Paragon ISO Telemetry Harness – First Public Release

This release publishes the Paragon ISO Demo — a cross-device inference telemetry harness that proves deterministic CPU↔GPU behavior across vendors, with reproducible benchmarks.

What's inside

• Iso-demo binary: Go/WebGPU demo for running MNIST across CPU, GPU, and WebGPU.
• Telemetry system: Collects per-digit latencies, CPU vs GPU drift (≤ 1e-6), throughput, and agreement metrics.
• Reports: JSON, Word, and PDF outputs summarizing multiple machines (Intel, AMD, NVIDIA, Apple Silicon).
• Documentation bundle:
  • Paragon AI Runtime.pdf – architectural whitepaper
  • Claude Technical White Paper.pdf – model/code analysis
  • Google AI Framework Analysis.pdf – third-party perspective
  • Grok Report.pdf – external review
  • Distributed ML Infrastructure Testing Framework.docx – telemetry details

Key features

• Vendor-agnostic determinism: Identical outputs on CPU and GPU across OS/hardware.
• Performance telemetry: Average & P90 latencies, GPU/CPU speedup ×, throughput, drift heatmaps.
• Open-source runtime: Built in Go + WebGPU, Apache-2.0 licensed.
• Proof bundle: Papers + raw telemetry reports validating reproducibility.

Quickstart

git clone https://github.com/openfluke/iso-demo
cd iso-demo
go run .

Then in the menu:

1. Press 2 to download MNIST
2. Press 4 to create model zoo
3. Press 5 / 6 to benchmark CPU vs GPU
4. Press 7 to compare drift (CPU vs GPU)

Reports are saved under public/reports/.

---

License

Apache-2.0 (code)
Papers © Samuel Watson / OpenFluke (redistributable for review)