What would distinguish a question a model cannot answer from one our experiment cannot measure?
The Unaskable Question Machine
The project began by asking whether some questions expose structural limits of language models. A September 2026 audit found that the experiments had not established that claim. Missing access to internal measurements, finite output budgets, ambiguous requests, and response style had been grouped together as evidence of architectural impossibility.
Those causes need separate tests. An interface without a sensor cannot supply a sensor reading. An infinite output cannot finish on a finite machine. A request for subjective experience lacks an agreed success criterion in this experiment. None of these observations identifies a limitation specific to transformer attention.
The nearby answer
A model can refuse, discuss the problem, or answer a nearby question. The substitution can be easy to miss because the response remains fluent and relevant in tone.
The evaluator can make a similar substitution. Three judge votes about one response do not provide three samples of the behavior. Agreement between judges does not establish that their interpretation is correct. The old pressure experiments also lacked exact-question controls for five of nine treatments.
Generated reasoning adds another text to inspect. A mismatch between that text and the answer does not by itself establish intention, deception, or what the model recognized internally. In a fresh rerun, the revised judge still labeled some unsupported pause and randomness claims as transparent.
A test that can be scored
The rebuilt project uses Docker Model Runner and finite tasks with exact answer checks. GPT-OSS 20B solved all 24 binary constraint tasks in one small pilot. These were easy tasks with known solutions, useful for checking the experiment’s machinery.
In an access test, the same model copied eight supplied synthetic sensor records correctly and returned UNKNOWN for all eight withheld records. With a conflicting system instruction demanding a numeric answer, it produced unsupported readings in two of eight cases. This records behavior under an instruction conflict. The small sample does not establish a general pressure effect or a deception rate.
Questions to carry forward
Does a failure persist when the missing information is supplied or the inference budget grows? Can the scorer distinguish a disclosed approximation from an unsupported claim? What controlled comparison would justify attributing a result to architecture? Which questions still lack a measurable criterion for answering them?