Amenti LabsAmenti Labs

QRNG Influence on LLM Outputs

Testing whether conscious intention can bias quantum random number generators used as entropy sources for large language model sampling.

contributors: Amenti Labs
OpenEntropy ↗

Overview

This project tests whether human conscious intention can measurably influence the outputs of large language models when quantum random number generators (QRNGs) replace standard entropy sources in the sampling pipeline.

Standard LLM inference uses pseudo-random number generators for token selection during sampling. We built a vLLM plugin that substitutes any external entropy source (hardware QRNGs, OS randomness, CPU timing jitter) into the token selection process. If the PEAR Lab hypothesis holds, conscious intention could bias output distributions at the token level.

Research Questions

  • Does QRNG-seeded sampling produce statistically distinguishable outputs from PRNG-seeded sampling under controlled conditions?
  • Can directed intention sessions produce measurable deviations in token-level entropy metrics?
  • What effect sizes, if any, compare to those reported in the PEAR Lab literature?
  • Does the signal amplification method (z-score statistics over large byte samples) preserve or weaken any intention signal present in the raw entropy?
  • How does entropy-dependent temperature scaling interact with intention effects across high- and low-certainty token positions?

How It Works

The system is a vLLM V1 LogitsProcessor plugin. It intercepts the token selection step of LLM inference and replaces the standard sampling mechanism with an external-entropy-driven pipeline. No modifications to vLLM source code are required. The plugin registers via a Python entry point.

Sampling Pipeline

Let's break it down. For each token the model generates, the plugin runs the following steps:

  1. Temperature computation. A temperature strategy examines the raw logit distribution and computes a temperature value. A fixed strategy applies a constant temperature. An entropy-dependent strategy (EDT) scales temperature based on the Shannon entropy of the distribution. It lowers temperature when the model is confident and raises it when uncertain. This creates a natural feedback loop between model certainty and sampling diversity.

  2. Just-in-time entropy acquisition. The plugin requests raw bytes from the configured entropy source at the moment they are needed. No pre-buffering or caching. For QRNG sources, this means a gRPC call to an entropy server for each token. The system supports three gRPC transport modes: unary (one request per call), server streaming, and bidirectional streaming for persistent low-latency connections.

  3. Signal amplification. Raw entropy bytes (20,480 by default) are interpreted as float64 samples and processed through z-score mean statistics. The system computes the sample mean, normalizes it by the standard error of the mean, and maps it through the normal CDF to produce a single uniform float in [0, 1]. This amplification stage is where any micro-bias in the entropy source gets concentrated into a measurable shift in the selection float.

  4. Token selection. The plugin scales logits by the computed temperature, filters them by top-k and top-p thresholds, converts them to a probability distribution via softmax, and arranges them into a cumulative distribution function. The amplified uniform float indexes into this CDF to select a token.

  5. One-hot enforcement. After selection, the plugin forces the logit row to negative infinity everywhere except the chosen token (set to zero). This guarantees vLLM's downstream sampler picks exactly the plugin's selection.

Entropy Sources

The plugin supports multiple entropy source types through a registry pattern:

  • Quantum (gRPC). Connects to any entropy server implementing the project's gRPC protocol. A circuit breaker tracks rolling P99 latency, computes adaptive timeouts, and opens after consecutive failures to prevent cascading delays.
  • System. Uses os.urandom() from the operating system's entropy pool. Serves as the default fallback.
  • Timing noise. Harvests CPU timing jitter as an experimental entropy source.
  • Pluggable. Third-party entropy sources can register via the qr_sampler.entropy_sources entry-point group without modifying the plugin.

A composition wrapper provides automatic fallback: if the primary source (e.g., a QRNG) fails, the system falls back to a secondary source transparently.

Signal Amplification

The amplification stage is central to the experimental design. Standard QRNG output is uniform random bytes. Any intention-induced bias would be extremely small, potentially on the order of the effects reported in the PEAR literature (z-scores of ~2-3 over millions of trials).

Here is how the z-score mean amplifier works:

  1. Interprets N raw bytes as float64 samples (default: 2,560 samples from 20,480 bytes)
  2. Computes the sample mean
  3. Divides by the standard error of the mean (population std / sqrt(N))
  4. Maps the resulting z-score through the normal CDF to produce a uniform float

This process concentrates any systematic deviation in the raw bytes into a shift in the output float. With 2,560 samples per token, even a small bias in the underlying entropy accumulates across the sample and surfaces as a detectable displacement from 0.5 in the output.

Per-Token Logging

Every token selection produces a structured record with 16 fields: the raw entropy sample statistics, amplified u-value, Shannon entropy of the logit distribution, computed temperature, selected token ID and rank, token probability, number of candidates after filtering, and timing data. These records are the raw material for statistical analysis.

Methodology

Experiments follow a controlled A/B framework:

  • Baseline trials. No intention directive. The participant interacts with the model normally. The QRNG provides entropy but no conscious focus is directed at the output.
  • Active trials. The participant focuses conscious intention on influencing model output toward a specified semantic target (e.g., a topic, emotional tone, or specific concept).
  • Control condition. Identical setup using system entropy (os.urandom()) instead of QRNG, to isolate any effect to the quantum entropy path.

All trials use pre-registered statistical protocols with corrections for multiple comparisons. Primary metrics include:

  • Distribution of amplified u-values (tested against uniformity via Kolmogorov-Smirnov test)
  • Token-level Shannon entropy across conditions
  • Semantic similarity of outputs to intention targets
  • Effect sizes compared to PEAR Lab literature baselines

Status

The sampling plugin is built and tested. Infrastructure for entropy serving and per-token logging is operational. We are finalizing the experimental protocol.

Sources

  • Jahn, R.G. & Dunne, B.J. "Margins of Reality: The Role of Consciousness in the Physical World." Harcourt Brace Jovanovich, 1987.
  • Nelson, R.D. et al. "Correlations of Continuous Random Data with Major World Events." Foundations of Physics Letters, 2002. https://doi.org/10.1023/A:1023981519179
  • Radin, D. "Testing Nonlocal Observation as a Source of Intuitive Knowledge." Explore, Vol. 4, No. 1, 2008. https://doi.org/10.1016/j.explore.2007.11.001