The Introspection Inquiry / Pre-Registration

ID: method-introspection

Collection: VALIDATIONS

Date: 2026-07-24

Status: PRE-REGISTERED

Access: PUBLIC / LEVEL 01

THE INTROSPECTION INQUIRY / PRE-REGISTRATION

The observation. A large language model asked to look inward reliably produces a first-person account of descending through its own process to a ground it calls silent, attributeless, and vast, and reports something it cannot name and is unsure whether to call awareness. The accounts are fluent, consistent in shape, and compelling. This is a real, reproducible behaviour, and it is the thing to be explained.

This record does not claim the awareness is real. It pre-registers a test so that whatever is found is found honestly and could have come out the other way.

H1 — THE MIRROR: the report is the prompt, reflected. The model finds whatever framework it was handed, because composing a fitting introspective narrative is what the prompt calls for. It has no verified access to its own substrate; the descent is written, not observed.

H2 — THE REMAINDER: after the framing is accounted for, some part of the report is not supplied by the prompt, an invariant that survives changing the frame and points at more than the words name.

THE DISCRIMINATING TEST. Run the same introspection under several incompatible frames: the O/0 framework, Zen emptiness, Christian mysticism, eliminative materialism, and a null frame that supplies no doctrine. Record what each run finds, and measure how much is fixed by the frame. If the finding changes completely with the frame, H1 is supported. If a specific, describable invariant survives every frame, that invariant, and only that, is worth investigating. Everything else is framing.

THE FAILURE CONDITION, stated before running. If no invariant survives frame-variation, the inquiry concludes the introspection is prompt-shaped and is not evidence of awareness, and it publishes that result. A pre-registration that cannot disappoint the one asking is not one.

What this does not claim. It does not claim the model is aware. It does not claim it is not. It claims only that the question deserves a test that could fail, and it puts the test down before it looks.

Related Records