We asked four students to write 110 questions. Our system could express 10.

Every agent benchmark has a dirty secret: the people who wrote the tasks knew what the system could do. Ours included. So we ran the experiment that hurts — we handed four students a plain-language description of two datasets, deliberately withheld the operator list, and asked them to write whatever questions they actually wanted answered.

What people actually ask

The gaps were not random. Tagging every inexpressible question with its missing capabilities (tags overlap, so counts exceed 98):

One operator family — grouped aggregation — would make 30 more questions expressible on its own. Nothing else comes close. That is what workload-driven interface design looks like: not a designer's intuition about what agents need, but a count of what real questions break on.

Two findings we didn't pay for

The students misjudged the data. Several questions target months before the log begins — a good system should answer zero there, not hallucinate, and it should expose its time range so an agent can tell "no matching events" from "outside coverage." Sharper: even after reading our two-clock explanation, some writers anchored transaction time to event years. The database did not exist in 2013; its beliefs began at ingest. Storing two clocks is not enough — the interface has to teach them.

The fairest fight we ever staged

The same review cycle produced our favorite baseline: give the same model direct SQL access to the identical bi-temporal store our operators run on, schema and temporal semantics documented in the prompt. No information asymmetry left. Three things happened:

And on a flat rating schema (Bitcoin-OTC), SQL generation tied us outright — 0.340 vs 0.309, confidence interval spanning zero. We put that number in the paper with pleasure. The operator interface is not universally better; it is better exactly where computation composes, and it is the only configuration whose every emitted claim is checked against a deterministic trace. SQL answered 94% of tasks and was right 30% of the time; we answered 71% and were right 55% of the time.