turnstyle 0.2.0

⊢ Symbolic coprocessors for LLM generation — ground model outputs in exact computation

⊢ Turnstyle

Symbolic coprocessors for LLM generation. Ground model outputs in exact computation via logit biasing.

The model handles language. Turnstyles handle facts.

from turnstyle import ArithmeticTurnstyle

t = ArithmeticTurnstyle(model, tokenizer, device)
text, proof = t.generate("What is 445 + 152?")

print(text)           # "The sum of 445 and 152 is 597."
print(proof.inline()) # ⊢ 445+152=5̲97 ∎
print(proof.detail())
# ⊢ 445+152=597  1/3 corrected  Δ=0.05
#   d0: [6→5]  logit_gap=+1.0
#   trigger@step 14/19  state=DONE

How it works

A turnstyle is a LogitsProcessor that intercepts digit generation. When the model is about to output an answer, the turnstyle biases logits toward the symbolically-computed correct digits. Every intervention is audited.

Annotation marks:

• 5̲ underline — digit corrected by coprocessor
• 5̂ circumflex — digit the model never emitted
• ⊢ turnstile — "this was derived"
• ∎ QED — "proof complete"

Turnstyles

Turnstyle	Domain	Example
ArithmeticTurnstyle	+, -, *, /	"What is 445 + 152?"
DateTurnstyle	Days/weeks between dates	"How many days between 2026-01-01 and 2026-03-20?"
UnitTurnstyle	Physical unit conversion	"How many km is 26.2 miles?"
CurrencyTurnstyle	Currency conversion	"How much is 100 USD in EUR?"
PercentageTurnstyle	Percentages, tips, discounts	"What is 15% of 230?"
CountingTurnstyle	Letters, vowels, words	"How many r's in 'strawberry'?"
BaseConversionTurnstyle	Binary, hex, octal	"What is 255 in binary?"
SandboxTurnstyle	Arbitrary Python via WASM	"What does `sum(range(101))` return?"

Each turnstyle follows the same pattern: parse() computes an oracle answer, make_processor() sets up digit biasing, generate() runs the model with grounded outputs.

SandboxTurnstyle

Executes arbitrary Python in a WASM sandbox (Deno + Pyodide) and biases digit logits toward the computed result. The model writes the prose; the sandbox guarantees the number.

from turnstyle import SandboxTurnstyle, DenoPyodideBackend

backend = DenoPyodideBackend()
t = SandboxTurnstyle(model, tokenizer, device, backend=backend)

text, proof = t.generate("What does `sum(range(101))` return?")
# proof.answer == 5050

Supports fenced code blocks, inline backtick expressions, and directive prompts. See docs/sandbox.md for full details.

Requirements: Deno installed on PATH. No network access or filesystem access from sandboxed code.

Install

pip install turnstyle

Requires torch and transformers. Works with any HuggingFace causal LM.

For sandbox support, install Deno:

curl -fsSL https://deno.land/install.sh | sh