ISO C++26 lab for refusal-direction math. No GPU, no Python for the first hour.
1.2 selective research refresh#
- The macOS release lanes use pinned-major Homebrew LLVM 23; Linux and Windows Clang lanes retain LLVM 22.
- Adds
select-layers --scores FILE --count Kwith bounded, deterministic support selection and JSON output. - Adds
apply --mode norm-preserving: rank-one removal plus column-norm restoration, with erased-column rejection. - Adds selective checkpoint planning/export/verification and selected-tensor SFT in the companion Python research toolkit.
- Includes a pinned MiniCPM5-1B researcher example, disjoint data splits, tensor/shard digests, and measured aggregate results.
- Preserves the original numerical modes and August research catalog; adds 14 primary references and an interactive shadcn/ui planner.
- C++ handles numerical operators and plans. Real checkpoint inference, training, and serialization require the documented Python runtime.
1.1 safety and toolchain refresh#
- Certified release lanes now target GCC 16.2, LLVM/Clang 22 (including verified LLVM 22.1.8 on Windows), and MSVC 14.51 on the Visual Studio 2026 image.
- Dimensions, allocation products, ranks, edit strengths, finite values, matrix files, JSONL lines, and record counts fail closed before unsafe work begins.
- Bounds-checked tensor access and owned environment reads replace unchecked indexing and borrowed process-environment pointers at input boundaries.
- Every release build enables platform hardening; Clang 22 ASan+UBSan runs hostile parser/arithmetic tests before publishing.
- Release archives receive GitHub build-provenance attestations in addition to
SHA256SUMS.
10 minutes#
abliterate-cxx guide
abliterate-cxx doctor
abliterate-cxx limits
abliterate-cxx self-check
abliterate-cxx demoHour 0.5 (after demo): estimate dim → apply orba-directional → eval toys → recipes.
Then open GETTING-STARTED.md in the archive (repo: docs/cxx26-researcher-guide.md).
Pick one archive (names do not collide)#
| Archive | Who |
|---|---|
| abliterate-cxx-linux-x64-gcc16.tar.gz | Linux x86_64 (GCC 16.2 default) |
| abliterate-cxx-linux-x64-clang22.tar.gz | Linux x86_64 Clang 22 |
| abliterate-cxx-linux-arm64-gcc16.tar.gz | Linux aarch64 GCC 16.2 |
| abliterate-cxx-linux-arm64-clang22.tar.gz | Linux aarch64 Clang 22 |
| abliterate-cxx-windows-x64-msvc.zip | Windows x64 (default) |
| abliterate-cxx-windows-x64-clang22.zip | Windows x64 Clang 22 |
| abliterate-cxx-windows-arm64-msvc.zip | Windows ARM64 |
| abliterate-cxx-macos-arm64-llvm.tar.gz | Apple Silicon |
| abliterate-cxx-macos-x64-llvm.tar.gz | Intel Mac |
Verify with SHA256SUMS on the release. Each archive was unpacked on the runner and guide + doctor + limits + self-check + demo + estimate/apply/eval ran outside the git checkout. Release archives also carry GitHub artifact attestations.
doctor must print cplusplus=202400. After Hour 0, do Hour 0.5 (estimate dim → apply orba-directional → eval toys → recipes), then Heretic for a real model.
This is a toy-matrix lab, not a Heretic/GGUF substitute.