AI make claims about code—tests pass, no secrets, behavior preserved—but these claims become stale when code changes. Assay ties each claim to a Merkle hash of the minimal dependency cone of the code it covers, so staleness becomes a simple hash comparison rather than a judgment call.
On five public repositories, a 600-token context brief costs 14 to 114 times less than an exploration proxy, warm rebuilds are up to 5 times faster than cold ones, and cone binding re-verifies only 7.9% to 81.9% of claims where repository binding re-verifies all and per-module binding misses 23% to 68% of required invalidations. A merge gate with no model in the loop blocks all nine scripted adversarial behaviors while admitting honest ones. This provides a cheap, precise floor for accountability in AI-assisted software delivery.