Ledger-First Observability: Audit-Before-Mutation for Deterministic Replay and Compliance in Production AI

Modern LLM-based systems execute non-deterministic computations with side effects that span APIs, data stores, and organizations. Traditional observability treats logs as an afterthought, which prevents faithful replay, frustrates incident response, and forces expensive manual compliance audits. We present ledger-first observability: a simple, enforceable law—audit before mutation—that makes executions replayable and compliance by construction.

Concretely, K-12 Law 5 mandates writing an immutable span to the ledger before any mutation of ephemeral state or durable memory; if the append fails, the episode fails closed. Combined with pure transition/output functions, seeded randomness, and idempotent routing in G-9, this yields deterministic replay at the application boundary, tamper-evident provenance, and direct mappings to regulatory clauses (EU AI Act, GDPR).

We specify operational semantics, provide a replay algorithm with a proof sketch, map representative regulatory obligations to runtime invariants, and discuss an implementation in a typed effect runtime (Effect TS). Case analyses (e.g., Knight Capital) illustrate how ledger-first would have prevented or sharply bounded impact.