v0.1.1

Patch release from the first live smoke (200 rows of the UCI SMS Spam Collection, jev-1.13.0).

• Token estimates are now calibrated on real usage: ~3 chars/token plus ~200 tokens of per-request framing. The smoke's dry-run estimate went from 2.45x under to 1.04x (91,871 estimated vs 95,313 actual).
• README documents the smoke results: 97.4% accuracy across comparable rows, and every auto-accepted row was correct — all 5 errors landed in the review queue. $0.0040 for 200 rows.

v0.1.0

AI columns for CSV/JSONL files, backed by TypeSafe's Jev: label every row with typed answers, probabilities and a review queue — deduplicated, resumable, with a cost preview before you spend anything.

What's in v0.1.0

• jev-table data.csv --spec pack.yaml — columns come from a jev-packs format v0 spec (choice/score/noul questions + thresholds), parsed by the canonical jevassert.packs loader.
• Output: *.jev.csv (answers + _confidence + review + error), *.corrections.csv, *.stats.json/.md, resume cache.
• Review queue: a label auto-accepts only above its threshold floor; no floor (e.g. unknown) or no thresholds means review.
• Flywheel: edit corrections.csv (or delete an id from review to confirm as-is), replay with --corrections over the cache (zero API calls), export --emit-cases as cases.jsonl for jevassert/jev-packs.
• Dedupe identical rows, resume cache, SDK retries with 429/529 backoff, auto-capped concurrency.
• --dry-run cost preview (sends nothing), actual-vs-estimated tokens per run, 10k-row safety cap, pinned-model mismatch warning.
• Input: CSV or flat JSONL. Local-first: row data goes only to the configured endpoint; no telemetry.
• 50 offline tests (mock transport), CI on every push.

Install: uvx jev-table --help or pip install jev-table.
Spec reference: https://github.com/dtduc-git/jev-packs/blob/master/SPEC.md