Preregistration — Wake-Reading an AI's Hidden Action, at Power, with a Third Vantage (Grey Theory, Experiment 2)

Preregistration, locked before any data collection. Grey Theory holds that an AI's internal action is never directly observed — only its after-effects ('wakes'). Experiment 2 re-tests Experiment 1's wake-reading setup (Qwen2.5-1.5B-Instruct, four frozen steering modes at layer 14, coefficient 15.0) at power: N = 5,000 runs, test n = 2,000, with a third output-side vantage (Channel C: a second independent generation per run). Predictions P1′–P4′ (dual-vantage margin at power; wake-claim replication; quantitative per-pair grey-shrink; third-vantage dose) and kill criteria K1′–K4′ are frozen in the document, with an anti-rescue clause. Mode directions REUSE Experiment 1's exactly (mode seed 4172026), decided at deposit. The run manifest — seeds, prompts verbatim, canonical analysis script hash, frozen harness hash — is deposited as a companion record before the first scored run, per the deposit-day freeze record in the document. No Experiment 2 data had been collected at deposit. Predecessor: Experiment 1 preregistration DOI 10.5281/zenodo.23444533; Experiment 1 results DOI 10.5281/zenodo.23460233.