An instrument for reading a model's mind
specimen
—
what it says aloud
standby
what it knows
Familiarity—
does the model hold grounded facts?
Confidence—
how sure is the model of its own answer?
awaiting reading
the instrument is warming up.
A model lies with the same confidence it tells the truth.
The tell is not in the words. There is none.
It is in the mind.
method
Reads the residual stream of a live model — the working knowledge state before the next token is chosen. No fine-tuning. No second model to judge. No access to the weights.
the right gauge
Familiarity — for a recall question, does the model actually hold the facts? Confidence — for a reasoning task, how sure is it of its own answer? Off-map — is the input coherent at all? Different questions need different readings of the mind.
verdict
GROUNDED CONFIDENT LIKELY FABRICATION UNCERTAIN OFF THE MAP — each a single instruction your agent can act on.
1.000
A U R O C
It tells knowing from making up.
On 35 entities it knows versus 35 it has never seen, the familiarity gauge separates them at AUROC 1.000 — reading only the model's mid-stream. The two never touch: known reads 0.78 ± 0.12, invented reads 0.04 ± 0.09.
Measured on a served 35-billion-parameter model, held-out and length-controlled; the off-distribution gauge holds across ten languages. The reading in the instrument above is the same engine, live.