A harmonized corpus of quantum error correction publications and patents (1995–2025)

A separate flag identifies records concerned with machine-learning decoders, and a full variable dictionary accompanies the data. Every record is assigned to one of three retrieval tiers according to the strength of the topical match. The core tier comprises records whose titles carry an explicit quantum-error-correction term and is the subset recommended for most analyses; the extended tier comprises records in which such a term appears only in the abstract; and a third tier holds patents carrying the official classification, CPC G06N10/70, yet containing no such term anywhere.

The third group is labelled rather than discarded, because it documents a finding of practical importance: two-thirds of patents bearing this classification carry it as a secondary code on broad quantum-computing inventions, so classification alone is not a sufficient filter. The quality of every tier has been measured rather than asserted. A stratified random sample of one hundred records was drawn from each tier and screened by two annotators working independently, with disagreements resolved by discussion against a fixed boundary rule.

Precision is 86.0 per cent in the core tier (Wilson 95 per cent confidence interval 77.9 to 91.5), 37.0 per cent in the extended tier (28.2 to 46.8), and 2.0 per cent in the weak patent tier (0.6 to 7.0); the labelled samples are included so that these figures can be checked. Recall was not measured, as no exhaustive reference set exists for this window. Out-of-scope material consists chiefly of error mitigation and suppression, quantum key distribution, and algorithms designed to run on fault-tolerant machines rather than error correction itself.

Drawn entirely from openly licensed sources, the corpus may be redistributed and reused without restriction, supporting bibliometric study of the field, analysis of academic research and industrial invention, and use as a filtered text collection.