frontier-select-merge

Selection & Merge

Select top-K beams and merge their insights into accumulated knowledge.

Process

1. Read scored beams ranked by composite score
2. Select top-K beams as "winners"
3. Merge insights from selected beams:
   • Deduplicate claims that express the same insight differently
   • Resolve contradictions between beams (prefer higher-evidence claims)
   • Identify cross-beam patterns — insights independently confirmed by multiple beams
4. Record dead ends from eliminated beams to prevent re-exploration
5. Compute insight delta:
   • Count genuinely new unique insights this epoch
   • Compare to total accumulated insights
   • insightDelta = newUniqueInsights / totalAccumulatedInsights

Insight Delta Interpretation

• > 0.5 — highly productive epoch, major new territory explored
• 0.2 - 0.5 — productive epoch, meaningful new insights
• 0.1 - 0.2 — diminishing returns, approaching convergence
• < 0.1 — very little new, likely converged

Outputs

• selection.json — selected beams, eliminated beams, merged knowledge, insight delta

Quality Criteria

• Exactly K beams selected (matching selectionWidth input)
• Selection is strictly by composite score ranking
• Insight deduplication is genuine (not just string matching — semantic dedup)
• Dead ends from eliminated beams are preserved
• Insight delta computation is traceable (show the counts)