9. Quality Assurance and Oversight
Run-level diagnostics, calculation controls, and their limits.
Each run records the number of prices removed as inconsistent with the consensus for their product category, the product categories excluded in full for persistent inconsistency, and the regions whose cost score was suppressed for insufficient basket coverage. Those counts are reviewed against previous runs, because a sharp movement in any of them indicates a change in the input data and not a change in cloud prices.
The band boundaries fitted for each provider in each run are recorded alongside those counts, so that a grade published in any past version can be traced back to the boundaries that produced it.
Three controls in the calculation itself limit the influence of bad input. The median is used in place of the mean at every point where prices are combined, both within a region and across the basket. Prices inconsistent with their product category are removed before any statistic is computed. Regions with too little evidence are suppressed. None of the three detects a provider price list error that is internally consistent across all regions, since the method has no reference outside the published lists against which to test one.