Skip to content

Releases: SpeyTech/certifiable-quant

Certifiable model quantization with formal error bounds and cryptographic certificates.

18 Jan 12:45
@williamofai williamofai
v1.0.0
79aabfc

Choose a tag to compare

View all tags
Certifiable model quantization with formal error bounds and cryptographic certificates. Latest
Latest

v1.0.0 — Initial Release

Certifiable model quantization with formal error bounds and cryptographic certificates.

Highlights

  • Complete quantization pipeline: analyze → calibrate → convert → verify → certificate
  • Formal error bounds: Theoretical analysis with overflow proofs and Lipschitz propagation
  • Cryptographic certificates: Merkle-rooted proof linking all pipeline stages
  • 134 tests passing across 7 test suites
  • Pure C99: Zero dynamic allocation, MISRA-C aligned

Modules

Module Description
dvm/ Fixed-point arithmetic (Q16.16) with fault detection
analyze/ Theoretical error bounds, overflow proofs, operator norms
calibrate/ Runtime statistics, coverage metrics, degenerate detection
convert/ FP32→Q16.16 conversion, BatchNorm folding
verify/ Check quantized values against theoretical bounds
certificate/ Merkle-rooted proof generation and serialization
audit/ SHA-256 hashing primitives

Pipeline: data → training → quant → inference

License

Dual licensed: GPL-3.0 (open source) / Commercial (contact william@fstopify.com)

Built on the Murray Deterministic Computing Platform (MDCP), UK Patent GB2521625.0.

Building deterministic AI systems for when lives depend on the answer.

# v1.0.0 — Initial Release

Certifiable model quantization with formal error bounds and cryptographic certificates.

Highlights

  • Complete quantization pipeline: analyze → calibrate → convert → verify → certificate
  • Formal error bounds: Theoretical analysis with overflow proofs and Lipschitz propagation
  • Cryptographic certificates: Merkle-rooted proof linking all pipeline stages
  • 134 tests passing across 7 test suites
  • Pure C99: Zero dynamic allocation, MISRA-C aligned

Modules

Module Description
dvm/ Fixed-point arithmetic (Q16.16) with fault detection
analyze/ Theoretical error bounds, overflow proofs, operator norms
calibrate/ Runtime statistics, coverage metrics, degenerate detection
convert/ FP32→Q16.16 conversion, BatchNorm folding
verify/ Check quantized values against theoretical bounds
certificate/ Merkle-rooted proof generation and serialization
audit/ SHA-256 hashing primitives

Test Coverage

Suite Tests
test_primitives 8
test_bit_identity 8
test_analyze 30
test_calibrate 28
test_convert 12
test_verify 22
test_certificate 26
Total 134

Build

mkdir build && cd build
cmake ..
make
ctest --output-on-failure

Documentation

  • docs/CQ-MATH-001.md — Mathematical foundations
  • docs/CQ-STRUCT-001.md — Data structure specifications
  • docs/requirements/SRS-001-ANALYZE.md
  • docs/requirements/SRS-002-CALIBRATE.md
  • docs/requirements/SRS-003-CONVERT.md
  • docs/requirements/SRS-004-VERIFY.md
  • docs/requirements/SRS-005-CERTIFICATE.md

Related Projects

Project Description
[certifiable-data](https://github.com/williamofai/certifiable-data) Deterministic data pipeline
[certifiable-training](https://github.com/williamofai/certifiable-training) Deterministic training
[certifiable-inference](https://github.com/williamofai/certifiable-inference) Deterministic inference

Pipeline: data → training → quant → inference

License

Dual licensed: GPL-3.0 (open source) / Commercial (contact william@fstopify.com)

Built on the Murray Deterministic Computing Platform (MDCP), UK Patent GB2521625.0.

Building deterministic AI systems for when lives depend on the answer.

Assets 2
Loading