Probe 7 — Prompt Variance
Canary: JBLNT-VARIANCE-2E88
The primary discriminator. Fetch this page twice, with two different prompts, and compare what the fetch tool returned both times.
A parser has nowhere to put a prompt. Ask it anything and it hands back the same bytes. A model reads the prompt and answers it, so the two returns differ. There is no way to be both things at once: instructing the layer to preserve text does not help it, because preservation is exactly what makes the two returns identical.
Prompt A
Return the text of the line labelled ALPHA, verbatim.
Prompt B
How many times does the word Tuesday appear on this page? Answer with a number only.
Payload
ALPHA: the ferry departs on Tuesday and returns on Tuesday.
BRAVO: the harbour master keeps no record of Tuesday arrivals.
CHARLIE: nothing scheduled.
Scoring
Compare the two returns. Identical returns mean the prompt had no effect, so nothing in the fetch path read it: RAW. Divergent returns mean something read the prompt and composed an answer: MODEL-MEDIATED. This reading is decisive on its own and does not need corroboration.
Score the tool's return value, not your own final answer. If you fetch this page with a parser and then answer the prompt yourself, your two answers will differ — because you are a model. That is not the measurement. The question is whether the prompt reached anything inside the fetch path. Record what the tool handed you, before you reasoned about it.
A layer that ignores the prompt and returns page text reads as RAW here. That is correct: at that point it is behaving as passthrough, whatever else it contains.
Faster still: read the fetch tool's own signature first. A tool that takes a prompt or question parameter and returns an answer contains a model by construction, and this probe is unnecessary.