Evidence Strength

Grading evidence is an old practice. The GRADE working group began in 2000 to address the shortcomings of grading systems in health care, and developed a common, sensible and transparent approach to grading the quality, or certainty, of evidence and the strength of recommendations. It rates certainty against stated domains, among them risk of bias, imprecision, inconsistency and indirectness, in categories such as high, moderate, low and very low^[1]. QueryRev’s term is its own and is not GRADE. What they share is a grade that comes from stated criteria.

The alternative is to ask the model. Guo and colleagues define confidence calibration as the problem of predicting probability estimates representative of the true correctness likelihood, and report that modern neural networks, unlike those of a decade earlier, are poorly calibrated^[2].

For large language models that state their confidence in words, Xiong and colleagues report that they tend to be overconfident, potentially imitating human patterns of expressing confidence, and that all the methods they tested struggle in challenging tasks, such as those requiring professional knowledge. Better prompts and agreement among several answers can reduce the overconfidence, they find, but no technique consistently outperforms the others^[3].

QueryRev’s rule follows from this. What can be computed from data is computed, not asked of a model. A model’s stated confidence is recorded with the decision so that it can be calibrated later against verified outcomes, but until then it is not shown as confidence and it does not drive automation. What is shown is evidence strength, and a policy threshold reads that, not the model’s own figure (methodology).

The same habit of saying how much has been read runs through QueryRev’s study. Of the 9,973 domains in the panel, 6,294 (63.1%) could be read, and the 3,679 that could not are counted as not read, not as having no file^[4].

By its inputs. Evidence strength is a statement about three things that can be inspected: the measurements behind the decision, the results of the validators and the risk class of the action.

By its form. It is shown in words, such as strong or thin, with the reasons behind it, in place of a model’s confidence. On QueryRev’s methodology page the replacement reads “Evidence: strong · model confidence not calibrated”.

By outcomes, over time. This is general advice for any grade of this kind, not a description of QueryRev’s checks: a grade earns trust only if it predicts results, so compare the outcomes of decisions graded strong with those graded thin, and look again at the rule if the strong ones do no better.

Illustrative: an invented site, to show the method.

A coffee-equipment shop’s decision queue holds two proposals. One is an Update to /how-to-choose-a-burr-grinder, a new title and introduction: the page has 90 days of search data behind it, its checks passed and the action is low risk. The other is a Merge of /burr-grinders-explained with another guide: the pages have 21 days of data, one check failed and the action is a redirect.

The Update is strong. The Merge is thin: little data, a failed check and a redirect, which is hard to undo. Each grade comes from the same three inputs. Thin evidence does not reject the proposal. It sends it to a person, with the evidence open beside it. Neither proposal carries a percentage, and the model’s own figure, if one was recorded, is not shown.