Verification for the Buyer, Not the Seller: A Deployed Construction-Quote Gate

The current generation of AI tools for construction cost is built, almost without exception, to help the contractor quote faster, protect margin, and suggest upsells; it puts the informational advantage on the seller's side of a market where the buyer already cannot tell whether a quote is fair. This paper reports a tool built for the other side. Residential construction is a credence good, a purchase whose fair price the buyer cannot verify even after the work is done, and the economics of such markets holds that giving the buyer a means of verification would close the asymmetry. Yet when a large laboratory experiment (Dulleck, Kerschbamer, and Sutter, 2011) granted buyers verifiability, it changed market efficiency at best slightly, leaving a twenty-year puzzle. A feature of that work is easy to miss and central here: verifiability was handed to the buyer as a rule of the experiment, whereas in a real market no buyer has any instrument with which to verify a construction quote at all, and none has been deployed and studied. That tool is a verification gate, running in a live construction-cost service, that admits buyer-contributed quotations to a shared comparison board while preventing a malicious or mistaken contribution from corrupting it. The gate treats the model that reads a contribution as untrusted, allowing it only to propose structured fields with verbatim source spans and no capability to admit, write, or act; it verifies each field against the contribution's own text, re-derives reliability from agreement across repeated extractions rather than from the model's confidence, and decides by a deterministic, content-hashed function. Because the deciding component cannot act on an instruction, an instruction embedded in a contribution has no path to the decision, which we show at runtime as an invariance: the decision is unchanged whether the instruction is present or absent. The gate's components, untrusted proposal, source-grounded verification, deterministic re-verification, capability denial, and cryptographic auditability, are drawn from existing work and none is claimed as novel. The contribution is the position and the setting: to our knowledge the first deployed, buyer-side, verification-only gate in this credence-good market, offered as a field artifact for a verifiability question the discipline has examined only in the laboratory, and demonstrated on a live deployment through a controlled injection pair, a fabricated-field refusal, a fail-closed response to extractor failure, and reproducible decisions.