STATE.json or state/COMPLETION_DECISION.json.
• Read packages/agent_contracts/pipeline_runtime.py before updating state/PIPELINE_STATE.json or state/CURRENT_STAGE.json.
• Read packages/agent_contracts/hypothesis.py before writing any evolved hypotheses/<id>/HYPOTHESIS.json artifact.
• Read packages/agent_contracts/state.py before any island state validation, and read packages/run_artifacts/island_state.py before invoking the run-level island persistence helper.
• Read packages/agent_contracts/ranking.py when the round will write tournament artifacts through downstream ranking work.
• Read packages/agent_contracts/evolution_round.py before appending state/EVOLUTION_ROUNDS.jsonl.
• Read the active state/STRATEGY_PLAN.json, state/PIPELINE_STATE.json, state/CURRENT_STAGE.json, state/EVOLUTION_STATE.json, the selected parent hypotheses, and the latest parent review bundles before starting or resuming a round.

Execution Contract:

• Resume decisions must come from artifact state, not hidden process memory.
• The active evolution route must be consistent with state/STRATEGY_PLAN.json.
• Refresh state/STRATEGY_PLAN.json through python -m tools.policy.plan_strategy <run_dir> before every evolution round that is resuming from persisted state.
• Use an explicit phase override only when the caller is intentionally forcing a new stage transition instead of restoring the current persisted round or substage.
• Use from tools import sync_pipeline_stage_artifacts as the canonical paired write surface when entering Evolution, Reflection, Insights from Reviews, Proximity, or Ranking.
• The stage-sync helper is implemented in packages/run_artifacts/stage_sync.py.
• Before resuming or starting an evolution round, call tools.sync_pipeline_stage_artifacts(...) for Evolution.
• Before dispatching review, insights, proximity, or ranking work for the active child, call tools.sync_pipeline_stage_artifacts(...) for the exact substage being entered so dashboard state reflects the live round status.
• Read state/RESOLVED_RUN_CONFIG.json before deciding whether the loop is capped or completion-driven:
  • convergence.max_iterations > 0 means a user-visible capped run.
  • convergence.max_iterations == 0 means completion-driven mode; do not invent a small hard iteration cap.
  • convergence.safety_max_iterations is still a hard safety ceiling and must be recorded if reached.
  • Persist EVOLUTION_STATE.safetyMaxIterations exactly from RESOLVED_RUN_CONFIG.convergence.safety_max_iterations; never rewrite it from the current iteration count.
• Treat iteration_policy = completion_driven together with human_checkpoint = auto as an autonomous loop contract: after one round closes, refresh routing and continue automatically instead of returning to the user for per-round confirmation.
• human_checkpoint can still intentionally pause the loop, but only at the configured checkpoint boundary (before_overview, before_completion, or every_major_stage).
• If the host-agent turn stops before a terminal route is reached, the handoff must state that convergence has not been reached, state is persisted, and the recommended next action is continue evolution through resume or an explicit continue request.
• One evolution round may create only one new child hypothesis. Do not batch-refine every seed candidate once and then stop.
• Do not synthesize placeholder hypotheses, reviews, tournaments, proximity receipts, embeddings, or evolution-round receipts to make progress.
• If an evolve, review, proximity, ranking, convergence, or round-receipt substage cannot run through its canonical skill or tool, stop and report a resumable blocked state instead of writing low-information artifacts.
• When state/STRATEGY_PLAN.json reports next_action: continue_evolution, use signals.selected_parent_ids as the parent set for the next child hypothesis and persist the resulting parent linkage in HYPOTHESIS.json and ORIGIN.json.
• Before generating that child, run evolution-strategy-supervisor and choose exactly one concrete evolution strategy allowed by:
  • signals.selection_strategy
  • selected_evolution_strategies
  • the parent hypotheses and their latest review bundles
• For single-island rounds, use from tools import update_run_single_island_reward as the canonical run-level island reward / decay update and persistence surface after the child has finished review and ranking update receipt coverage.
• For seed island initialization and multi-island child island creation, use from tools import ensure_run_islands_for_hypotheses as the canonical run-level island initialization surface.
• The run-level helper is implemented in packages/run_artifacts/island_state.py and delegates the deterministic reward mechanics to packages/agent_mechanics/island_reward_update.py.
• Do not hand-edit islands/ISLANDS.json, and never write state/ISLANDS.json. The run-level helper must load hypotheses, apply decay/reward, and persist the canonical islands/ISLANDS.json artifact.
• After the run-level helper returns, reload canonical islands/ISLANDS.json and confirm the selected island now has nonzero visit_count and nonzero decayed_visits; the selected island visit_count must also cover the completed single-island round receipts that selected it.
• If the selected island metrics do not validate after a single-island closeout, stop before the next routing refresh, run python -m tools.validation.contract_validation <run_dir> --skill hypothesis-evolution-loop, and report a resumable blocked state instead of continuing on stale UCB state.
• Newly created islands must remain unvisited until they are later selected by a completed single-island round: verify decayed_reward = 0.0, decayed_visits = 0.0, and visit_count = 0 after calling the initialization helper.
• 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.
• The lower-level tools.update_single_island_reward(...), tools.compute_single_island_reward(...), and tools.apply_decayed_island_update(...) surfaces are internal/testing fallbacks only for in-memory mechanics; do not use them as the host-agent evolution-loop persistence path.
• Use from tools import evaluate_convergence as the canonical convergence-counter update surface after the child's final top-k membership is known.
• The convergence helper is implemented in packages/agent_mechanics/convergence_check.py.
• Use from tools import append_evolution_round_record as the canonical append-only write surface for state/EVOLUTION_ROUNDS.jsonl after the child has completed review, proximity, ranking, and convergence updates.
• state/EVOLUTION_ROUNDS.jsonl is the completed round receipt. It does not replace state/STRATEGY_DECISIONS.jsonl, and it must not be used to rewrite prior routing decisions.
• The round receipt's placement_match_ids and ranked_match_ids must be child-owned closeout refs for that round only:
  • placement refs require match_strategy = placement_tournament and hypothesis_1_id = child_hypothesis_id
  • ranked refs require match_strategy = ranked_tournament and hypothesis_1_id = child_hypothesis_id
  • duplicate match IDs are invalid
  • later matches where the child appears as hypothesis_2_id may remain in the child's lifetime HYPOTHESIS.json refs, but they must not be copied into this earlier round receipt
• After appending a completed round receipt, update state/EVOLUTION_STATE.json from the same convergence helper output and refresh state/STRATEGY_PLAN.json; do not leave pre-round router signals as the current plan.
• The referenced continue_evolution decision must be a canonical router record from python -m tools.policy.plan_strategy <run_dir> with router-level signals. Do not add child IDs, chosen concrete strategies, proximity statuses, tournament IDs, top-k entry results, or convergence transitions to state/STRATEGY_DECISIONS.jsonl.
• Treat the referenced router signals as a replayable pre-round state contract:
  • signals.hypothesis_count and signals.viable_hypothesis_count are the counts before the child is created.
  • signals.convergence_count is the value before the child is created and must become the round receipt's convergence_count_before.
  • signals.entered_top_k_last_round must reflect the prior completed round when one exists.
  • signals.top_hypothesis_ids must become the round receipt's previous_top_k_ids.
  • signals.selected_parent_ids must reference viable hypotheses already persisted before the child is created.
  • signals.selected_island_ids must match the selected parents' persisted island_id values.
• A completed evolution round must be replayable from exactly one router decision, one evolved child, one review bundle, one proximity receipt, completed ranking artifacts with ranking update receipt coverage, one convergence update, and one appended round receipt.
• After emitting the child hypothesis, you must:
  • run hypothesis-review-pipeline
  • run insights-from-reviews when the child is viable
  • run hypothesis-proximity-update for each viable child hypothesis
  • run hypothesis-ranking-pipeline
  • for signals.selection_strategy == single_island, call tools.update_run_single_island_reward(run_dir, selected_island_id, candidate_hypothesis.id, decay_factor) exactly once after the child's final Elo and viability state are known, and use the helper-persisted canonical islands/ISLANDS.json artifact as the source of truth
  • after that single-island reward update, verify the helper-persisted selected island has nonzero visit_count and decayed_visits; do not append a completed round receipt or refresh routing if the selected island metrics are stale
  • for signals.selection_strategy == multi_island, assign the child a new non-empty island_id, call tools.ensure_run_islands_for_hypotheses(run_dir, [candidate_hypothesis.id]), verify the created island remains unvisited, and do not apply the single-island reward helper to that cross-island child
  • call tools.evaluate_convergence(candidate_hypothesis.id, previous_top_k_ids, current_top_k_ids, current_convergence_count) exactly once after the final ranked frontier is known, then persist both enteredTopKLastRound and convergenceCount exactly as returned
  • call tools.append_evolution_round_record(run_dir, record) exactly once with an EvolutionRoundRecordContract that records the decision index, selected parent IDs, selected island IDs, chosen concrete evolution strategy, child hypothesis ID, child island ID, proximity receipt status, child-owned placement/ranked tournament match IDs, previous/current top-k IDs, entered_top_k, and convergence counts before and after the round
  • update state/EVOLUTION_STATE.json
• When next_action: return_to_generation, leave the evolution loop, run the requested generation pass, then refresh the evolution plan again.
• Do 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. Let ranking use its documented receipt-gated missing-graph fallback only after hypothesis-proximity-update has recorded a skipped or failed receipt/status.
• The evolution-round proximity_receipt_status must be copied from the persisted per-child state/proximity_receipts/<child_hypothesis_id>.json receipt. Do not infer this field from environment variables, provider availability, or dashboard state.
• The loop is complete only when state/EVOLUTION_STATE.json records a terminal stop reason and the routing audit log contains one decision record per evolved hypothesis.
• When writing state/EVOLUTION_STATE.json, persist the active stopPolicy, iterationPolicy, iterationBand, and safetyMaxIterations metadata from the effective run configuration and policy artifacts.
• Only record stopReason: safety_iteration_limit_reached when iterationCount >= RESOLVED_RUN_CONFIG.convergence.safety_max_iterations and safetyLimitHit is true.
• When consumed by a host agent, run python -m tools.validation.contract_validation <run_dir> --skill hypothesis-evolution-loop after each evolved hypothesis is persisted.
• When the loop reports status: completed, run python -m tools.validation.verify_pipeline_completion <run_dir> --skill hypothesis-evolution-loop.
• The final research overview remains a separate step in research-overview-pipeline.
• Treat complete as a completion-verifier outcome after overview, not as a state/STRATEGY_PLAN.json route.