https://jepela.com/feed.xml2026-10-15T00:00:00ZJepela https://jepela.com/blog/when-not-to-use-an-llm.html2026-10-15T00:00:00Z

If your code needs one judgment from a known set of answers, a text generator is more than you need. When to use a decision engine, and when not.

https://jepela.com/blog/what-our-engine-gets-wrong.html2026-10-15T00:00:00Z

The Jepela engine does not compare numbers or read negation, and lines move answers by being there. Measured, with what we check in code instead.

https://jepela.com/blog/jepela-is-open.html2026-10-15T00:00:00Z

Since 2026-10-15 businesses can sign up for Jepela, a decision API with memory: typed answers with probabilities, and output is free.

https://jepela.com/blog/forgetting-with-proof.html2026-10-15T00:00:00Z

Jepela's forget removes a fact from a subject's memory, then checks that recall no longer finds it. What it removes, what stays, and a measured run.

https://jepela.com/changelog.html#2026-10-15-accounts-open-with-card-top-ups2026-10-15T00:00:00Z

Business customers can open an account at app.jepela.com. Jepela approves each new account. Prepaid credit can be added by card, on the payment provider's checkout page where VAT is added, or by invoice. The price is unchanged: $0.042 per million input tokens, output free. See Pricing.

https://jepela.com/changelog.html#2026-09-29-calibration-for-yes-no-questions2026-09-29T00:00:00Z

Calibration fitted on your own labels now also sets a shift for yes/no questions, for a question the engine leans towards yes or no on. Measured on public benchmark yes/no questions: fitted on 593, the other 461 went from 59.2% to 70.1% right. Details in Quality.

https://jepela.com/changelog.html#2026-09-28-public-with-live-demos2026-09-28T00:00:00Z

jepela.com, the documentation and the live demos are open to everyone. A try-it box on the home page asks the live engine three questions, with and without one account's memory. Jepela Arcade: seven games where two players make the same decisions and only one has a Jepela memory. Station 7, a first-person game, and a gallery of twelve small apps that decide with memory.

https://jepela.com/changelog.html#2026-09-27-matching-for-choices-with-many-options2026-09-27T00:00:00Z

A choice with more than 20 options is matched by vectors by default. "method": "engine" still reads every option. Measured on 600 bank messages with 77 intents: 65.2% right, against 62.2% for the engine reading all 77. A matched choice is billed for the state and memory it read, not its options: 23 tokens against 812 on that test. Details in Choice.

https://jepela.com/changelog.html#2026-09-26-memory-fine-tuning-and-shared-passes2026-09-26T00:00:00Z

Recall also searches the subject's names and aliases, and the question's own words. Pinned lines go first at every decision about a subject, and exclusion rules remove options in code. Fine-tuning: a model of your own from your labelled cases, used only when it beats the standard engine on cases it never saw and still listens to memory. Requests that arrive together share one pass on the engine.

https://jepela.com/changelog.html#2026-09-24-the-jepela-api-goes-live2026-09-24T00:00:00Z

The hosted API at api.jepela.com, in the United States region. Keys, prepaid credit with a ledger, and a memory per subject. Three question types: choice, score and noul (yes or no), each with probabilities and a confidence. $0.042 per million input tokens. Output and memory writes are free.