Disposable Programs under an Immutable Specification: What a Validator Actually Guarantees — A Study on Data Extraction

Data extraction pipelines break when sources change format. Intent-Driven Adaptive Extraction (IDAE) makes the extraction program a disposable artifact synthesized by an LLM, while an immutable final intention (the telos), checked by a validator, is the permanent artifact; approved programs are reused for every record of the same format. We evaluate IDAE with a ground-truth oracle for every record, applied equally to all approaches, which separates acceptance by the validator from correctness. Seven experiments cover two synthetic domains, 560 real USGS earthquake records, two models (Claude Opus 4.5 and Claude Haiku 4.5) and six repetitions per domain. With Opus 4.5, IDAE converted all 560 USGS records correctly with 2 LLM calls, whereas per-record LLM transformation reached 62.7% with 560 calls and ~120× more tokens. The oracle, however, refutes the claim that zero hallucination is an architectural property: with Haiku 4.5, one faulty program amortized over its format corrupted 98% of a phase; reconstructing intermediate validators caused 74% silent corruption in the phases where it fired; a telos requiring a field absent from the source led synthesized programs to fabricate it in 330 records and to clamp valid negative depths to zero; and, because IDAE cannot tell a failing program from an invalid record, a single invalid record can poison a whole format. Amortizing explicit rejections removes this effect at no extra cost. IDAE's guarantee is exactly as strong as its specification: it buys cost, determinism and auditability, not semantic correctness. English version of the paper. The original Portuguese version is available at https://doi.org/10.5281/zenodo.23050791. Code, data and experiment logs: https://github.com/carloseduardodb/idae.