Chunking: 20-candidate operating point, 14-of-14 validated
A previously-recommended chunking rule, validated at real scale (14 of 14 acceptable), including a self-correction to its own sizing model.
A recommended 20-candidate operating point, independently validated at real scale, with balanced (not greedy) chunking that preserves every candidate exactly once. Large sources overwhelm both cost and quality if fed to a review step unbounded; naive truncation loses the meaning that lives past the cut point. Evidence: CHUNKED-OPERATING-POINT.md, which states a measured 14-of-14 acceptable result and one real correction to its own sizing model, found by checking predicted output against actual output.
Technologies
- Section-bounded chunking
- Balanced (non-greedy) candidate packing
Applicability
Useful for a technical or executive evaluator assessing Chunking specifically; drawn from internal source evidence, not an independently public-published artifact.
Evidence
Evidence class: documented-statement
Frequently asked
What is the evidence behind "Chunking: 20-candidate operating point, 14-of-14 validated"?
CHUNKED-OPERATING-POINT.md, which states a measured 14-of-14 acceptable result and one real correction to its own sizing model, found by checking predicted output against actual output.