Interactive seeded demo · synthetic fixture data · no upload required
Operate the Temporal Stress Chamber
Set up a scenario and compare its effect on discrimination, calibration and explanation stability.
Open sample project →Uses the same synthetic dataset and includes saved scenarios, a search and an assurance case.
Scenario control deck
Signal substitution + feature staleness · seed 20260903
Share of rows whose live velocity_24h_txn_count input is partially replaced by correlated merchant_risk_score, followed by a genuine model rerun.
Share of rows whose velocity_7d_amount_sum is replaced with a lagged value (feature_staleness operator).
How many rows back the lagged value is drawn from.
Preview: valid over 300 synthetic rows.
signal_substitution: 256 rows · velocity_24h_txn_count
feature_staleness: 108 rows · velocity_7d_amount_sum
Temporal Stress Chamber
How to read this
Left = reference window, right = stressed comparison window. Each channel keeps its own colour throughout, cyan for discrimination, blue for calibration, violet for explanation stability. Only the dashed line at the gate is the injected disruption itself. Vertical deflection and line thickness encode the size of the change in that channel's own terms, not a shared severity scale. Chips above the gate show the actual armed scenario operators.
View as table
| Channel | Metric | Baseline | Stressed | Note |
|---|---|---|---|---|
| Discrimination | PR-AUC | 0.667 | 0.667 | Computed by src/lib/metrics.ts#computePRAUC over the actual scenario-executed rows for the current parameters. |
| Calibration | ECE | 0.037 | 0.037 | Computed by src/lib/metrics.ts#computeECE over the actual scenario-executed rows for the current parameters. |
| Explanation stability | Attribution-profile cosine similarity (proxy) | 1.000 | 1.000 | Computed by src/lib/metrics.ts#explanationStabilityProxy comparing the baseline and scenario-executed attribution profiles. This is a cosine-similarity proxy, not the dissertation's validated ESI. |