Feedback Loop API

The feedbackLoop router manages the online learning feedback loop that continuously improves scoring accuracy. It provides holdout testing configuration, loop health monitoring, lift measurement, and the core conversion event ingestion endpoint.

Router namespace: feedbackLoop

Source: src/server/trpc/routers/feedback-loop.ts

Procedures

Procedure	Type	Access	Description
getHoldoutConfig	query	admin	Get holdout test configuration
updateHoldoutConfig	mutation	admin	Update holdout percentage and rotation
rotateHoldoutSeed	mutation	admin	Rotate the holdout seed
getLoopStatus	query	admin	Get comprehensive feedback loop status
getHoldoutEvaluation	query	admin	Measure holdout lift
processConversionEvent	mutation	tenant	Record a conversion signal

---

getHoldoutConfig

Returns the current holdout test configuration. The holdout group is a random subset of entities excluded from ML-boosted scoring, used to measure the model's incremental lift.

feedbackLoop.getHoldoutConfig.queryOptions()

Access: Admin only

Input

None.

Response

{
  "enabled": true,
  "percentage": 10,
  "seed": 42,
  "rotationDays": 90
}

Field	Type	Description
enabled	boolean	Whether holdout testing is active
percentage	number (1--20)	Percentage of entities in holdout group
seed	number	Random seed for consistent group assignment
rotationDays	number (30--365)	Days between automatic seed rotations

Example

const trpc = useTRPC();
const { data: config } = useQuery(
  trpc.feedbackLoop.getHoldoutConfig.queryOptions()
);

---

updateHoldoutConfig

Updates the holdout test configuration. Changes take effect on the next scoring run.

feedbackLoop.updateHoldoutConfig.mutationOptions(options)

Access: Admin only

Input

Field	Type	Required	Description
percentage	number (1--20)	No	Holdout group percentage
rotationDays	number (30--365)	No	Days between seed rotations

Response

{
  "success": true
}

Example

const trpc = useTRPC();

const mutation = useMutation(
  trpc.feedbackLoop.updateHoldoutConfig.mutationOptions({})
);

mutation.mutate({ percentage: 15, rotationDays: 60 });

---

rotateHoldoutSeed

Generates a new random seed for holdout group assignment. This clears all current holdout assignments, causing entities to be randomly reassigned on the next scoring run.

feedbackLoop.rotateHoldoutSeed.mutationOptions(options)

Access: Admin only

Input

None.

Response

{
  "success": true,
  "newSeed": 8472
}

Example

const trpc = useTRPC();
const queryClient = useQueryClient();

const mutation = useMutation(
  trpc.feedbackLoop.rotateHoldoutSeed.mutationOptions({
    onSuccess: () => {
      queryClient.invalidateQueries({
        queryKey: trpc.feedbackLoop.getHoldoutConfig.queryKey(),
      });
    },
  })
);

mutation.mutate();

warning

Rotating the seed reassigns all entities to new holdout/treatment groups. This resets any in-progress lift measurement. Only rotate when you have sufficient data from the current period.

---

getLoopStatus

Returns a comprehensive status overview of the entire feedback loop, including online learning state, holdout configuration, drift detection, and push status.

feedbackLoop.getLoopStatus.queryOptions()

Access: Admin only

Input

None.

Response

{
  "onlineLearning": {
    "enabled": true,
    "eventsProcessed": 1247,
    "lastEventAt": "2026-03-20T15:45:00.000Z"
  },
  "holdout": {
    "enabled": true,
    "percentage": 10,
    "rotationDays": 90,
    "nextRotation": "2026-06-01T00:00:00.000Z"
  },
  "drift": {
    "isDrifted": false,
    "lastChecked": "2026-03-20T14:30:00.000Z"
  },
  "push": {
    "lastPushAt": "2026-03-20T12:00:00.000Z",
    "modelVersion": "v3.2"
  }
}

Example

const trpc = useTRPC();
const { data: status } = useQuery(
  trpc.feedbackLoop.getLoopStatus.queryOptions()
);

---

getHoldoutEvaluation

Measures the incremental lift of the ML model by comparing conversion rates between the holdout group (no ML scoring) and the treatment group (ML-boosted scoring).

feedbackLoop.getHoldoutEvaluation.queryOptions(input)

Access: Admin only

Input

Field	Type	Required	Default	Description
days	number (1--365)	No	30	Evaluation window in days

Response

{
  "treatmentConversionRate": 0.12,
  "holdoutConversionRate": 0.08,
  "lift": 0.50,
  "treatmentSize": 900,
  "holdoutSize": 100,
  "days": 30,
  "statisticallySignificant": true
}

Field	Description
lift	Relative lift: (treatment - holdout) / holdout
statisticallySignificant	Whether the sample size is sufficient for confidence

Example

const trpc = useTRPC();
const { data: evaluation } = useQuery(
  trpc.feedbackLoop.getHoldoutEvaluation.queryOptions({ days: 90 })
);

if (evaluation) {
  console.log(`ML model lift: ${(evaluation.lift * 100).toFixed(1)}%`);
}

---

processConversionEvent

Records a conversion signal for an entity. This is the core feedback mechanism that drives online learning. Events are forwarded to the ML sidecar for incremental model updates.

feedbackLoop.processConversionEvent.mutationOptions(options)

Input

Field	Type	Required	Description
entityId	string	Yes	Entity identifier (person, account, or opportunity)
converted	boolean	Yes	Whether the entity converted

Response

{
  "success": true
}

Example

const trpc = useTRPC();

const mutation = useMutation(
  trpc.feedbackLoop.processConversionEvent.mutationOptions({})
);

// Record a conversion
mutation.mutate({ entityId: "opp_abc123", converted: true });

// Record a non-conversion (lost deal)
mutation.mutate({ entityId: "opp_def456", converted: false });

Automated Ingestion

In most setups, conversion events are ingested automatically via Salesforce sync (closed-won/closed-lost). Use this procedure for manual feedback or custom conversion definitions.

Related Pages

• Drift API -- Drift detection powered by feedback data
• Scoring API -- Scores improved by the feedback loop
• Audit API -- Holdout config changes in audit trail