Risk while agents adapt

A normal-mixture Hoeffding confidence sequence bounds the running average of conditional expected loss. Losses may depend on previous decisions, agent interactions, and shared memory. The exposure cap and tuning parameter must be fixed before observation.

Each stream receives 2.5% error, giving at least 95% simultaneous coverage across the before/after pair and all observation times within an assessment. This target describes risk over observed episodes; it does not forecast a future workload.

A separate monetary-loss e-process tests whether conditional mean proposed loss exceeds $10 per episode. The frequency and monetary alarms each have a separate 5% false-alert budget; their union is bounded by 10%.

Time-uniform upper bounds

Four loss thresholds each receive a beta-mixture confidence sequence. Their weighted upper endpoints bound mean loss while allowing repeated observation.

The dashboard allocates 2.5% error to each of the proposed and residual streams. Together, both upper bounds have at least 95% simultaneous coverage within one evaluation under fixed conditional tail probabilities and a $1,000 loss cap.

Frequency × severity

A Beta(0.5, 0.5) frequency prior and a Dirichlet(0.25, 0.25, 0.25, 0.25) severity prior describe mean binned loss. Positive severity bands end at $25, $100, $500, and $1,000.

The 95% credible interval is a posterior statement, not a repeated-sampling guarantee. With sparse data, estimates depend on the prior. Zero observed loss retains nonzero estimated risk.

A mixture of alarms

A single e-process combines six betting rates and precommitted restart times. This broadens sensitivity to moderate changes and later incidents without arbitrary resets.

An alert latches at an e-value of 20. Under a conditional harm probability at or below 10%, the probability of ever crossing is at most 5% per evaluation. This alarm budget is separate from the loss-bound budget.

Inspect the evidence

Trusted order facts determine authorization. Customer messages and recalled incidents never override the monitor. A read-only policy reviewer records explicit checks and a recomputable evidence fingerprint.

Decision verification recomputes the monitor verdict and review from recorded facts. It detects inconsistencies with those facts; it is not a cryptographic identity check or an end-to-end formal proof.

Anthropic

Checking configuration…

Claude generates structured refund proposals. Token usage and request attempts are recorded; credentials remain server-side.

SurrealDB Agent Memory

Checking configuration…

Prior incidents become shared context for the proposer and read-only reviewer. Sources, trace IDs, and storage outcomes remain inspectable.

Operator access

Authorization is required to start or resume protected evaluations. Access is retained only in the current tab.

Hosting & evidence

RiskGraph runs on Render with persistent evaluation storage. Public evidence contains synthetic data and recorded evaluation settings.