refined-types

First-class refinement types for Python — with native Pydantic v2 integration. A practical projection of Σ-types into a language without dependent types.

Status: 0.1.0a2 — alpha. Core, predicates, and the Pydantic v2 adapter are live; Hypothesis strategy registration is a stub and lands in a later alpha. Not for production yet.

What it gives you

A Refined[T, P] value is a T that has been checked against predicate P at a parse boundary, with that evidence preserved by the type checker. Downstream code does not re-validate; it cannot accidentally accept an unrefined T either. The discipline is Σ-type-shaped: a smart constructor is the only way in, the witness is phantom (erased at runtime, distinct at the type level), and the refinement transports through function signatures.

Quickstart

from refined_types import (
    Err, NonEmptyString, Ok, Positive, Refined, parse, try_parse,
)

# 1. Smart-constructor flow — total handling via Result
match parse(Positive, 5):
    case Ok(value=qty):
        ...  # qty: Refined[int, Positive]
    case Err(error=e):
        print(e)  # value -1 failed predicate Positive: x > 0

# 2. Ergonomic flow when failure is a bug — raise on the spot
qty = try_parse(Positive, 5)        # qty: Refined[int, Positive]

# 3. Functions can demand refined inputs
def buy(qty: Refined[int, Positive], item: Refined[str, NonEmptyString]) -> None:
    ...

buy(qty, try_parse(NonEmptyString, "apple"))   # ok
buy(5, "apple")                                # type-checker rejects
#   ^^^ Argument of type "int" cannot be assigned to parameter "qty"
#       of type "Refined[int, Positive]" in function "buy"

The last call site is the load-bearing one: Refined[int, Positive] and plain int are distinct types under pyright strict and mypy strict in invariant positions. The library's main promise is that you cannot silently bypass the predicate by handing a function a raw value — that is enforced statically, before any code runs.

Pydantic v2 integration

Optional via pip install "refined-types[pydantic]". Importing the adapter installs schema-generation hooks; the core stays free of any pydantic import.

import refined_types.pydantic  # side-effect: install pydantic hooks
from pydantic import BaseModel
from refined_types import Email, NonEmptyString, Positive, Refined

class User(BaseModel):
    name:  Refined[str, NonEmptyString]
    age:   Refined[int, Positive]
    email: Refined[str, Email]

user = User.model_validate({"name": "Ada", "age": 30, "email": "a@b.co"})
# user.age is typed Refined[int, Positive], not int — see tests/test_pydantic_e2e.py

The emitted JSON Schema carries the predicate's constraints:

User.model_json_schema()["properties"]["age"]   # {"exclusiveMinimum": 0, ...}
User.model_json_schema()["properties"]["name"]  # {"minLength": 1, ...}

We instantiate via model_validate({...}) rather than User(name=..., age=30, ...) because the latter would require the caller to hand pydantic an unrefined int for a Refined[int, Positive] field — and the type checker correctly refuses. That refusal is the value prop, not a quirk.

Architecture

refined_types/
├── _core/          ← L1 + L2. Zero deps beyond typing_extensions.
├── pydantic/       ← L4 adapter. Optional via [pydantic] extra.
└── hypothesis/     ← L4 adapter stub. Strategy registration in a later alpha.

The core is framework-agnostic: it works with plain Python, dataclasses, attrs, msgspec, or any modeling layer. Adapters are isolated modules that import the core plus their framework and nothing else. A static test under tests/test_core_isolation.py enforces that _core/ never imports pydantic, hypothesis, attrs, or msgspec.

What ships in 0.1.0a2

• Core: Result[T, E] (Ok/Err discriminated union with match / then / map / map_err), Refined[T, P] phantom marker, ParseError.
• Predicates: Positive, Negative, NonNegative, NonPositive, Zero, NonEmptyString, MinLength.of(n), MaxLength.of(n), Email, NonEmptyCollection, Unique, Sorted. Twelve in total.
• Parse: parse(P, value) -> Result[Refined[T, P], ParseError] and try_parse(P, value) -> Refined[T, P] (raises). Both ship on purpose — parse for total handling, try_parse for ergonomic call sites where a failure is a bug.
• Pydantic v2 adapter (optional extra): refinement-aware __get_pydantic_core_schema__ + JSON Schema constraint emission.
• Hypothesis adapter: stub that raises ImportError with install instructions. Strategy registration lands in a later alpha.

Why not phantom-types?

phantom-types is the incumbent and a good library. We differ on axes that matter for Pydantic-v2-native, type-checker-strict workflows:

Aspect	phantom-types	refined-types (0.1.0a2)
DSL	inheritance	Refined[T, P] generic
Core dependencies	typeguard, numerary	typing-extensions only
Pydantic	v1 hook (legacy)	v2 native, schema-aware
Type-checker distinctness	weak (subclass of T)	strong (distinct in invariant slot)
Parse failure	raises only	parse → Result and try_parse → raise
Predicate algebra	functions: both(p,q)	planned: operators P & Q (0.2.0)
Refined collections safe API	absent	planned (0.2.0)
Σ-type framing	absent	first-class in docs

Honest scope: anything marked "planned" is not in 0.1.0a2.

Σ-type framing — one paragraph

In MLTT/HoTT a Σ-type is Σ(x : T). P(x) — a value of T paired with a proof that P holds. Python has no dependent types, so we keep the shape of Σ rather than its full power: a smart constructor is the only entry, the witness is phantom (erased at runtime, distinct at the type level), and the refinement transports through function signatures. Cross-field dependent constraints, proof-relevant equality, type-level computation — none of that. The discipline Σ encourages — parse-don't- validate, illegal states unrepresentable, transport through types — is exactly what this library gives. See .context/SPEC.md §8 for the longer discussion.

Installation

# Core only — no pydantic, no hypothesis
pip install "refined-types==0.1.0a2"

# With Pydantic v2 integration
pip install "refined-types[pydantic]==0.1.0a1"

# Hypothesis extra installs hypothesis but the adapter is a stub for now
pip install "refined-types[hypothesis]==0.1.0a1"

# Everything
pip install "refined-types[all]==0.1.0a1"

Minimum Python: 3.10.

Roadmap

• 0.1.0a1 — first alpha: core, 12 predicates, pydantic v2 adapter, type-check guards
• 0.1.0a2 — docs alpha (you are here): honest README quickstart, type-checker rejection example, spec carrier rebaselined to drop the hypothesis overclaim
• 0.1.0 (stable) — real Hypothesis strategy registration or explicit defer to 0.3.0, accumulated real-world feedback applied
• 0.2.0 — predicate algebra (&, |, ~), refined collections with safe API (NonEmptyList.head() -> T)
• 0.3.0 — Hypothesis strategy registration (if not in 0.1.0), FastAPI recipes, attrs/msgspec integration examples
• 1.0.0 — stable API, full docs site

License

MIT — see LICENSE.