Congealing A/B Explainer Lab

Rebuilt from your corrected files. Treatment path models decay-weighted contextual snippet expansion around anchor snippets.

Sources: Technical Description_ Decay-Weighted Contextual Snippet Expansion (Congealing).docx • Congealing Provisional Patent.docx • Congealing_PPA_final.docx

A: Baseline Retrieval

  • C4 • Technical Description • 0.667
    Congealing score combines distance decay and a relevance/similarity boost: C(S0,Si)=β·D(di)+γ·(R0·Ri·Sim(S0,Si)).
  • C5 • Technical Description • 0.333
    Candidate neighbors may be selected by threshold τ, top-K ranking, or probability sampling based on congealing score.
  • C10 • Congealing Final • 0.167
    The method can be evaluated by comparing answer quality from baseline retrieval versus decay-weighted contextual expansion.

Answer to "How is the congealing score computed?": Congealing score combines distance decay and a relevance/similarity boost: C(S0,Si)=β·D(di)+γ·(R0·Ri·Sim(S0,Si)). Candidate neighbors may be selected by threshold τ, top-K ranking, or probability sampling based on congealing score. The method can be evaluated by comparing answer quality from baseline retrieval versus decay-weighted contextual expansion.

B: Decay-Weighted Expansion

  • C4 • Technical Description • 0.377 • expanded
    Congealing score combines distance decay and a relevance/similarity boost: C(S0,Si)=β·D(di)+γ·(R0·Ri·Sim(S0,Si)).
  • C5 • Technical Description • 0.216 • expanded
    Candidate neighbors may be selected by threshold τ, top-K ranking, or probability sampling based on congealing score.
  • C2 • Technical Description • 0.216 • expanded
    Anchor snippets are highly relevant sentences identified after initial retrieval and before final synthesis.
  • C3 • Technical Description • 0.216 • expanded
    For each anchor S0, surrounding sentences Si are evaluated inside a fixed window of neighboring sentences.
  • C6 • Congealing Provisional • 0.216 • expanded
    Semantic chunking improves boundaries but can be computationally expensive and still miss physically distant critical context.
  • C10 • Congealing Final • 0.127 • direct
    The method can be evaluated by comparing answer quality from baseline retrieval versus decay-weighted contextual expansion.

Answer to "How is the congealing score computed?": Congealing score combines distance decay and a relevance/similarity boost: C(S0,Si)=β·D(di)+γ·(R0·Ri·Sim(S0,Si)). Candidate neighbors may be selected by threshold τ, top-K ranking, or probability sampling based on congealing score. Anchor snippets are highly relevant sentences identified after initial retrieval and before final synthesis.

A/B Program Summary (proxy metric)

Metric = keyword recall against expected concept terms per test query. Use this as a prototype harness before swapping in human eval or LLM-as-judge scoring.

Control Avg: 0.900
Treatment Avg: 0.650
Lift: -0.250