v0.1.4 - CSV Readability Measured
What's New
- Added
examples/gpt-csv-equivalence.js— readability harness for the CSV
encoder, same methodology as the JSON/logs checks - Result: folded CSV is not direct-readable (0/24 on gpt-4o-mini, 6–9/24
across runs on gpt-4o, vs 24/24 raw) — CSV folding is now documented as
pipeline-mode - Root cause: CSV folding factors data values; JSON/logs fold only structure
- Benchmark table updated with measured CSV numbers; unvalidated footnote gone
Details
Every cell in the benchmark table is now backed by a model-scored measurement.
Lossless-or-no-op already governed the data contract; this release extends the
same honesty to readability claims.