Top 46%
en este catálogo (1067 prácticas puntuadas)
Las puntuaciones se concentran en valores altos, por lo que la posición dentro del catálogo suele ser más reveladora que el número por sí solo.
Evidence of impact / measured public value1/3
Only internal accuracy/match-rate benchmarking has been published for both workstreams; neither the SIC/SOC model nor ScannerAI has reached a deployment stage with measured cost, quality or timeliness gains for public statistics.
Transparency, fairness & accountability3/3
ONS published full comparative performance tables across five surveys, open-sourced the ACE2 codebase, and explicitly disclosed weaknesses including gold-coded-data scarcity for SIC and the model's failure to transfer to the DLHE survey population.
Transferability / demonstrated replication2/3
The same pipeline was tested across five distinct surveys with directly comparable published results, demonstrating reuse of one method across use cases, though the sharp accuracy drop on DLHE shows transfer is not automatic.
Scalability beyond pilot1/3
Both the autocoding model and ScannerAI remain in research/prototype status; ONS's own project page states no deployment decision or timeline exists yet.
Governance, capability & sustainability2/3
The design specifies a human-in-the-loop safeguard routing low-confidence records to expert manual coders, but no independent oversight body, ring-fenced budget, or continuity plan beyond the Data Science Campus's own research programme is documented.
Evidoria. La Apuesta de la ONS por la IA en las Estadísticas Oficiales — de la Investigación en Codificación Automática SIC/SOC al Prototipo ScannerAI de Lectura de Recibos. ID persistente: 78258360-2794-49e0-8695-2da36192b55a.