Question 1

Is this a replacement for my database's access control, or a layer on top?

Accepted Answer

A layer, and the project argues the layer matters. Ordinary tooling returned 1250 for a tenant id interpolated from the request and found no tenant column on a field named account_id, while JevSQL takes the tenant from the authenticated actor and refuses to compile until the column is classified. Permissions stay in code and SQL; the model supplies the semantic call.

Question 2

Will a saved decision be locked into a table I cannot query without the library?

Accepted Answer

No. A materialized decision is an ordinary SQLite table with a primary key, so other applications can query it with no library and no API key. Refreshing reuses the saved SQL, parameters, model name and cache namespace, and unchanged judgment inputs reuse cached answers. The source query still scans its candidate rows, so this is incremental work, not change-data capture.

Question 3

Do the 31 comparisons prove it is safe to switch on?

Accepted Answer

No, and the suite says so. Two of the 31 are deliberately not wins: the secret scanner and the typed review each score 2 of 3 alone and 3 of 3 together. The final comparison reports that nothing here is ready to switch on, the ticket sample is ten rows showing shape rather than calibration, and the offline workflow tour's fixture results demonstrate software behavior, not model accuracy.

Question 4

Can I use it on Postgres, or is this SQLite only?

Accepted Answer

SQLite is the core. It builds on the built-in node:sqlite module, needs Node 22.16 or newer and ships zero runtime dependencies. For PostgreSQL or other already-authorized result sets the README points to a separate native row API with typed decisions, explicit projections and review queues, and for exact totals over encrypted data to a streaming money API.

Question 5

Which SQL functions can I call in JevSQL?

Accepted Answer

jev_noul returns a yes-probability, jev_bool applies a threshold, jev_decide sends low and high bands to 0 or 1 with NULL in between, and jev_choice selects a label with confidence, probabilities and top-probability helpers.

Question 6

What limits does JevSQL's planner impose?

Accepted Answer

Planner defaults cap a request at 25 states, 120 questions and 60,000 serialized UTF-8 bytes; maxJudgments defaults to 1,000 new judgments per query and maxEstimatedCostUsd stops before dispatching another round.

Question 7

How are JevSQL decision tables updated?

Accepted Answer

materialize writes an ordinary SQLite table, refresh reuses the saved SQL, parameters, model and cache namespace inside a savepoint, and changes lists added, changed and deleted output rows; nothing runs on a timer automatically.

JevSQL — EugeneBoondock

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

Other listings of this project

From the Jev guides

From the Jev AI model community