Sign in required

Scoped to experiment —. One line per MODEL_VERSION — every getnet_psp_router:* release plus the heuristic-v0 baseline. Shadow-mode predictions ARE included so we get signal during the still-in-shadow rollout — the model line reflects what the router would have chosen if it had routed live.

Same experiment, split by the downstream PSP the router selected (PREDICTED_VALUE in ATHIA_FEEDBACK). Bars = volume, lines = approval rate. One card per processor discovered in the last 14 days — no hardcoded processor list.

Shadow-mode traffic is excluded. Charts count only predictions where response_payload.is_shadow_mode = false — i.e. decisions that actually influenced routing. Experiment —. Bars = daily volume (right axis). Lines = approval rate (left axis, 0–100%).

Same experiment, denied ÷ total. Bars = daily volume. Lines = decline rate (0–100%).

For each Getnet model, correlate its declared feature set against APPROVED over the last 30 days of PAYMENT_ML.ABTESTING.GETNET_PROSA_MCO_TRAINING_DATA. Numerics use Pearson r², categoricals use η² — both [0,1], higher = stronger signal. Feature lists come from data/getnet_*_features.txt. Refreshed every 15 min.

Single-line daily approval rate across all Getnet merchants in PAYMENT_ML.ABTESTING.GETNET_PROSA_MCO_TRAINING_DATA. Sanity check for platform-wide trend and outages.

One line per Getnet merchant that had at least 100 transactions in the last 14 days. Same series as the aggregate chart above, just decomposed.

Per-training-run quality metrics for the getnet_psp_router models (dnn, gbdt, logreg) across the three downstream PSPs (cybersource, getnet, prosa). Source: —. Pick a metric to compare all archs side-by-side per PSP — click a legend entry to hide/show that arch. Line = per-training-run metric, faded bars = training-set size (right axis) so you can eyeball whether a spike is a real improvement or a small-sample fluke. Use the Version dropdown to pin the score cards, Radar, and Best-Arch table to a specific version across all 3 architectures (default = latest).

Cross-PSP volume + approval + decline-code breakdown across the 8-merchant GetNet cohort. Fed by an hourly upstream cron in DATA-Athena-Snowflake that writes CSVs the portal syncs each refresh. Full drill-down + tables: Explore PSP Analysis →