A research pipeline for comparing hardware entropy sources in embedding space.
It takes raw byte recordings from OpenEntropy, cuts them into short windows, embeds them with OpenAI text-embedding-3-large, and audits what structure survives strict evaluation.
Live results: amentilabs.com/research/entropy-embedding-atlas
Hardware entropy on a single Apple Silicon machine is not a collection of independent streams. 19 out of 58 sources produce nearly identical embeddings despite coming from different hardware subsystems. They share structure because they share host state.
PLL oscillators, I/O timing, network latency, GPU operations, scheduling, microarchitecture, IPC, signal sources, timing, and an external quantum RNG all land within cosine distance 0.01 of each other. The human-assigned categories do not predict which sources cluster together. fsync_journal sits at the center — its timing touches the entire storage stack, making it sensitive to every kind of system load.
Sources that measure CPU internals (preemption_boundary, mach_continuous_timing, icc_atomic_contention) produce genuinely unique byte patterns. preemption_boundary is the most isolated source, sitting 0.41 cosine distance from the average. On 38 sources with enough sessions for leave-one-out evaluation, the model identifies the correct source 46.0% of the time (chance is 2.6%). Stat-distance correlation is 0.80.
Across 4 perturbation campaigns (CPU scheduling, memory/network, GPU/CPU, software), one factor explains 50-80% of cross-source movement. dram_row_buffer responds the most. PLL sources are the most temporally stable (0.003-0.005 drift between sessions). System-level sources like usb_enumeration drift up to 0.29.
The Crypta Labs QCicada USB QRNG sits in the dense core (16 neighbors within cosine distance 0.01), likely because high-entropy hex looks similar regardless of origin. Under stress tests, QCicada barely moves (max shift ~0.01, vs ~0.14 for dram_row_buffer) and its movement direction does not correlate with PLL sources. This is what you would expect from an independent quantum source: consistent high-entropy output that does not respond to host state. The real coupling story is between host-based sources.
- New physics or quantum influence on classical sources
- That visual cluster overlap proves mechanism
- That results generalize beyond this one machine
- That this coupling has cryptographic security implications
openentropy sessions → 256-byte windows → text-embedding-3-large → audits → reports
- Read OpenEntropy sessions from disk (
raw.bin+raw_index.csv). - Cut into 256-byte windows with 128-byte stride.
- Embed with OpenAI
text-embedding-3-large(3,072 dimensions). - Evaluate: retrieval accuracy, clustering, stability, drift, cross-source coupling.
- Generate reports, figures, and data bundles.
git clone https://github.com/amenti-labs/openentropy-embed.git
cd openentropy-embed
python3.12 -m venv .venv
source .venv/bin/activate
pip install -e ".[local,dev]"Requires Python 3.12+. Set OPENAI_API_KEY in .env for the embedding backend.
# Index sessions into windows
openentropy-embed index /path/to/sessions/session-dir --out runs/windows
# Embed with default backend
openentropy-embed embed runs/windows --out runs/embedded
# Evaluate and visualize
openentropy-embed evaluate runs/embedded
openentropy-embed visualize runs/embedded
# Or run the full pipeline in one command
openentropy-embed run /path/to/sessions/session-dir --out runs/fullembed/ Core pipeline (indexing, embedding, evaluation, reporting)
backends/ OpenAI embedding backend
experiments/ Experiment scripts and data bundle exporters
tests/ Regression tests
runs/ Generated artifacts (gitignored)
Expects OpenEntropy sessions on disk with session.json, raw.bin, and raw_index.csv. Reads these directly without requiring OpenEntropy internals.
- OpenEntropy — the recording framework
- Entropy Embedding Atlas — interactive explorer and writeup
- Crypta Labs QCicada — the USB QRNG used as a case study
MIT