Retrieval Depth and Contradiction Search — Run 001
Status · SUPPORTED
RESEARCH REPORT
July 23, 2026
Research question
Does bounded breadcrumb traversal plus contradiction search recover materially more relevant evidence than nearest-page retrieval on one long document?
Prior hypothesis
Following adjacent evidence and then searching explicitly for contradiction and revision language would recover more predeclared evidence anchors than the nearest page alone.
Method
A 24-page authoritative paper was extracted page by page. One fixed question and five evidence-anchor groups were declared before the run. Nearest-page retrieval, adjacent breadcrumb expansion, and the same breadcrumb plus a maximum of four contradiction-marker pages were compared.
Observations
- Nearest-page retrieval selected one page and recovered two of five anchors.
- Adjacent breadcrumb retrieval selected three pages and recovered four of five anchors.
- Breadcrumb plus contradiction search selected seven pages and recovered all five anchors.
- Retrieved characters increased from 2,522 to 7,720 to 17,503.
Supporting evidence
- The source contained 24 extracted pages.
- The source hash was recorded before testing.
- All variants used the same fixed query and predeclared anchors.
- Anchor coverage increased from 0.4 to 0.8 to 1.0.
- Contradiction search recovered a limit missed by adjacent traversal.
- The machine-readable result is preserved.
Counterevidence
- Anchor coverage measures retrieval, not review quality.
- Only one document and one question were tested.
- The anchor set was selected by the experiment designer.
- No model conclusion, latency, or provider cost was compared.
Conclusion
For this document and question, bounded breadcrumb traversal recovered more declared evidence than nearest-page retrieval, and contradiction search recovered one additional limit at greater context cost. The run supports deeper retrieval as a candidate method, not a universal optimal stopping rule.
Confidence
High for this run’s retrieval counts; low for generalization to review quality or other documents.
Limits
- One source cannot establish a general retrieval policy.
- The experiment did not assess conclusion stability.
- Character count is only a rough cost proxy.
- The optimal stopping condition remains unresolved.
Production consequence
No production retrieval path is replaced. Contradiction search enters the Review Lab backlog for independent replication.
Next unresolved question
Does the coverage gain persist across independent long documents, and when does additional context stop changing the conclusion?