Measuring consciousness-adjacent constructs in artificial systems is difficult. Step-level proxies can correlate within one architecture while independent tests fail to transfer, and small factorials can produce r > 0.99 for constructs the architecture cannot express. We introduce Construct-Architecture Matching (CAM), a seven-field machine-checkable pre-registration template for catching four failure modes: insufficient signal dynamic range, probe-family mismatch, missing output channels, and protocol self-contradiction. We apply CAM to five case studies from a single consciousness-benchmark project. Three stopped in less than 60 seconds of smoke compute, including one in which the authors' own pre-registration contained three independent protocol-design errors and the gate failed in 32 seconds. The contribution is methodological. CAM did not add a sixth validated construct to the benchmark; it made the negative result auditable and stopped unsupported promotion. The same template can flag author-side protocol errors as well as substrate-side mismatches, helping prevent unsupported claims from reaching publication.
Fuwang Feng (2026) studied this question.