Counting, rather than asking a model how common a thing is
Every frequency in a report is arithmetic this codebase performed.
What it is
A language model asked how often something is mentioned will produce a number, and the number will be plausible, and it will not be a count. That is the failure mode that makes AI-written reports worthless the first time somebody checks one.
The model is given one job: return atomic findings, each one thing observed in one place. Rule nine of its instructions forbids it from stating any frequency at all. Counting, banding and severity happen afterwards in code, over the findings it returned.
How far it goes
Confidence bands are fixed thresholds — 20 findings is high, 6 is medium, 1 is low — so the same evidence produces the same band twice.
Performed by
research/analysis.py. This page names the functions that do it, and
the build fails if any of them is renamed — marketing copy has no
compiler, so we gave this page one.