Blinding the AI Co-Scientist: Mechanical Pre-Commitment for AI-Assisted Empirical Research

AI assistance has collapsed the cost of producing an empirical analysis, and the cost of producing the next one under a different specification. When a new specification can be issued as a single instruction, what separates honest iteration from specification search is intent, and intent is not auditable. Pre-registration is built for that gap, but it is enforced socially: a reader compares the finished analysis against the original plan and trusts that deviations were reported honestly. We present a protocol that enforces the pre-commitment mechanically, together with a deployment in which it bound against our own interest and required us to accept a failed outcome. Six mechanisms carry it: hash-frozen registration with dated, append-only amendments and third-party timestamping; structural blindness, in which the estimator package contains no loader for real data; a scripted unblinding order; a pre-decided kill table that never fires on a result; launch certificates that the one-shot production run must clear; and two-seat review, in which the seat that produced an artifact never certifies it. Together these make key forms of specification search technically unavailable within the registered pipeline rather than merely discouraged. The deployment is a causal event study in finance whose pre-registered admission gate refused its own headline analysis twice, and the refusal shipped under the pre-registered fallback. We evaluate the protocol on what its artifacts let an outsider check: the refusal, the amendment record under ambiguity, an adversarial review conducted inside the blindness contract, measured enforcement cost, and one failure the machinery did not prevent. That failure taught the design rule we close on: every check on an outward-facing claim must run from the reader's vantage point.