The only way to prove a contract test can fail is to ship a server that lies

Someone building an MCP server shipped a second binary whose only job is to break the first one's advertised contract: mcp-worse mirrors two of mcp-better's tool names, omits the list-cache stamps, and serves tools in the wrong order on purpose ^[1]. The test that compares them, contrast-smoke, exits 0 only if the good server passes the contract and the bad one fails it ^[2].

That second condition is the whole argument. A suite that only runs against the happy path proves the happy path exists; it does not prove the checker would catch a violation if one arrived, which is why, according to the post, mcp-worse exists at all ^[3]. Most servers never answer the question of what breaks when the docs are wrong, because nothing in the codebase is built to fail deliberately ^[4].

The gap being tested is real and unenforced. Server documentation routinely claims statelessness, a cacheable tools list, and stable tool order, and nothing in the protocol stops a server from claiming all three and honouring none of them ^[5]. From the client side the tool names are identical whether the server behind them is honest or not ^[6].

The mechanics are unglamorous, which is the point. The lying list function sorts tools into a hardcoded anti-order, echo before health, where the honest server returns health then echo ^[7], and deliberately skips the with_ttl_ms and with_cache_scope calls that stamp the response ^[8]. The predicate for a passing server checks three things on the wire: the expected tool names, a ttlMs greater than zero, and a cacheScope of Public ^[9]. The predicate for a lying surface is the inverse: missing either stamp, or wrong order ^[10]. contrast-smoke spawns both binaries as child processes and speaks MCP over stdio, so it reads the wire rather than the source or the README ^[11].

Captured output from 16 August 2026 against v0.4.3 shows better returning health, echo, confirm_echo with ttl 60000 and scope Public, and worse returning echo, health with both fields absent ^[12]. That ttl is a 60 second list cache lifetime ^[13]. The honest server has since grown a third tool, confirm_echo, that the lying companion was never updated to match, so tool count is now part of the divergence as well ^[14] - a one-tool gap on top of the ordering and stamping breaks ^[15].

Two build commands: cargo build --bins produces both binaries, since contrast-smoke needs both on disk to probe them ^[16], then cargo run --example contrast-smoke ^[17].

The maintenance obligation cuts both ways, and this is the part worth stealing. If mcp-better regresses, someone dropping the ttlMs stamp in a refactor or letting tool order go nondeterministic, the same probe fails loudly against the good server ^[18]. If mcp-worse ever accidentally starts passing the contract, that fails too, because a companion that stops lying reliably makes the test worthless ^[19]. The lying binary is labelled in its own health response with tier LYING-DEMO and a warning that it deliberately fails the contract for teaching, and the author says it never ships to a registry ^[20].

Watch whether the known-bad binary keeps pace with the good one. It has already fallen a tool behind ^[14], and a stale liar is the failure mode that quietly turns this from a contract test back into a smoke test.