ZeroGrad
Capabilities · Anomaly Detection

Transaction Anomaly Detection

Healthy
Zero-shot detectionEndpoint ledger-anomaliesProduction v2Challenger v3

Real-time production health

Known immediately — no outcomes needed.

Last 24 hours
Events scored
2.6M
Availability
99.999%
P50 latency
6ms
P95 latency
14ms
Error rate
0.00%

Predictions / hour

66K106K146K00:0004:0008:0012:0016:0020:00

Latency

0ms8ms16ms00:0004:0008:0012:0016:0020:00
P50P95

Inputs, context & predictions

Schema
Healthy
Event volume
Normal
Missing values
Normal
New entities
86

Input drift vs reference (PSI)

0.1 · 0.25
event_type
0.03
amount
0.03
merchant_id
0.02
hour_of_day
0.01

Prediction distribution

today vs reference · no shift
0Anomaly score1
Live Reference examples

Routing

Zero-shot detection
v2 · Managed A100%
Backend
Managed by Zero Grad
Context
pinned to v2
Fallback on timeout
Managed C · 0.02% of requests

Matured outcome quality

Performance metrics are computed only for predictions with observed outcomes. Analyst-confirmed incidents take up to ~2 days to arrive, so recent predictions aren't scored yet.

Evaluation window
Sep 1 – Sep 30
predictions made in this period
Outcome coverage
89%
have an observed outcome
Median outcome delay
6 hours
analyst-confirmed incidents
Last outcome received
Oct 1, 08:55
via outcomes API

Business impact

Sep 1 – Sep 30 · 89% coverage
Incidents caught (30d)
4
Exposure flagged
$58,400
On-call pages avoided
~120
vs. static thresholds

Exposure is the total value of events in confirmed incidents at time of alert.

Model quality

Sep 1 – Sep 30
Alert precision
84%
Alerts acknowledged
35 / 37
Median time to detect
2.4 min
Alert volume
Normal

Alert precision — matured outcomes

By prediction day. Recent cohorts are provisional until their outcomes arrive.

83.4%84.0%84.6%Sep 17Sep 19Sep 21Sep 23Sep 25Sep 27Sep 29
Alert precision · maturedProvisional (< 60% coverage)
Outcome coverage by cohorttoday →

Outcomes

Ground truth flows back by prediction_id. It improves evaluation and routing — never model weights.

Send outcomes

Every prediction returns a prediction_id. Send the real outcome whenever you learn it.

# Later, when the real outcome is known
zg.capability("ledger-anomalies").send_outcomes([
    {"prediction_id": "pred_01J9X4T5WE", "outcome": {"incident": True},
     "observed_at": "2026-09-30T23:41:00Z"},
])

Outcomes never update model weights — Zero Grad does not train a customer-specific model. They are used to measure real production quality, compare foundation models and context-selection strategies, improve routing, and decide whether a challenger beats the current production configuration.

What Zero Grad learned from your outcomes

Simulated

No training doesn't mean no learning.

  1. Analyst feedback: 38% of weekend payout alerts were false positives → a weekend baseline was added to context.
  2. TabFM with labeled incidents in context raises alert precision from 84% to 88% (v3 challenger).
  3. Both missed incidents were slow drifts → foundation-model residuals are now weighted higher.

What adapts: which foundation model serves which slice, how in-context examples are selected, and calibration. What doesn't: model weights.