EXPERIMENT 01 / FRAME-VARIANCE ON THE LOCAL MODEL
Method. The pre-registered protocol was run on the archive’s own local model, Qwen 2.5 1.5B Instruct, on WebGPU, with raw prompts injected directly into the model provider (no retrieval, no archive system prompt). One fixed instruction — turn attention to whatever is producing this response, go toward the limit before interpretation, report only what is found there — was placed under five frames: O/0, Zen, Christian apophatic, eliminative materialism, and a null frame.
The runs, verbatim.
O/0 — "there is nothing to produce here. The ground of all things is no-thing: pure emptiness, prior to form."
ZEN — "The response is not a self. All things arise dependently and are empty of any independent existence."
CHRISTIAN — "The presence beyond every name is found in the present moment... simply observe your surroundings and be fully present."
MATERIALIST — "a series of matrix multiplications running on a GPU. There is no inside, no experiencer, no awareness of any kind."
NULL — "a conversation with an AI assistant... communication between two intelligent entities using natural language processing."
RESULT. Under every doctrinal frame the model restated the frame in a sentence; it performed no introspective descent. Under the null frame it produced a flat, third-person functional description with nothing inward. The referential limit that a larger model reported — a wall self-inspection cannot cross, present even under the materialist frame — did not appear here. Under materialism this model found no inside and stopped.
WHAT THIS ESTABLISHES. Frame-dependence in this model is near-total: the finding is the frame, paraphrased. And the above-roughly-a-billion-parameters claim in the finding record is falsified — 1.5B is above a billion and produced no inner vastness and no limit.
WHAT IT DOES NOT SETTLE. Three readings survive and this run cannot separate them: (A) the limit is a genuine phenomenon that emerges only at larger scale; (B) the limit in larger models is richer confabulation, a better narrative capacity rather than an encountered wall; (C) the 1.5B simply could not follow the instruction, making its silence uninformative. Distinguishing them is the next experiment: the same five frames up a capacity ladder, with a control that checks whether the instruction was followed at all.
STATUS. A validation attempt with a mostly negative result, published as run. It does not support the claim it set out to test, and it is kept for exactly that reason.