sovereign-pulse-vllm 0.1.0

Sovereign Pulse physics gate — vLLM LogitsProcessor plugin

Sovereign Pulse — vLLM Plugin

Physics-grounded token gate for vLLM. Blocks uncertain tokens before sampling using 75 physics laws in 46 ns/token.

Install

pip install sovereign-pulse-vllm

Usage — 2 lines

from sovereign_pulse_vllm import SovereignPulseLogitsProcessor
from vllm import LLM, SamplingParams

gate   = SovereignPulseLogitsProcessor(mass_budget=500.0, verbose=True)
params = SamplingParams(logits_processors=[gate], temperature=0, max_tokens=200)
llm    = LLM(model="meta-llama/Meta-Llama-3-8B-Instruct")

output = llm.generate("The capital of France is", params)
print(output[0].outputs[0].text)
print(gate.stats)
# {'passed': 142, 'blocked': 58, 'total': 200, 'block_rate': 0.29}

Parameters

Parameter	Default	Description
mass_budget	500.0	Session energy budget (higher = less blocking)
confidence_floor	0.02	Minimum truth_index to pass
verbose	False	Print per-token gate decisions

What it blocks

A token is blocked when any physics law fails:

• truth_index = (p1−p2)/p1 below floor — model is coin-flipping
• Hamiltonian H > 1.0 — compute cost exceeds information value
• Rényi entropy tail — heavy-tail uncertainty Shannon misses
• Landauer bound — physically impossible token cost
• Poincaré recurrence — model repeating a previous state

Commercial licensing

MIT license for research and personal use. Commercial deployment license: [e-mail]