Filho (2026) — ESAA the pattern behind engineered agent evidence

Why it matters for NETEVO

ESAA specifies a four-element event-sourcing pattern for autonomous LLM agents: agents emit only structured intentions in validated JSON; a deterministic orchestrator validates and persists events in an append-only log; the orchestrator applies file-writing effects deterministically; and a verifiable materialised view is projected from the log. The agent's cognitive output is thereby separated from every mutation of project state, so audit and observability are properties of the architecture rather than assertions made after the fact.

The pattern produces evidence, not testimony. An append-only, hash-verified event log is a record of agent activity that does not depend on the agent's own account of what it did. Where a governance control names what must be tracked about every agent action, ESAA names the architectural pattern that produces a record that can be tracked deterministically — the difference between requiring an auditable register of agent actions and engineering one.

This is a preprint, not peer-reviewed work. Two validating case studies demonstrate the pattern at small scale: a landing-page project with 9 tasks and 49 events, and a clinical dashboard with 50 tasks, 86 events and 4 concurrent agents. Until the pattern appears in peer-reviewed form, it is best read as a well-specified architectural proposal with early small-scale validation.