Insuring Every Action: An Authority Frontier Framework for Runtime Actuarial Control of Autonomous AI Agents

Battle Count: 50

The paper is primarily about AI agent runtime safety and actuarial control of side-effect-bearing actions. While it borrows actuarial vocabulary (reserve, exposure, boundary, premium) and uses risk measures (VaR, ES) and conformal prediction from quantitative finance, the application domain is autonomous agent tool-use risk, not trading. The risk-measure machinery and capital allocation concepts have conceptual parallels to quantitative risk management, but the paper does not address trading strategies, portfolio optimization, or market microstructure.

The AAI framework requires: a deterministic quote-bind-commit protocol with atomic budget ledgers, a seven-class action taxonomy with typed predicates, a safe-default compiler, toll-bounded capability tokens with cryptographic signing, deterministic stateful canonicalization with canonical state hashing (SHA-256), conformal envelope calibration, alpha-spending schedule management, ambiguity-set construction for interface failures, three separate audit/telemetry/read-budget streams, and a policy automaton. The six implementation points for bitwise determinism are non-trivial. The cross-domain normalization and evaluation framework add further complexity. Production deployment would require integration with existing tool APIs, database backends, and human escalation workflows.

The paper emphasizes bitwise replay determinism as a core property (P1), with six specific implementation points securing it. Paired proposal-replay design fixes the proposal stream across budgets. However, the full grid-replay determinism harness is slated for the next implementation pass, and the live panel uses only 5 seeds per cell. The τ-bench bridges use public historical traces. The AAI contract parameters (taxonomy, safe defaults, alpha-spending schedule) are fully specified.