The Specifications
The architecture and math behind the canonicalization layer.
1. The Divergence Baseline
Everything else on this page is measured against this one number.
Five environments produced three different sums, spread across 1.25390625. An Intel Xeon and an AMD EPYC agree; two Intel Xeons disagree. The vendor on the chip does not predict the answer.
Matching hashes mean the two machines behaved identically on this test. They do not prove your fleet is safe, and they do not mean the hardware is the same: Result C contains both an Intel and an AMD chip. This test does not isolate the cause. For path-dependent divergence on GPUs, run the GPU harness.
2. CAGM (Deterministic ML Training)
Standard ML training pipelines rely on floating-point math, meaning the exact same dataset and hyperparameters will produce different model weights depending on the GPU or CPU used.
CAGM is a fixed-point integer classifier (Q16.16) that completely removes floating-point arithmetic from the training loop.
- The Integrity Engine (IE): Canonicalizes logits at layer boundaries to ensure bit-exact hashes across compilers and hardware.
- Trajectory Proof: Emits a 64-character SHA-256 proof that binds the weights, gradients, activations, and epoch order into a single cryptographic commitment.
- Output Contract: Strict stdout output:
accuracy,proof, andie_verify.
3. Test Suites & Validation
- SPCMP: 42/42 adversarial assertions passed (NaNs, subnormals, extreme boundaries).
- CAGM: 41/41 verification assertions passed (determinism, overflow guards, adversarial data).
- CAGM vs Sklearn: Benchmarked against 6 standard ML models across 5 datasets, maintaining 100% deterministic proof matching.
Test Reports
Primary source documents. Not marketing copy.
| Source | Scope | Link |
|---|---|---|
| float-divergence-script | Sorted-sum experiment, 5 environments, 3 distinct results. Runs anywhere Python runs. | Repo · Script |
| prove-model-drift | Same PyTorch script, same seed, 4 environments: identical final loss, 4 different weights hashes | Repo · Script · Page |
| gpu-float32-divergence | GPU harness: FMA, fast-math and atomic-reduction divergence on Tesla T4 and RTX 3050, 1,000,000-element corpus | Repo |
| Report | Scope | Link |
|---|---|---|
| CAGM Verification Report 41/41 PASS | Determinism, IE canonicalisation, degenerate inputs, overflow guards | |
| CAGM vs sklearn | 6 sklearn baselines, 5 datasets, stratified k-fold CV, ctypes-native library | |
| CAGM Cross-Device Benchmark | Accuracy and proof-hash comparison, 4 machines, CPU/GPU/TPU -- all proofs identical |
| Report | Scope | Link |
|---|---|---|
| PCMP / SPCMP Verification Report 42/42 PASS | Every IEEE-754 edge case and degenerate array shape, checked against an independent Python oracle | |
| SPCMP Cross-Hardware Verification | Google CPU, TPU, GPU, and a native GPU node -- identical Merkle root and hash on all four | |
| Cross-Machine Determinism | 20 sort runs, 4 heterogeneous machines, 5 files each -- zero divergent hashes | |
| SPCMP Determinism Report | Top-k / bottom-k convergence; documents the honest scalar-reduction limit | |
| Aerospace Verification Proof | 40/40 assertions -- radiation bit-flip absorption, telemetry compression, GNC-loop determinism | |
| Scale Test Report | 32 runs, 100k to 1B float32 elements, 8 data patterns -- zero failures | |
| Samayuktam Sorting Benchmark | CPU/GPU sort timing vs NumPy quicksort/mergesort/stable; 4.3–4.8× on uniform and normal at 100M elements, up to 27× on duplicate_heavy_tail |