Founder story: Quaesitor

[sandovabarbora]

Adds a founder story for Quaesitor, an independent review service measuring how often an AI assistant answers over a warehouse confidently, fluently and wrongly.

One reason it may fit here: the piece carries a measurement that bears on the argument in "ETL is Dead". Giving an agent the company's own schema documentation did not make it more correct in our runs. It made the errors more plausible, moving them from visibly absurd to clean and still wrong. That is a narrower thing than the Context Store you describe, schema docs rather than context and decision objects, but it points at the same problem from the measurement side: context appears to change how failures look before it changes how often they happen.

77 hand-verified questions, 114 runs, 5,533 classified answers, all reproducible from the public repository at https://quaesitor.eu.

Happy to adjust length or framing.

[sandovabarbora]

One more reason this might fit, and it is a narrower point than the story itself.

"Data Engineering After AI" routes work by confidence: high-confidence inferences go to an LLM-as-Judge, medium-confidence ones to a domain expert, and as you put it, the humans are reviewing the uncertain ones rather than every artifact. That architecture needs a usable confidence signal to route on, and the cheapest one available is self-consistency: ask the model more than once and keep what repeats.

We measured that, because our harness asks every question three times anyway. Across 1,644 answered question-instances, restricting to answers where all three attempts agreed leaves the error rate at 84.9% against 85.1% at full coverage. Twenty points of coverage for two tenths of a point of risk. The risk-coverage curve is flat, which is not a selective classifier with a badly chosen threshold, it is the absence of a signal.

That is not our result alone. Julka (arXiv:2606.22728) reports the same thing in retrieval rather than SQL: 42% of knowledge-graph errors and 59% of dense-retrieval errors showed zero dispersion across five samples. He calls it agreement blindness.

So the Contextualize layer's routing depends on something nobody has shown works, and the obvious candidate does not. That seems worth someone measuring properly, and it is the direction the founder story ends up pointing at.