Model Cards

A card for every AI feature that touches money

Each card states the intended use, the training data, the headline metric against a baseline, the human-in-the-loop boundary and the things the model is explicitly not for. Reviewed quarterly and signed by the model owner.

Live models

Pacing, anomaly, forecast, FX, coach

Review cadence

Quarterly

Signed by owner + CISO

Auto-execute share

< 38%

Rest gated by materiality

Last full audit

2026-Q2

External: Trail of Bits

pacer-v7

Pacer

Probabilistic intra-day budget pacing across ad platforms

v7.4.2

Inputs: Hourly spend, platform delivery curves, calendar features, FX
Outputs: P10 / P50 / P90 end-of-day spend, suggested cap delta
Training data: 8.4B impressions, 19 verticals, 2022-01 → 2026-Q1 (rolling)
Headline metric: MAPE (P50, daily): 3.1% (baseline Linear pacing: 11.4%)
Oversight: Auto-execute below materiality; human approval ≥ €25k delta
Not used for: Creative selection, audience targeting, bid manipulation
Last reviewed: 2026-06-01

sentinel-v4

Sentinel

Anomaly detection on platform spend & reconciliation streams

v4.1.0

Inputs: Per-account spend deltas, latency, FX, invoice deltas
Outputs: Anomaly score 0–1, suggested action, evidence trace
Training data: 11.2M labelled incidents, 2023-06 → 2026-Q1
Headline metric: Precision @ 0.9 recall: 0.91 (baseline Rules-only: 0.62)
Oversight: Notify-only by default; auto-pause requires policy opt-in
Not used for: User-level scoring, fraud attribution to individuals
Last reviewed: 2026-05-22

forecast-v3

Forecast

Multi-horizon spend & cash forecast (1d / 7d / 30d / 90d)

v3.8.1

Inputs: Historical spend, contractual commitments, seasonality, FX forwards
Outputs: P10/P50/P90 spend & cash bands, calibration report
Training data: 1,180 customers · 4.1 years of spend ledgers
Headline metric: CRPS (30d horizon): 0.054 (baseline Naïve seasonal: 0.121)
Oversight: Read-only; never executes payouts on its own
Not used for: Revenue or LTV forecasting, board guidance
Last reviewed: 2026-04-30

fx-leak-v2

FX Leak

Detection of FX spread anomalies in platform invoices

v2.3.0

Inputs: Daily mid-market FX, platform billing FX, invoice line items
Outputs: Per-invoice basis-point leakage, monthly attribution
Training data: 2.3M invoice lines · 31 currency pairs
Headline metric: Coverage of ≥ 5bps leakage events: 98.4% (baseline Manual sampling: ~22%)
Oversight: Reports only; flagged items routed to controller queue
Not used for: Hedging recommendations, automated FX trading
Last reviewed: 2026-06-10

policy-coach-v1

Policy Coach

Plain-language explanation of policy & guardrail decisions

v1.2.4

Inputs: Policy YAML, evaluation trace, user context
Outputs: Natural-language rationale, suggested remediation
Training data: Internal policy corpus, no customer content
Headline metric: Faithfulness (vs evaluation trace): 0.97 (baseline n/a (new capability))
Oversight: Explanatory only; never grants access or relaxes policy
Not used for: Policy authoring, legal advice, compliance attestation
Last reviewed: 2026-06-18

What changes trigger a new card version

Any change to inputs, outputs, or intended use.
Any retraining on new data, or material drift in the headline metric.
Any change to the human-in-the-loop boundary or materiality threshold.
Any extension into a new vertical, region or currency.