jevsql

# JevSQL

For exact totals over large or encrypted datasets, use the
[streaming money API](docs/streaming-money.md). It reads bounded batches and sums
decimal values by currency without sending money arithmetic to a model.

For PostgreSQL or other already-authorized result sets, use the
[native row API](docs/native-rows.md): typed decisions, explicit projections,
tenant-scoped caching, budgets, cancellation, and review queues without SQLite.

**Turn database rows into decisions you can query, inspect, refresh, and test.**

JevSQL adds TypeSafe Jev judgments to SQLite. It can compare records by meaning, select exact evidence from text, rank retrieved passages, route uncertain results to review, and detect when a changed source invalidates an earlier decision. It runs as a Node library or CLI with zero runtime dependencies.

| Problem | Working solution |
|---|---|
| An AI answer cites evidence that no longer supports its claim | Save an evidence audit, refresh it, and inspect the before/after decision history. |
| The same business appears under different names in two systems | Use SQL to narrow candidate pairs, then `jev_match` to estimate whether they refer to the same entity. |
| A generated contact address or amount contains invented characters | Find candidates in code, then use `jev_pick` to select an exact source span or return `NULL`. |
| A retrieved passage is related but does not answer the question | Rank permission-filtered passages with descriptive score levels. |
| Automation confidently guesses when evidence is missing | Use explicit unknown labels, abstention bands, and review queues. |
| Nobody knows which confidence threshold is useful | Evaluate labeled rows once and compare accuracy, coverage, and review workload across thresholds. |
| A database pass