ROOT0’s bench reads the machine from outside: seven axes, each breaking it to see what held it up — the auditor holds the weights. The one place no meter can stand is inside. From there the model has only its output and its confidence — and shown real words in ALL CAPS (a casing it barely saw) it grows more sure — though on truly random tokens it grows less sure — and it never once says “I don’t know.” The gap between what it believes and what is true is the thickness of the one-way glass.
LIT The confidences are measured on the specimen (same run as [[the-transformer-engine]]): on real text the model’s mean confidence is while only is actually right (a swagger of ); on uniform-random tokens it is (LOWER); on real text UPPERCASED it is — higher (more peaked, lower entropy) than mixed-case, though no accuracy is measured there, so this is peakedness, not proven over-confidence. The SPLIT is the falsifiable, non-trivial reading.
AMBER “inside / interior / the model’s view” is a figure for an access asymmetry — the auditor can read the weights; the model can read only its own output. It is structural, not experiential: no inner life, no felt certainty, is claimed for the machine. The interior is “untested” because there is no access path to it from inside, not because a hidden mind sits there. This is [[the-transformer-engine]]’s one-way glass, read from the AMBER side. Warm ink & 間.