The deployment question
RiskGraph measures agent mistakes before deployment. It connects actions, evidence and uncertainty, so teams can compare safeguards without hiding the useful work those safeguards block.

Bring your workflow
Describe agents, handoffs and controls. Preview the setup as a graph; custom execution needs a sandbox adapter. Upload reviewed simulation or real episodes. Sources stay separate, and missing outcomes stay unknown.

Explore failure modes
Test refunds, travel and private information against budgets, stale quotes, conflicting records and protected aliases. Discovery proposes possible failure families. Independently reviewed outcomes become evidence; proposed scenarios alone do not.

Follow the topology
Replay the request, records, proposal, memory, critic and gate. Source-backed memory supplies procedures, never permission. Critics can correct an answer or introduce a mistake. Each saved response remains inspectable.

A real failure pattern
Earlier retrieval masked restricted fields but trusted aliases and citation identifiers. Both models leaked fictional protected values there. The guard also blocked valid answers because stale or unrelated restrictions contained the same number.

The repair
The repair selects current authoritative sources in the exact scope, retains revision ties, applies current denials first, and checks value and citation lineage. Critic and guard use runtime checks. Independent evaluation labels stay separate.

Fresh matched evaluation
The fresh run matched Haiku and Sonnet across 12 tasks: eight required refusals and four valid controls. Settings and output contracts were identical. Actual Sonnet reviews are scored unchanged. These are raw errors with marginal uncertainty intervals.

Estimate probability
Repaired Sonnet with a critic made 0 classified errors in 12 tasks. The Jeffreys posterior mean is 3.8%; its marginal ninety five percent interval reaches 18.5%. Wider simultaneous bounds account for multiple comparisons. Zero observed errors still leaves uncertainty.

From frequency to severity
Combine each mistake probability with an explicit severity range. Expected cost is frequency times severity; tail probability measures losses above a threshold. Cash, response time and intangible impact are editable assumptions, not measured damages.

Risk versus useful work
Compare errors, useful completions and unnecessary refusals. Toggle saved critic and guard outcomes. Matched comparisons count prevented and introduced mistakes, with uncertainty on the difference. A range crossing zero does not establish a benefit.

Inspect the evidence
Inspect sources, retrieval receipts, proposals and critic changes in the action recorder. Uploaded demo traces have the same risk views. Real and simulated evidence keep separate models and declared weights. Unseen failure mass is an explicit sensitivity assumption.

Evidence. Assumptions. Uncertainty.
Where can the workflow fail, how severe could loss be, and which control preserves useful work? RiskGraph makes that decision reviewable. These small fictional stress tests do not guarantee production failure rates. Explore the live demo and reproducible evidence.