Repository navigation
Product readiness follow-up from AITG portfolio audit #66
Description
Context
AITG completed a product audit of compressionKIT.
Agreed decisions
- Toolkit/source stays pre-v1 on 0.2.x; each Hugging Face bundle has independent v1.0+ versions with compatibility metadata.
- Future Python distribution/import is
compression-kit/compressionkit. - Toolkit source is BSD-3-Clause; neural weights/codebooks are Ambiq-production-only; permissive DSP retains Apache/BSD; hybrid bundles require component-level license manifests.
Checklist
- Align
compression-kitdistribution,compressionkitimport, README, and install examples. - Publish package or clearly keep source install until publication.
- Generate one complete Hugging Face bundle inventory from canonical manifests.
- Link toolkit/export schema versions from every independently versioned bundle.
- Publish component-level code/weights/codebook/data/license/provenance manifests.
- Preserve historical MESA lineage for old bundles and current MESA-free lineage for new goldens.
- Define versioned scorecard acceptance thresholds by modality/method/ratio.
- Verify checksums, reference vectors, LiteRT/C-header artifacts, and downstream integration.
- Keep codec bundles separate from downstream clinical inference claims.
Done when
Package identity, dual versioning, artifact inventory, provenance, licensing, and validation are machine-readable and consistent.
Activity
apage224 commentedon Oct 6, 2026
Proceeding with the authorized corrected RVQ release after #65 and #75 merged. Scope: repackage the 11 saved single-stage RVQ checkpoints from main without retraining; validate synthetic sample artifacts, independent trained-model parity, complete packets, precision reports and checksum integrity; publish new v1.1 repositories while preserving v1.0; update canonical and downstream release references.
Release inspection found the current RVQ publisher drops reference vectors/checksums and stages renamed files without a complete canonical artifact contract. This must be corrected before publication. Historical physiological scorecards will be explicitly identified as historical; two-stage entropy priors are outside this release.
apage224 commentedon Oct 6, 2026
All eleven corrected RVQ v1.1 bundles have been published and downloaded by immutable revision. Every payload matched its published checksum manifest; all local, staged and downloaded packages passed strict validation with ten independent training-reference frames each. The original local checkpoint/deploy hashes and v1.0 repository heads remain unchanged.
Evidence inventory: https://github.com/AmbiqAI/compressionkit/blob/9a3b078/releases/rvq-v1.1.json
440 complete-packet round trips covered all four encoder precisions; the downstream vendored runtime passed another 44 round trips against all eleven published revisions. All eleven codebook headers compile as C99. Full source tests: 445 passed, 78 optional artifact/data skips.
Source/publication/reference updates: #76. Downstream selection/pinning updates: https://github.com/AmbiqAI/compressionkit-dev/pull/2. Remaining product-readiness items in this issue are still open; no retraining, physiological requalification, MCU qualification or entropy-prior release was performed.