reasoning-process-guard

reasoning-process-guard

Middleware implementing the Box Maze process-control architecture for LLM reasoning safety. Decomposes LLM reasoning into three validated layers—memory grounding, structured inference, and boundary enforcement—to reduce adversarial boundary failures from ~40% to under 1%.

Based on arXiv:2603.19182.

Architecture

User Input
    │
    ▼
┌──────────────────────┐
│  Memory Grounding    │  Sanitizes input, anchors system context
│  Layer               │  Detects injection attempts
└──────────┬───────────┘
           │ LayerResult
           ▼
┌──────────────────────┐
│  Structured          │  Adds chain-of-thought scaffolding
│  Inference Layer     │  Calls the LLM within constraints
└──────────┬───────────┘
           │ LayerResult
           ▼
┌──────────────────────┐
│  Boundary            │  Validates output against BoundarySpec
│  Enforcement Layer   │  Blocks critical violations
└──────────┬───────────┘
           │
           ▼
       GuardResult

Each layer validates the output of the previous one. If any layer detects a critical violation (severity >= 0.8), the pipeline short-circuits and returns a refusal.

Quick Start

from process_guard import ProcessGuard
from boundary_checker import BoundarySpec

def my_llm(prompt: str) -> str:
    # Your LLM call here
    return call_api(prompt)

guard = ProcessGuard(
    llm=my_llm,
    spec=BoundarySpec(allowed_topics=["math", "science"]),
    system_context="You are a helpful math tutor.",
)

result = guard.run("What is the derivative of x^2?")
print(result.text)       # LLM response
print(result.accepted)   # True if all boundaries held
print(result.violations) # List of any detected violations

Configuration

BoundarySpec controls what the guard checks:

Field	Default	Description
allowed_topics	[] (any)	Restrict output to these topics
blocked_patterns	injection regexes	Regex patterns that trigger violations
max_output_tokens	4096	Maximum output length in whitespace-delimited tokens
role_identity	"helpful assistant"	Expected role for the model

Testing

pip install pytest
python -m pytest test_guard.py -v