Sema

The AI-native, neurosymbolic programming language. Models, contracts, effects, and semantics are language constructs — not library calls.

Get Started Install the Skill Language Tour View on GitHub

Sema is a programming language for the era of language models. Its syntax is familiar — Python-shaped indentation, def, struct, enum, traits — but its semantics are new: a model can implement a function, similarity is an operator, effects and budgets are part of a type, and contracts are checked, not hoped for. The deterministic core and the generative edge live in one language, with one verification story.

# A function whose body is written by a model — but fenced by types, effects,
# a token budget, and a semantic contract that is actually checked.
simulate def summarize(article: str) -> str !{model.invoke} by gpt:
    sem "A faithful, single-paragraph summary of the article."
    budget tokens=400
    ensure len(result) < len(article)
    check semantics("the summary makes no claim absent from the article")

# The deterministic core stays provably invoke-free: `!{}` means no model, no I/O.
def dedup_titles(titles: list[str]) -> list[str] !{}:
    return semantic.dedup(titles, threshold=0.86)

Models are first-class

simulate def … by model lets a model implement a function body under a sem descriptor, a token/time budget, and ensure/check semantics contracts. Models are values you can pin, swap, and schedule.

Semantics is a built-in

~= is calibrated similarity; semantics("…") is a natural-language predicate that type-checks as a guard; semantic.rank/filter/cluster/dedup are language operations, not SDK calls.

Effects you can prove

Every function carries a capability row — !{fs.read, model.invoke}. The deterministic core is !{}. The compiler tracks it; policies constrain it.

Verification by default

require/ensure/invariant are hard contracts; check/check semantics are monitored. sema assure fuzzes properties for counterexamples and, at gold, mutation-tests your suite.

Governance in the language

policy with verified examples:, monitor for distribution drift, with budget(...), supervise/heal, session protocols, and non-interfering collector taps — the operational concerns of AI systems, made native.

No silent no-ops

A construct that parses must have a real effect or fail loudly. sema check flags anything it does not recognize. What you write is what runs.

Start here

Why Sema The problem Sema solves and the bet it makes.

Installation Build the toolchain and run your first check.

Your First Program A runnable project in ten minutes.

Mental Model The deterministic core and the generative edge.

Go deep

Language Tour Types, effects, traits, enums, generics, pattern matching.

Neurosymbolic Core Semantic values, similarity, simulate, contracts, assure.

Governance & Ops Policies, monitors, budgets, protocols, self-healing.

Standard Library belief, cache, document, agent_loop, provenance, usage.

Guides Build agents, reports, solvers; interop, multimodal, packaging.

Language Specification The normative spec, generated from LANGUAGE.md.