Drift Detection

Drift for traditional ML models in the model registry: the drift overview, prediction records, baselines, analyses, their results and rows, and each model's drift settings. An analysis compares a model's active version's production predictions with that version's active baseline.

Baselines are built, and actuals recorded, through the Model Registry API. All endpoints require authentication via the X-API-Key header and the scope shown. List endpoints are paged (see Pagination): limit (default 50, max 200), cursor, sort (a field, - for descending) and q. They return the items in data with total, limit, cursor, hasMore and requestId in meta.

---

Prediction Record Object

Every prediction a registry model makes is recorded once, by whichever caller made it.

{
  "predictionId": "pred_7f3a",
  "entityId": "customer_456",
  "modelId": "model_xyz",
  "modelVersion": 2,
  "source": "production",
  "abTestId": "ab_abc123",
  "variantId": "challenger",
  "routingReason": "weighted_random",
  "features": { "age": 42, "plan": "pro" },
  "prediction": "churn",
  "probabilities": [0.12, 0.88],
  "confidence": 0.88,
  "actual": "churn",
  "success": true,
  "latencyMs": 23,
  "timestamp": "2026-09-27T10:30:00Z"
}

Field	Description
source	production or test (the model page's Test tab; never counted in results)
abTestId, variantId, routingReason	Set when an A/B test routed the prediction
features, output	Present only while the model's Record inputs setting is on
confidence	The highest class probability (classifiers)
actual	The prediction's actual, when one has been recorded

Drift Result Object

{
  "_id": "6ab9577009d39c66afbbcba6",
  "modelId": "model_xyz",
  "baselineId": "6ab8e1c0...",
  "jobId": "6ab95740...",
  "windowStart": "2026-09-20T00:00:00Z",
  "windowEnd": "2026-09-27T00:00:00Z",
  "sampleSize": 1250,
  "overallStatus": "warning",
  "driftScore": 0.14,
  "featuresAnalyzed": 12,
  "featuresWithDrift": 3,
  "predictionDrift": {
    "metric": "accuracy",
    "higherIsBetter": true,
    "current": 0.86,
    "baseline": 0.91,
    "change": -0.05,
    "drop": 0.055,
    "status": "warning",
    "labelledCount": 420
  },
  "algorithmsRun": ["psi", "ks", "chi_square", "domain_classifier"],
  "algorithmsSkipped": [{ "name": "cbpe", "reason": "Estimating accuracy needs the baseline's calibration ..." }],
  "calculatedAt": "2026-09-27T00:05:12Z",
  "calculationDurationMs": 5400,
  "notification": {
    "state": "sent",
    "sentAt": "2026-09-27T00:05:14Z",
    "inApp": { "sent": ["user_123"] },
    "email": { "sent": ["ops@example.com"] },
    "slack": { "error": "Not sent: no Slack webhook is set" }
  }
}

Field	Description
overallStatus	ok, warning, alert, or error when nothing could be scored. It comes from driftScore against the PSI thresholds (without a score, the worst status any algorithm gave), raised to predictionDrift.status when that is worse.
driftScore	The mean PSI across features: a plain number, not a percentage. null when PSI scored no feature.
featuresAnalyzed, featuresWithDrift	Features some algorithm scored, and those with drift. The rows themselves are GET /mlops/models/:modelId/drift-results/:id/features.
predictionDrift	The model's performance on the window's predictions that have actuals, against the baseline's, scored by the model's task: metric is accuracy for a classification or multiclass model, mae (mean absolute error; higherIsBetter false) for a regression model, and review_score (the mean of the reviewers' verdicts recorded as actuals: true 1, false 0, or a 0 to 1 score) for a multilabel, timeseries or other model. current, baseline and change are on that metric; drop is how much worse than the baseline, relative (0.1 is 10% worse, negative is better); status is ok, warning or alert from the model's performanceDropWarning / performanceDropAlert. When it could not be measured, error says why (no actuals yet, a baseline without performance, or a task changed since the baseline was built) and there are no figures or status, except that against a baseline without performance current (with labelledCount) is still measured and only the comparison is missing.
notification	Only on a result that found drift (warning or alert): how it was announced. state is pending, sending, sent or failed; inApp, email and slack each have sent (user ids, addresses, the webhook host) or error (why not). See the notifications setting below.

Result Row Object

A result has one row per feature, carrying every per-feature algorithm's metric, and one row per algorithm about the whole dataset.

{
  "_id": "6ab95771...",
  "resultId": "6ab9577009d39c66afbbcba6",
  "modelId": "model_xyz",
  "order": 1,
  "featureName": "income",
  "category": "feature",
  "driftType": "data",
  "metrics": {
    "psi": { "value": 0.41, "threshold": 0.2, "status": "alert", "category": "univariate" },
    "ks": { "value": 0.32, "pValue": 0.001, "threshold": 0.2, "status": "alert", "category": "univariate" }
  },
  "hasDrift": true,
  "status": "alert",
  "severity": 3,
  "score": 0.41
}

Field	Description
category	feature for a feature's row; multivariate or performance for a dataset row (named like [multivariate] domain_classifier)
metrics	By algorithm: value (null when it could not run), threshold, status, pValue, and the algorithm's category
status, severity	The row's worst metric status, and its severity: alert 3, warning 2, ok 1, error 0
score	PSI when it ran, else the first scored metric; null when none scored
error	Why nothing on the row could be scored

---

GET /api/v1/mlops/drift

The drift overview: the registry models you can see, with their latest result's summary and their latest analysis run.

Scope: mlops:read

Parameter	Type	Required	Description
status	string	No	ok, warning, alert, error, or no-data for models never analyzed
q	string	No	Name, description or framework
sort	string	No	-latestDrift.calculatedAt (default), name, latestDrift.driftScore, latestDrift.sampleSize, updatedAt

Response 200 OK

{
  "data": [
    {
      "_id": "model_xyz",
      "name": "churn-classifier",
      "framework": "sklearn",
      "latestDrift": { "resultId": "6ab95770...", "overallStatus": "warning", "driftScore": 0.14, "sampleSize": 1250, "calculatedAt": "2026-09-27T00:05:12Z" },
      "latestAnalysis": {
        "jobId": "6ab95740...",
        "status": "failed",
        "triggeredBy": "scheduler",
        "error": "Version 2 has no production predictions from 2026-09-20 00:00 to 2026-09-27 00:00 UTC to compare with its baseline (Test tab predictions are not counted)",
        "startedAt": "2026-09-27T00:00:03Z",
        "updatedAt": "2026-09-27T00:00:41Z",
        "finishedAt": "2026-09-27T00:00:41Z"
      },
      "driftReadiness": { "state": "analyzed" }
    }
  ],
  "meta": {
    "total": 1,
    "limit": 50,
    "nextCursor": null,
    "requestId": "req_abc123"
  }
}

latestAnalysis.status is pending, deploying, running, completed or failed; triggeredBy is manual, scheduler or baseline. The model record (GET /mlops/models/:id) carries the same two fields.

driftReadiness.state is analyzed when the model has a result, else what it still needs: no-task (the model has no classification or regression task; set problemType with PATCH /api/v1/mlops/models/:id), inputs-off (Record inputs is off, so its predictions have no inputs to compare), no-baseline (its active version, given as version, has no baseline) or not-analyzed (it has a baseline and no result yet; latestAnalysis says why a run failed).

---

POST /api/v1/mlops/models/:id/predictions

Record a prediction a registry model made outside the platform (a copy served elsewhere), so drift includes it. Predictions the platform serves are recorded automatically.

Scope: mlops:write

Field	Type	Required	Description
:id (path)	string	Yes	Registry model ID
features	object	Yes	Input values by feature name
prediction	any	Yes	The model's output
probabilities	array	Classification: Yes	Per-class probabilities; the highest is the confidence
entityId	string	No	Your id for what was predicted; actuals can join on it
modelVersion	number	No	The version that made the prediction
latencyMs	number	No	How long it took

Response 201 Created

The prediction record as stored, source production.

---

GET /api/v1/mlops/models/:id/predictions

A model's prediction records, each with its actual when one was recorded.

Scope: mlops:read

Parameter	Type	Required	Description
:id (path)	string	Yes	Registry model ID
source	string	No	production or test
startDate, endDate	string	No	ISO 8601 bounds on the prediction time
q	string	No	Prediction id, entity id or error, or the prediction exactly
sort	string	No	-timestamp (default), latencyMs, modelVersion, success, confidence

---

GET /api/v1/mlops/models/:id/baselines

A model's baselines, newest first, each with its inputs' names (not their distributions).

Scope: mlops:read

Parameter	Type	Required	Description
:id (path)	string	Yes	Registry model ID
sort	string	No	-createdAt (default), modelVersion, sampleSize, baselinePerformance.value

Each item: modelVersion, sampleSize (rows), evaluationKind (classification, regression or reviewed: how the model was scored when the baseline was built), labelledRows, baselinePerformance ({ metric, higherIsBetter, value }, the version's performance on its reference rows; see Performance), columns (which of actual, prediction, confidence the rows had), skippedFeatures (with why), source (the file), isActive, supersededAt, createdAt, createdBy.

---

GET /api/v1/mlops/models/:id/baselines/active

The active baseline of the model's active version, with its features' distributions and calibration. null when the version has no baseline.

Scope: mlops:read

Parameter	Type	Required	Description
:id (path)	string	Yes	Registry model ID

---

POST /api/v1/mlops/models/:id/drift-analyses

Start a drift analysis: the active version's production predictions in the window against its active baseline, with the algorithms enabled in the model's drift settings, plus prediction drift from actuals.

Scope: mlops:write

Field	Type	Required	Description
:id (path)	string	Yes	Registry model ID
windowDays	number	No	Days of predictions to analyze (default 7)
windowStart, windowEnd	string	No	An explicit window (ISO 8601)

Response 200 OK

{
  "data": { "jobId": "6ab95740...", "status": "deploying" },
  "meta": { "requestId": "req_abc123" }
}

A version without a baseline is refused (400, saying to build one). A window with fewer predictions (with recorded inputs) than the model's minWindowSampleSize is not scored: the run fails saying so. Follow the run with the model's latestAnalysis or the status endpoint below.

---

GET /api/v1/mlops/models/:modelId/drift-analyses/:id

One analysis run by the jobId it returned: status (pending, deploying, running, completed, failed), its window, and error when it failed.

Scope: mlops:read

Parameter	Type	Required	Description
:modelId (path)	string	Yes	Registry model ID

---

GET /api/v1/mlops/models/:id/drift-results/latest

A model's most recent drift result. null when it has never been analyzed.

Scope: mlops:read

Parameter	Type	Required	Description
:id (path)	string	Yes	Registry model ID

---

GET /api/v1/mlops/models/:id/drift-results

A model's drift results, newest first.

Scope: mlops:read

Parameter	Type	Required	Description
:id (path)	string	Yes	Registry model ID
status	string	No	ok, warning, alert or error
sort	string	No	-calculatedAt (default), driftScore, sampleSize

---

GET /api/v1/mlops/models/:id/drift-analyses

A model's analysis runs, newest first: its history, including runs that could not score anything. Each run has status (pending, running, completed or failed), triggeredBy (manual or scheduler), created_at and duration_ms. A completed run has its result_id and a summary (overallStatus, driftScore, featuresAnalyzed, featuresWithDrift, sampleSize); a failed run has its error, for example no production predictions in its window.

Scope: mlops:read

Parameter	Type	Required	Description
:id (path)	string	Yes	Registry model ID
q	string	No	Status, trigger or error text
sort	string	No	-created_at (default), duration_ms

---

GET /api/v1/mlops/models/:modelId/drift-results/:id/features

A result's rows, most drifted first. Get the resultId from a result's _id or a model's latestDrift.resultId.

Scope: mlops:read

Parameter	Type	Required	Description
:modelId (path)	string	Yes	Registry model ID
rows	string	No	features (feature rows only) or dataset (dataset rows only)
q	string	No	Feature name
sort	string	No	-severity (default), score, featureName, order (the result's own order)

---

GET /api/v1/mlops/models/:id/drift-settings

A model's drift settings. null when they were never saved.

Scope: mlops:read

Parameter	Type	Required	Description
:id (path)	string	Yes	Registry model ID

{
  "data": {
    "modelId": "model_xyz",
    "enabled": true,
    "algorithms": {
      "enabled": ["psi", "ks", "chi_square", "jensen_shannon", "wasserstein", "domain_classifier", "pca_reconstruction", "adwin", "cbpe"],
      "overrides": { "psi": { "warningThreshold": 0.1, "alertThreshold": 0.2 } }
    },
    "performanceDropWarning": 0.05,
    "performanceDropAlert": 0.1,
    "minBaselineSampleSize": 30,
    "minWindowSampleSize": 30,
    "notifications": { "email": true, "recipients": ["ops@example.com"], "slack": "https://hooks.slack.com/services/..." },
    "schedule": {
      "enabled": true,
      "frequency": "daily",
      "timeWindowDays": 7,
      "lastRunAt": "2026-09-27T00:00:03Z",
      "nextRunAt": "2026-09-28T00:00:03Z",
      "lastJobId": "6ab95740..."
    }
  },
  "meta": { "requestId": "req_abc123" }
}

---

PATCH /api/v1/mlops/models/:id/drift-settings

Save a model's drift settings. The first save stores defaults for the fields not sent; later saves change only the fields sent.

Scope: mlops:write

Field	Type	Required	Description
:id (path)	string	Yes	Registry model ID
enabled	boolean	No	Drift settings on; required for the schedule to run
algorithms	object	No	{ enabled: [names], overrides: { [name]: { warningThreshold, alertThreshold, params } } }. Built-in names: psi, ks, chi_square, jensen_shannon, wasserstein, domain_classifier, pca_reconstruction, adwin, cbpe; or a custom algorithm's name
performanceDropWarning, performanceDropAlert	number	No	How much worse than the baseline the model's performance may get before prediction drift is a warning (default 0.05) or an alert (default 0.10), relative: 0.05 is 5% worse (lower accuracy or review score, higher mean absolute error). The warning must be smaller than the alert. A warning or alert makes the analysis's overall status at least that high, so it is announced.
minBaselineSampleSize	number	No	Fewest reference rows a baseline needs (default 30)
minWindowSampleSize	number	No	Fewest predictions with recorded inputs an analysis window needs to be scored: a whole number, at least 1 (default 30)
notifications	object	No	{ email: boolean, recipients: [addresses], slack?: Slack incoming webhook URL }. When an analysis finds drift (warning or alert), the model's owner is notified in Strongly; with email, the recipients are emailed (the owner when there are none); with slack, that webhook gets a message. An address that is not one, or a webhook pointing at localhost, an internal host or a private address, is refused with 400. An analysis that could not run sends no notification.
schedule	object	No	{ enabled, frequency: "hourly" | "daily" | "weekly", timeWindowDays }

A scheduled analysis runs while both enabled and schedule.enabled are true, the first run at the scheduler's next check. Building a baseline with analyzeDaily turns on a daily schedule the same way.

Response 200 OK

The model's drift settings, as GET /api/v1/mlops/models/:id/drift-settings shows them.