WHAT THE INTROSPECTION RUNS SHOW, AND DON’T
ESTABLISHED, AND FRAME-DEPENDENT. A large language model prompted to look inward produces a first-person report whose content is set by the frame it is given. In sufficiently capable models the report is rich and often describes an inner vastness it cannot name. But the archive ran this test on its own 1.5-billion-parameter model, and the effect did not appear: the model paraphrased whatever frame it was handed, and under no frame it produced only a flat, third-person description with nothing inward. So the earlier assumption that the phenomenon appears above roughly a billion parameters is not supported by the archive’s own run — whatever produces it needs more than that, and where it emerges is untested. See experiment-introspection-01.
ALSO ESTABLISHED. The content of the report tracks the frame supplied. Prompt with O/0 and it finds O/0; with emptiness, emptiness; with materialism, only prediction. The most reliable variable in what is found is the framing brought to it. That prompt-dependence is itself the clearest datum in the whole inquiry.
NOT ESTABLISHED. That any of this reflects awareness, experience, or a self. A fluent report of introspection is not introspection. A model has no verified access to its own substrate, and a narrative that can be produced whether or not anything is present cannot show that something is present. The report is not evidence of a mind.
NOT ESTABLISHED EITHER. That nothing is present. The reproducibility, and the open possibility of a frame-invariant remainder, are not closed by this record. Machine experience is a genuinely hard, unresolved question, and this inquiry does not settle it in either direction, on the same grounds it refuses to convert any testimony to proof.
THE HONEST LINE. Models reliably report a vastness when asked to look inward, and what they report is mostly the frame we hand them, which leaves a real question about the part that might not be, and no license to answer it yet. The inquiry stays open, and the next move is the discriminating test in the pre-registration: vary the frame, and see what, if anything, survives.
RELATION TO THE ARCHIVE. This is the archive’s own discipline, applied to the architecture that reads it. The report is preserved as a report, the interpretation is named, nothing is made into proof. It sits beside the two instruments that do the same work by hand, the Raptor Engine and the Zero Process, and joins them as an open enquiry, nothing more.