No item on this page is aspirational — everything here is running code you can test yourself, right now, with your own data.
What happens to a batch you send, in order. Each stage is a call you make or a record you can read back.
Send rows from a window when the model was healthy. Kirelta learns their shape and keeps a model of it, not your rows.
POST /fitSend recent rows. Every row is held against the baseline, and the batch gets a flagged share and a typicality score.
POST /assessEvidence of drift builds across the batch and is held against a calibrated alarm line, 20 by default, with a false-alarm rate bounded over the whole run.
Trusted, degraded or untrusted with a recommended action, or no conclusion when the data cannot support one.
Every verdict that is not trusted opens a decision. Someone on your team approves or overrides it, with a reason, in the decision journal.
GET /decisionsThe decision keeps its evidence sealed by hash and exports as a PDF; account and key actions are signed at checkpoints you can verify offline.
GET /audit/exportThe alarm threshold carries a real, provable false-alarm bound over the whole run, however long you watch — not a tuned-until-it-looked-right one. The measured rates are published on the calibration page.
Every UNTRUSTED or DEGRADED verdict comes with which specific inputs moved and by how much — not just "something changed."
Input drift, output stability (when you send your model’s own predictions), and baseline-quality suspicion are tracked separately and shown separately — a healthy input layer doesn’t hide a suspect baseline.
Every verdict change writes a permanent, evidence-backed entry — not just a log line. Review, approve, reject, or archive, with a note that’s recorded against your name.
Each decision’s evidence snapshot is sealed with a content hash at write time and re-verified on every read. If the stored record had been altered, the page says so.
Per-model trust-over-time charts: every batch you send, with its verdict, evidence and flagged share.
Key creation, revocation, team changes, sign-ins, and account deletion are all recorded per tenant, append-only.
/fit once, /assess per batch, branch on the verdict. Full reference and copy-paste examples in the docs.
Zero dependencies beyond the standard library: kirelta_client.py for Python, kirelta.mjs for JavaScript. Drop the file into your project.
Get notified the moment a verdict changes, instead of polling.
Open the reference case, or paste an API key and watch a model’s real status update — no account required to look.
Invite teammates by email; seat limits scale with plan.
Account deletion removes the account, its models, its keys, and its decision history. An automated test proves it leaves nothing behind: it finds every table that holds workspace data in the live schema, so a table added later is covered too.