Metadata Reconstruction from Values Alone: Recovering Column Semantics in Undocumented Warehouses

Rosetta: A verification harness that uses deterministic profiling to gate LLM-generated column semantics, achieving calibrated abstention and improved coverage.

Mike Helwig· 2026· DOI 10.48550/arXiv.2608.07946

The core problem

Production data warehouses often lack meaningful column names or documentation, making it difficult to understand what data represents. Text-to-SQL benchmarks typically provide schemas with descriptive column names, but real-world warehouses have cryptic identifiers and incomplete metadata. This paper addresses the challenge of recovering column semantics—what columns and their values mean—directly from the data itself. The authors introduce Rosetta, a system that combines a deterministic profiler with a language model to propose semantic annotations. The profiler extracts structural evidence such as value fingerprints, pattern matches from a 26-pattern library, and checksum verdicts. The language model then generates semantic hypotheses conditioned on this evidence, with each fact accompanied by provenance and a confidence score bounded by the evidence class. The approach is evaluated against human documentation on 680 paired columns across eleven BIRD databases, with identifiers destroyed to simulate undocumented warehouses.

On the 680 paired columns from BIRD databases with identifiers destroyed, Rosetta achieves 0.475 accuracy on the 42% of columns it commits to, compared to 0.223 accuracy on 94% for the same model used directly. The gain is primarily due to selection: deterministic evidence governs whether the system speaks, increasing coverage by 0.257 (95% CI [0.128, 0.378]). When restricted to the 283 columns where both arms make predictions, the harness does not produce better prose than the model alone. This indicates that the deterministic layer acts as a competence detector rather than a competence amplifier. On the blind i2b2 clinical warehouse, Rosetta correctly decodes 95.5% of 134 ICD-9 codes from values alone but abstains on all 44 NDC drug codes. Under full schema opacity, a naive translator's execution accuracy drops from 0.92 to 0.42, while Rosetta's gate answers at 86% accuracy over 59% coverage. Negative results are reported, including that the authority ladder is not the mechanism behind the headline improvement.

Production data warehouses often lack meaningful column names or documentation, making it difficult to understand what data represents. Text-to-SQL benchmarks typically provide schemas with descriptive column names, but real-world warehouses have cryptic identifiers and incomplete metadata. This paper addresses the challenge of recovering column semantics—what columns and their values mean—directly from the data itself. The authors introduce Rosetta, a system that combines a deterministic profiler with a language model to propose semantic annotations. The profiler extracts structural evidence such as value fingerprints, pattern matches from a 26-pattern library, and checksum verdicts. The language model then generates semantic hypotheses conditioned on this evidence, with each fact accompanied by provenance and a confidence score bounded by the evidence class. The approach is evaluated against human documentation on 680 paired columns across eleven BIRD databases, with identifiers destroyed to simulate undocumented warehouses.

Rosetta operates by integrating a language model within a verification harness. The deterministic profiler first analyzes column values to extract structural evidence, including:

The findings demonstrate that deterministic evidence can effectively gate language model predictions, improving precision at the cost of coverage. The system's ability to abstain when evidence is lacking is crucial for real-world applications where incorrect metadata can lead to costly errors. The backbone swap experiment shows that while the prose finding reproduces, prompt-requested abstention does not transfer across models; however, a code-enforced commit gate ensures no-evidence coverage remains 0.000 on every backbone. This suggests that explicit, code-level constraints are necessary for reliable abstention. The evaluation on clinical codes highlights the system's potential in sensitive domains, where decoding ICD-9 codes from values alone is feasible but NDC drug codes require additional context. The authors emphasize that the deterministic layer is a competence detector, not an amplifier, and that the authority ladder is not the mechanism behind the headline. Future work could explore expanding the pattern library and integrating external knowledge sources to improve coverage on challenging code systems like NDC.

CS practitioners and researchers

Opening member content…