Arithmetic garbling

Arithmetic garbling represents values with labels per modulus. Free compatible arithmetic can avoid tables; nonlinear projection reconstructs signed values jointly from a coprime residue bundle and emits output labels. Gate material is shape-bound, strictly serialized, and consumed once.

PLLM's method configurations are clean-room adaptations informed by Garbling Gadgets for Boolean and Arithmetic Circuits, not wrappers around its implementation.

Reference exhaustive correctness, tamper rejection, plan commitments, and one-use process-local issuance checks establish implementation behavior only. The compiler authorizes the bounded primitive from explicit nonlinear and protected-scheduler components; baseline masked-linear components cannot activate it. Cryptographic review, cross-process or durable replay protection, complete-model coverage, and deployment assurance are not recorded.

The stable Python facade exposes method and scheduler configuration, but not raw labels or the arithmetic-garbling evaluator. The evaluator remains internal to the native compiled-plan boundary. See nonlinear methods, schedulers, and research evidence.

API: Nonlinear objects and signatures

Python SDK example

from pllm.nonlinear import ArithmeticGarblingSiluQ7
from pllm.schedulers import BoundedIndependentElementsProtectedTensorSchedule

method = ArithmeticGarblingSiluQ7()
schedule = BoundedIndependentElementsProtectedTensorSchedule(max_elements=128)

assert method.to_spec() == {
    "component": "pllm/arithmetic-garbling-silu-q7/v1",
    "params": {},
}
assert schedule.component == "pllm/bounded-independent-elements/v1"

These are immutable compilation choices. They do not expose or execute one-use evaluator material directly.