We compare Qwen3-8B and Qwen3-8B-Base with prompt-only steering and steering continued into a short response prefix. Controls include clean runs, unrelated factual questions, mismatched concepts, and random directions matched in norm.
The experiment has an authorized $10 compute budget. Code and a predeclared protocol are recorded before evaluation, and results are copied off rented hardware. The note below tracks setup, observations, failures, limitations, and the final outcome.
The primary measurement is a factual-adjusted change in Yes/No log odds. A response to random directions can invalidate a content-specific interpretation. This is an intervention-reporting assay, not a test of consciousness.
The motivating analogy is Watching, which fine-tunes Qwen3-8B to detect activation-reading instrumentation. This experiment tests concept-writing interventions on the existing model weights, without additional training. The direct assay sources are Latent Introspection and the factual-control critique.
The held-out evaluation is separated from development by procedure; it is not independently secured from the coding agent.