[SwineCoder101]

The same 20,000 WikiBio records written seven ways. Five controlled formats hold identical facts: templated prose, bullets, key-value, JSON and a table row. Two references: the article's own paragraph, and a key-value header above it. Tested against 500 lookup and 500 descriptive queries, on bge-small and MiniLM.

Results

• Layout matters more than any codec here. With the facts held fixed, it moves exact f32 nDCG@10 by up to 2× (MiniLM lookups: prose 0.179, table 0.353). Templated prose is last in every context.
• The model decides the format, not the query type. JSON is best on bge-small, the table row on MiniLM.
• There is no accuracy-for-speed trade-off. A binary-hamming scan plus 100 f32 re-scores (0.22 ms against 4.1 ms) keeps 0.96–0.99 of the best format's nDCG. It beats exact f32 on prose in every context.
• Noise ÷ band predicts the re-score depth on a held-out format within a factor of 1.16 at 95% recall (Spearman 0.985).

Changes

• hvec ingest --whole-files: one chunk per file, verbatim, line breaks kept. It conflicts with --chunk-words and --overlap-words. Covered by a unit test and a pipeline test.
• experiments/2026-11-text-format/: the WikiBio converter, run.sh, summarise.py, the write-up and the results (640 KB). The run took about an hour on an M4 Pro and used no API calls.
• Site: docs/_experiments/text-format.md and the WikiBio reference.
• README: the count of published experiments, corrected from four to ten.
• Roadmap: the entry for this experiment, plus two follow-ups (separate embedded and returned text, per-query scores for confidence intervals).

Caveats

• With 500 queries per type, gaps under about 0.02 are ties.
• The queries are templated and the judgments are exact-match, so absolute nDCG is low.
• The domain is biographies only, and only retrieval is measured.