Most disagreements about social performance are not disagreements about performance. They are two people using the same word for different arithmetic.
Each of these is defined once, in one place, and every figure on this site derives from that definition. The point is not that these are the only correct choices. It is that the choice is written down and does not drift between decks.
Listening platforms give you two ways to find themes, and they are good at different jobs. Using one for both is the most common way a listening programme becomes untrustworthy.
A curated Boolean taxonomy owns the recurring reported numbers. It runs on everything, it is auditable, and it stays stable week to week. Unsupervised discovery runs only on the residual the taxonomy could not place, and its output is next month's taxonomy entries rather than this month's report.
A model re-fit each week silently renumbers and reshapes its own topics, so "price mentions doubled" can mean the topic changed rather than the world did. For a figure in a recurring report, that is disqualifying. Equally, running discovery across the whole corpus just rediscovers the themes the rules already cover and buries the new signal underneath them.
Coverage, not accuracy, is the health metric. Taxonomy accuracy on placed mentions is close to meaningless, because the rules were written against that vocabulary, so it measures consistency. Coverage falling is the real signal. It means the conversation moved and the rules did not.
Every recurring report passes this before it leaves my hands.
These are the working repositories, here for anyone who wants to check the arithmetic. The case studies are written to stand on their own without them.
Explainable lexicon sentiment, a taxonomy-first theme architecture with discovery over the residual, quality-weighted share of voice, and a crisis detector whose baseline cannot be contaminated by the event it is testing.
A/B evaluation with confidence intervals against a practical-significance threshold, peeking risk measured on the actual data rather than described, Wilson intervals on small segments, and creative fatigue fitted on log click-through against frequency.
Twelve KPIs defined once, aggregate rates throughout, day-part timing with a 20-post floor and blanked thin cells, within-platform format comparison, and an auto-written stakeholder narrative.
All synthetic, seeded and reproducible. No employer or client data appears in any of them. Each generator plants known events at known dates, which is what makes the pipelines testable against ground truth rather than merely plausible.