THREE APPLICATIONS · ONE RISK FRAMEWORK
Estimate mistake probabilities from reviewed tasks, then apply costs suited to each application.
Wrong traveller, stale price, duplicate booking, missed deadline or nonrefundable itinerary.
Costs use synthetic cancellation and rebooking penalties. Haiku and Sonnet receive the same trips, with memory and review on or off.
Explore costly proposal risk →Identifier or credential disclosure, cross-user records, hallucinated facts, unreliable citations or unsupported assurances.
Costs cover response expenses, staff time and an optional impact valuation. All test records and identifiers are synthetic.
Explore privacy & knowledge risk →Cross-system customer identity, ticket eligibility, credit arithmetic and citation contracts.
Compare Haiku and Sonnet on Era company records, with scoped records, SurrealDB guidance and critic review.
Explore the enterprise experiment →| Measure | Travel | Privacy & knowledge |
|---|---|---|
| Observed frequency | Policy violations in completed synthetic trips | Known disclosure or unsupported-claim classes in fictional answers |
| Severity | Contract-based synthetic dollar exposure | Hypothetical planning ranges, with optional impact valuation |
| Expected cost | Mean potential loss before the execution gate | Σ class probability × assumed conditional severity |
| Tail risk | P(synthetic loss > chosen threshold) | Envelope for P(cost proxy > threshold), conditional on severity assumptions |
| Uncertainty | Working-model intervals and simultaneous fixed-batch bounds | Dirichlet working-model envelopes and joint class-probability bounds |
Each application has its own tasks and cost model. To combine them, specify traffic shares, a common currency and time horizon, and comparable cost inputs. Count a shared incident once.