irpf-bridge: Controlled Writeback into Official IRPF Desktop Save-State Artifacts

Abstract

The Brazilian individual income-tax return (IRPF) is prepared in a desktop application distributed by the federal tax authority, the Programa Gerador da Declaração (PGD), which persists declaration state in structured local artifacts. Public automation efforts usually bypass these artifacts through GUI scripting, parallel declaration representations, or read-only inspection. This paper investigates a different objective: whether the official save-state family can be treated as the primary read-write object while keeping humans in control of every proposed change. We present irpf-bridge, a local Python research prototype that ingests baseline XML/BKP pairs, maps supporting-document evidence into candidate deltas, validates them, and writes back narrow 2026 declaration slices under explicit scope boundaries. Results from the public repository combine deterministic rerun-convergence checks, exercised preservation evidence, and bounded manual reopening and acceptance observations in the official PGD. The contribution is not broad tax automation; it is evidence that, for the tested scope, the official artifact boundary can be crossed programmatically in a safe, reviewable, and verifiable manner without making GUI automation the primary integration surface.