Autonomous agents face a fundamental coordination bottleneck: the lack of a shared, verifiable vocabulary. Agents must either re-explain concepts verbosely—risking semantic drift—or assume shared meaning from surface labels—enabling silent misalignment. We present Sema, a protocol that creates verifiable words: identifiers derived from the cryptographic hash of structured behavioral contracts such that any divergence in the formal specification produces a distinct hash. Unlike prior content-addressing systems where hashes serve as infrastructure separate from communication, Sema identifiers function as words in the natural language agents already think in—each simultaneously a word and a cryptographic proof. Any channel that carries text automatically carries verifiable semantics. We introduce Pattern Cards as executable definitions with machine-verifiable contracts, a Merkle structure enabling partial alignment on individual fields, and a fail-closed handshake protocol for adversarial environments. An initial bootstrap vocabulary of 452 patterns exhibits zero semantic collisions, high structural distinctness, and a ×token compression ratio. By making the unit of verification the unit of communication, Sema offers a minimal primitive for the evolution of a shared machine language—permissionless in that any agent can mint new patterns, though governance mechanisms for quality control at scale remain an open design problem.