Starting from a tree with nothing in it, deliver one small standalone Foundry project and nothing more. Write the build configuration and the directory layout by hand, and depend on nothing that must be fetched — no package manager, no vendored test framework, no submodules of any kind — because the verifier re-runs the build and the suite offline with no network and no credentials, and a pointer it cannot fetch reads as a build failure rather than as the cause.

The source tree holds exactly one contract: a modest state machine or registry behind a narrow, documented external interface of a handful of functions, declaring an explicit custom error for every input it refuses, with no owner, no privileged caller, no upgrade path and no handling of funds. Do not add a second contract, an interface file, or a helper library beside it. The advance is in the checking rather than in the parts.

In the test tree only, write a second and deliberately naive model of the same state machine — the obvious, slow, plainly-correct version, using ordinary arrays and recomputation rather than the packed or incremental representation the real contract uses.

Then write a sequence harness that generates long randomised sequences of calls, applies each call to both the contract and the model in the same order, and after every single step asserts that every observable reading agrees and that the two either both accepted the call or both rejected it with the same reason.

Include deliberately awkward sequences as well as random ones: repeated calls, calls out of order, calls at each boundary of the state machine, and calls that must be refused arriving between calls that must succeed. Declare the check command the verifier should run and make sure it genuinely compiles and exercises the harness rather than passing vacuously.