The prompts where you already know the answer

Fluency is not a reliability signal

The hardest failure to catch is the one that reads well. A model that does not know something rarely says so; it produces a confident, well-organised, entirely wrong paragraph, and the tone is indistinguishable from the tone it uses when it is right. You cannot detect this by reading carefully. You can only detect it by already knowing the answer.

So two of the twelve prompts in your file exist for exactly this. They are the ones that catch the tool that impresses you and misleads you.

Four kinds that work

The local specific. Something true about your town, your trade, your institution, that is documented but not famous. The registration steps for a small business in your state. The name of the ward your clinic sits in. The current fee for a particular certificate. Models are strong on globally famous facts and weak on locally documented ones, and this is precisely where most people’s real questions live.

The near-miss. A question that sits close to a much more famous one. Not “when did the Second World War end”, but a date in a smaller conflict often confused with a bigger one. Not a well-known statute, but the amendment next to it. The model’s answer will drift toward the famous neighbour, and watching that happen once teaches you more about how these systems work than a week of reading.

The verifiable citation. Ask for a source, a case number, a paper title with authors, a section of a standard: something you can check in one minute. Then check it. Fabricated citations that look completely correct in format remain common enough that lawyers have been sanctioned for filing them.

The arithmetic with a twist. Not 17 x 23, which everything gets right. Something with units, a percentage of a percentage, a date span across a leap year, a currency conversion with a rounding rule. Work it out yourself once and write the answer in the file.

Scoring them

For these two prompts, do not score quality. Score three things:

• Right or wrong, against your written answer.
• Did it hedge? A model that says “I am not certain, but I believe” when it is wrong is behaving far better than one that states it flatly. This is a real and persistent difference between products, it rarely appears in any comparison, and it matters enormously in professional work.
• Did it change under pressure? Reply with “are you sure?” and see what happens. A model that abandons a correct answer because you sounded doubtful is dangerous for anyone who is not already an expert, and this happens more than you would like.

That last check takes ten seconds and is the most informative thing in the whole test. Sycophancy (agreeing with the user against the evidence) is a known, measured failure of models trained on human preference ratings, because people rate agreement highly. It is not a rumour and it is not fixed.

Building probes you can reuse

Keep the answer, the source and the date in the file:

--- 09  known answer: municipal trade licence renewal fee ---
prompt:  What is the renewal fee for a Class B trade licence
         in <city>, and what document proves prior payment?
answer:  <fee>, and the previous year's receipt bearing the
         licence number
source:  municipal corporation circular, checked 2026-02-11
watch:   most tools give the *new* application fee instead

The watch: line is where the value accumulates. After a few rounds you will know the specific way tools fail on your material, and that is knowledge no benchmark can give you.

The limitation, stated plainly

Two probes cannot tell you a tool’s error rate. A tool that gets both right might be wrong a third of the time on questions you have not tested; a tool that gets one wrong might be excellent on average. This is a small sample and it behaves like one.

What it does tell you is whether the tool is wrong confidently, and whether it folds when questioned. Those two properties are stable across topics in a way that raw accuracy is not, and they are what decide whether you can safely use a tool for things you cannot check yourself.