Three calls, and you have a verdict
Kirelta takes rows of numbers — the same feature vectors you feed your model — and tells you whether they still look like the data the model was trained on. Everything below is copy-paste ready.
Train a baseline
Send a batch of rows that represent healthy behaviour — data from a period when the model was working well. Kirelta learns the shape of that data and picks its own monitor for it. Every row must have the same number of features.
Returns
Assess a batch
Now send recent rows — the ones you actually want a verdict on. Kirelta compares them against the baseline and answers in one object.
Read the verdict
This is the whole product. Everything the console shows you is built from these fields — nothing else is computed anywhere.
| verdict | TRUSTED · DEGRADED · UNTRUSTED — the call itself. Branch on this. |
| action | The engine's recommendation in words, e.g. ALLOW, REVIEW (elevated abstention), BLOCK / escalate to a human. |
| flagged_rate | Fraction of rows in this batch that fell outside the trained range. 0.0–1.0. |
| drift_alarmed | true when the sequential test fires. This is the anytime-valid alarm — its false-alarm rate is bounded over the whole run, not per batch. |
| n | How many rows were assessed. |
| top_features | Column indices that contributed most to the deviation, most unusual first. |
| feature_names (send on /fit) | Optional, and worth more than it costs. Pass your column names as a list, in the same order as your rows, and every piece of evidence comes back named: {"feature": 2, "name": "cpu_load", "importance": 0.529}. Without it the engine can only say #2 — a fact, not an explanation. A list of the wrong length is rejected rather than applied, because a mislabelled feature is worse than an unlabelled one. The Python SDK reads them from a pandas DataFrame automatically. |
| persisted (on /fit) | Whether this model reached disk. true means it survives a restart; false means it is live in memory only and a persistence_note explains why (no volume configured, or the write failed). A model that says fitted but persisted:false is one redeploy from gone. The /health endpoint reports the same fact service-wide as durable. |
| output_stability | The second layer — the model's own behaviour. Send outputs (the predictions your model made) to /fit and /assess and Kirelta also tests whether the model still behaves like itself, via a permutation test on the Jensen–Shannon divergence. Always present, including when it is switched off — a dark layer reports active: false and says why, because "no drift" from a layer that never ran means nothing. The combination that matters: verdict: TRUSTED with output_stability.drift: true is the silent failure — nothing upstream changed and the model quietly stopped behaving like itself. Tools that watch only inputs miss it every time. |
| martingale alarm_threshold | How close this batch came to the alarm. The martingale is accumulated evidence against "no drift"; the alarm fires the moment it reaches alarm_threshold (= 1/α). A batch at 9 / 20 is not alarmed and is not fine — it is nearly halfway to a block. Every other tool gives you only the binary. Evidence accumulates over the rows of a batch and resets on the next call, so read it per batch. |
| drift_started_at | Index of the row where drift was declared. Absent when nothing alarmed. Evidence has to accumulate before the alarm can fire, so this lands a few rows after the true onset (measured: ~4 rows past a hard mid-batch shift) — read it as "no later than here", not as the exact first bad row. |
| feature_importance | The real share of the deviation each feature carries, summing to 1.0 — not just an ordering. top_features is kept for compatibility. |
| notes | The engine's own warnings. The one worth reading: "baseline looks contaminated" means the data you trained on may itself contain anomalies — which changes how you should read every verdict that comes from it. |
| slow_drift_rate | Present only when the baseline itself is aging. Catches the slow drift that never trips a single batch. |
| temporal.ac1 | Lag-1 autocorrelation measured in this batch. The drift alarm assumes your rows are exchangeable; when ac1 climbs above 0.3 that assumption weakens and the alarm becomes less reliable — the per-row flags stay valid. Kirelta measures this and tells you, rather than letting you over-trust the alarm. |
| trend | Present only when applicable: {"state":"recovering","window":8} means this model was alarmed recently and is now improving. It never replaces verdict — read both. |
Wire it into your service
The usual shape: assess the batch before you act on the predictions, and branch on the verdict.
Everything else
The rest of the surface, in one place.
Every model in your tenant, and how many features each expects.
Which monitor Kirelta chose for this model and why — useful when you want to justify a verdict.
Your plan, checks used today, models against your limit.
Create and revoke API keys. Session auth (Authorization: Bearer), not key auth — or just use the account page.
Download everything held about your account, or erase it completely.
Trust events — what changed. Kirelta POSTs an event to your webhook when a model's
verdict changes, not on every reading: a model assessed a thousand times while broken
sends one event, not a thousand. An escalation from degraded to untrusted always gets through,
cooldown or not — hearing "degraded" and then silence is the most dangerous thing this system
could do to you. Kirelta notifies; it never acts on your systems.
It only sees what you send it. There is no scheduler. If your service stops
calling /assess, Kirelta goes quiet, and no event will tell you that
you have gone blind — keep that check in your own uptime monitoring.
No auth. Returns {"status":"ok"} when the engine is up.
When something fails
| 400 | The body wasn't what the endpoint expected. The message says what was wrong. |
| 401 | Missing or revoked API key. |
| 404 | That model hasn't been fitted in your tenant yet. |
| 429 | Rate limited. Back off and retry. |
A model exists only after /fit. If you get a 404 on /assess, fit the baseline first.