Evolution Operators

This skill teaches you the four evolutionary operators used to derive improved hypothesis variants from existing hypotheses. Each operator targets a specific weakness identified by reviewers or tournament loss patterns. Every variant records its lineage — parent IDs and the operator used — so the system can track evolutionary history and enforce mandatory rematches.

This protocol contributes records to a run that the supervisor ultimately publishes with oase hypex analyze; workers do not publish a separate result.

Operator Reference

1. Ground

Purpose: Strengthen the weakest evidential claim in a hypothesis by finding new supporting or contradicting literature.

When to use: Review criticisms cite weak evidence, unsupported claims, or missing citations.

Procedure:

1. Read the hypothesis and its reviews. Identify the weakest evidential claim — the one reviewers flagged as poorly supported or speculative.

2. Run targeted literature searches to find evidence that supports or contradicts the weak claim:

   # PubMed for peer-reviewed biomedical evidence
   oase pubmed search "<mechanism>[MeSH Terms] AND <target>[MeSH Terms]" --max-results 20
   
   # arXiv for computational/theoretical evidence
   oase preprint search --source arxiv "cat:<category> AND abs:<keywords>" --max-results 10
   
   # bioRxiv for recent preprints
   oase preprint search --source biorxiv "<keywords>" --max-results 10

3. If new supporting evidence is found: Strengthen the claim with additional citations and refine the mechanism description to incorporate the new findings. Add the new literature to the evidence array.
4. If contradicting evidence is found: Rewrite the claim to accommodate the contradiction — narrow the scope, add a qualifying condition, or pivot the mechanism. Add the contradicting evidence with "role": "contradicts" or "role": "constrains".
5. Produce the variant hypothesis with the updated evidence array and revised statement/mechanism as needed.

Lineage: One parent.

"lineage": {"parents": ["H-0042"], "operator": "ground"}

2. Simplify

Purpose: Reduce experimental complexity and strengthen parsimony while preserving the core testable prediction.

When to use: Review criticisms cite overcomplexity, untestable experimental designs, too many moving parts, or difficulty distinguishing between alternative explanations.

Procedure:

1. Read the hypothesis and its reviews. Identify sources of complexity:
   • Experiments with "est_difficulty": "high" that could be replaced with simpler alternatives
   • Redundant experiments that test overlapping aspects of the hypothesis
   • Mechanism descriptions that invoke multiple untested intermediate steps
   • Predictions that are not independently testable
2. Simplify while preserving the core insight:
   • Reduce experimental complexity: Replace high-difficulty experiments with lower-difficulty alternatives where possible. Consolidate redundant experiments into a single, cleaner design.
   • Sharpen the mechanism: Remove speculative intermediate steps. Focus on the most directly testable causal chain.
   • Tighten predictions: Remove predictions that are derivative of or redundant with other predictions. Each prediction should add independent testable information.
   • Preserve falsifiability: The simplified hypothesis must still be refutable by the remaining experiments.
3. Produce the variant hypothesis with simplified experiments, mechanism, and predictions. Do not drop evidence — evidence items should remain unless they support a claim you removed.

Lineage: One parent.

"lineage": {"parents": ["H-0042"], "operator": "simplify"}

3. Combine

Purpose: Merge complementary aspects of two hypotheses from the same proximity cluster into a stronger unified variant.

When to use: Two hypotheses in the same cluster address complementary aspects of the same phenomenon — one explains mechanism A, the other explains mechanism B, and a combined A+B mechanism is more complete and testable than either alone.

Procedure:

1. Read both parent hypotheses and their reviews. Identify complementary strengths:
   • Does one hypothesis have stronger evidence where the other is weak?
   • Does one propose a mechanism that explains an observation the other cannot?
   • Do their experimental designs test different aspects of the same system?
   • Can their predictions be unified into a more powerful joint prediction?
2. Do not simply concatenate. A combined hypothesis is not "hypothesis A plus hypothesis B." It is a new hypothesis that synthesizes the complementary insights into a single coherent mechanism with unified predictions and experiments.
3. Build the combined variant:
   • Statement: A new falsifiable claim that integrates both parents' insights. It should be stronger or broader than either parent alone.
   • Mechanism: A unified causal mechanism that incorporates both parents' mechanistic contributions.
   • Predictions: Include predictions that test the joint mechanism — especially predictions that would distinguish the combined hypothesis from either parent individually.
   • Experiments: Design experiments that test the interaction or combined effect, not just each parent's claim independently.
   • Evidence: Merge evidence arrays from both parents. Deduplicate by lit_id — if both parents cite the same paper, keep one entry with the more informative note.
4. Produce the variant hypothesis with both parent IDs in the lineage.

Lineage: Two parents (required).

"lineage": {"parents": ["H-0042", "H-0047"], "operator": "combine"}

4. Out-of-the-Box (oob)

Purpose: Transfer an insight from an adjacent or distant field to create a novel variant with a fresh perspective.

When to use: The hypothesis space feels saturated — all hypotheses share the same literature base, the same mechanisms, and the same experimental approaches. Or when a hypothesis has stalled in tournament rankings without clear evidence-based weaknesses to fix.

Procedure:

1. Read the hypothesis and its reviews. Identify the core phenomenon or mechanism it addresses.

2. Search for analogies in adjacent or distant fields:

   # Cross-category arXiv search — look for structural analogies
   oase preprint search --source arxiv "cat:q-bio.MN AND abs:<analogous mechanism>" --max-results 10
   oase preprint search --source arxiv "cat:cs.AI AND abs:<phenomenon keyword>" --max-results 10
   oase preprint search --source arxiv "cat:physics.bio-ph AND abs:<system keyword>" --max-results 10
   
   # Cross-domain PubMed search
   oase pubmed search "<distant field mechanism>[MeSH Terms] AND <target system>[MeSH Terms]" --max-results 15
   
   # Broad exploration for unexpected connections (OASE fan-out)
   oase pubmed search "<phenomenon> <distant field keyword>" --max-results 15
   oase preprint search --source arxiv "<phenomenon> <distant field keyword>" --max-results 10
   oase preprint search --source biorxiv "<phenomenon> <distant field keyword>" --max-results 10

3. Identify a transferable insight: a mechanism, mathematical framework, experimental technique, or conceptual model from the distant field that maps onto the hypothesis's domain.
4. Build the variant:
   • Statement: Incorporate the cross-disciplinary insight into a new falsifiable claim. Make the analogy explicit — state what is being transferred and why it applies.
   • Mechanism: Describe the mechanism using the transferred framework. Explain both the analogy (why the systems are structurally similar) and the specific prediction it generates.
   • Evidence: Include literature from both the original field and the source field. The source-field evidence supports the analogy; the original-field evidence grounds the claim in the target domain.
   • Experiments: Design experiments that specifically test whether the transferred insight holds in the target domain.
5. Produce the variant hypothesis with a single parent and "oob" operator.

Lineage: One parent.

"lineage": {"parents": ["H-0042"], "operator": "oob"}

Lineage Rules

Every variant hypothesis must record its lineage accurately. The lineage tracks evolutionary history and enables mandatory rematches between parents and their offspring.

Structure

"lineage": {
  "parents": ["H-NNNN"],
  "operator": "ground|simplify|combine|oob"
}

Parent Count by Operator

Rules

1. Every variant must have at least one parent. A variant with an empty parents array is invalid (except for "operator": "null" originals).
2. combine must have exactly two parents. Both must be from the same proximity cluster.
3. Parents are retained. Variants compete against their parents in the next tournament round. The supervisor writes parent-vs-child pairings to a JSON file and passes it to elo pair --rematch <path>, which prepends them as forced rematches that bypass the rematch window.
4. Lineage is immutable. Once a hypothesis is submitted, its lineage cannot be changed.
5. Do not set operator to "null" for variants. Only original hypotheses (produced by generation agents) use "null".

Operator Selection Guide

Choose the operator based on the hypothesis's weaknesses as identified by reviewers and tournament outcomes:

When multiple signals apply, prioritize:

1. ground if evidence quality is the primary weakness — a well-grounded hypothesis survives all other criticisms better.
2. simplify if the hypothesis is strong in evidence but too complex to test — parsimony wins tournaments.
3. combine only when two hypotheses genuinely complement each other — do not force combinations between unrelated hypotheses.
4. oob as a strategic diversification move — use sparingly, not on every hypothesis.