Model Data Language

K.1

The problem

A raw model call takes a string and returns a string.

response = llm.chat("Extract title, author, date from this article")
# "The title is 'AI Revolution', written by John Smith on March 15, 2024"

The caller is left to parse prose: extract fields from natural language, absorb format drift ("Title:", "The title is..."), decide whether "March 15" is a date or a string, handle missing fields, catch malformed responses. It is fragile and it does not scale.

K.2

The contract

MDL replaces parsing with a contract. Schemas are Python dataclasses. MDL compiles them to model-safe specifications and validates every payload.

1. Define the schema in Python
2. MDL compiles it to a model specification
3. The model outputs conforming JSON
4. MDL validates and deserializes to a typed object

from dataclasses import dataclass
from typing import List

@dataclass
class ArticleInput:
    text: str

@dataclass
class ArticleOutput:
    title: str
    author: str
    date: str
    tags: List[str]

# guaranteed valid ArticleOutput, or an error
result = block(
    model_id="Qwen/Qwen3-4B",
    input=ArticleInput(text="..."),
    output=ArticleOutput,
)

print(result.title)  # typed str
print(result.tags)   # typed List[str]

K.3

Compilation and validation

input schema  → prompt specification
output schema → JSON structure specification + validation rules

The model receives clear structural requirements. MDL enforces them on every response before returning.

K.4

Type system

@dataclass
class Address:
    street: str
    city: str

@dataclass
class Person:
    name: str
    address: Address  # nested dataclass

K.5

Inside Conduit

MDL is integrated directly into Conduit's block interface.

from conduit.runtime import LMLiteBlock

block = LMLiteBlock(models=[...])

# option 1: typed MDL interface
result = block(
    model_id="Qwen/Qwen3-4B",
    input=MyInput(query="..."),
    output=MyOutput,
)

# option 2: OpenAI-style messages, no MDL
result = block(
    model_id="Qwen/Qwen3-4B",
    messages=[{"role": "user", "content": "Hello"}],
)

Both interfaces work. MDL provides type safety; messages provide flexibility.

K.6

Error handling

try:
    result = block(input=inp, output=OutputSchema)
except ValidationError as e:
    # schema violation: missing field, wrong type
    print(e.field, e.expected, e.received)
except ParseError as e:
    # malformed JSON from the model
    print(e.raw_response)

For transient failures, MDL retries with correction prompts.

result = block(
    input=inp,
    output=OutputSchema,
    max_retries=3,
)

K.7

Specifications