Skip to content

Reproducible package checks

Run these from a reviewed Git checkout with the locked development environment:

make package-check
make package-check-train
make package-check-guard

The first gate installs only core dependencies; it requires scikit-learn to be absent and runs all installed runtime/CLI fixtures. The second additionally creates a separate train environment, executes the synthetic fit and caller-pinned replay, and rejects an altered result. There is no skip path for installed training. Both use one BLAS/OpenMP thread. No model or research-data download occurs. The guard gate adds a separate environment for the optional ONNX/tokenizer runtime and executes a tiny authored graph, with complete chunk coverage. Linux/macOS CPython 3.12 CI exercises this additional gate without research assets.

Inputs and environment

Use uv 0.9.18 and a supported CPython 3.12, 3.13 or 3.14 interpreter. Linux ubuntu-24.04 and macOS macos-15 CI run each Python version's core/package and required-train/package contracts. A runner label is not an immutable OS image; the actual interpreter, OS/architecture and dependency inventory are recorded. No Windows support or cross-platform byte equality is claimed.

Hatchling 1.32.4, packaging 26.3, pathspec 1.1.1, pluggy 1.7.0, tomlkit 0.15.2 and trove-classifiers 2026.9.21.13 are exact in build-system.requires, the build dependency group and uv.lock. The gate exports hash-pinned requirements, creates one fresh build environment and installs only locked wheels. It verifies the actual installed inventory against the lock. Builds then run offline with build isolation disabled because that environment has already been provisioned. No inherited PYTHONPATH, UV_* or HATCH_* settings select another toolchain. Package installation may fetch locked software wheels; this is not an OS network sandbox for arbitrary backend code.

SOURCE_DATE_EPOCH=315532800 and PYTHONHASHSEED=0 are fixed by the gate. Hatch documents epoch-based reproducibility, and uv documents the build flow. The report records the interpreter binary, ABI, zlib runtime, uv binary and actual build package hashes. Reproduction means identical artifacts under those recorded inputs; a different environment must run its own comparison.

Reviewed source and archive boundaries

The source snapshot uses tracked or newly staged Git paths beneath the explicit sdist roots, with their current working bytes and Git file modes. Review and stage exact new files before checking. Untracked files are not silently admitted; local runs, credentials and the ignored agent harness stay outside the snapshot. Dirty tracked source is allowed for development, so the source-tree hash is the identity of the checked input; the recorded HEAD alone is not that identity.

Two new source directories receive the same bounded inventory, normalized mtimes and empty VCS boundary directories. The latter prevent Hatch from importing a parent directory's .gitignore or .hgignore; no Git history is copied. Two wheel/sdist builds must have identical SHA-256 values. The sdist's contents are independently checked, then rebuilt into the same wheel without the original checkout. Installation checks use fixtures/scripts from that extracted sdist.

Archives must contain exactly the reviewed source plus generated package metadata, or the exact runtime code/type files and wheel metadata/licence. Missing or extra files, changed payloads, incorrect wheel RECORD hashes, duplicate/case-aliasing paths, traversal, links and nonregular members fail. Source aliases and changes detected during snapshotting also fail. Input limits are 2,048 files, 8 MiB per file and 64 MiB of aggregate file content/archive bytes. These are parser bounds, not a peak-RSS or adversarial process-isolation guarantee.

Known private/generated paths and recognizable PEM, GitHub, AWS and Hugging Face credential patterns reject. This is a defence in addition to reviewed Git paths and exact inventory matching; no automated scanner can certify arbitrary content as non-confidential. Public synthetic fixture keys/weights remain fixtures and do not grant operational authority. Audit source changes before sharing artifacts.

Retained results and failures

Every invocation creates a fresh private directory under dist/package-check-*. It retains source inventory/hash, two source copies, build environment/inventory, requirements, raw command logs, both distributions, extracted source, installed core/train environments and the structured result. Failed invocations retain failure.json and available partial outputs; never reuse a failed directory. The final output identifies the retained directory and artifact hashes.

These are local build records, not uploaded CI artifacts or a signed release packet. They may contain paths and private operator context in command logs. Keep them outside git and follow the owner's local retention policy. Nothing publishes to a package index, adds reviewers, or changes repository visibility. The source, dependency and distribution inventories support engineering review; release approval, independent verification and scientific/hardware evidence remain separate gates. See licences and decision 0016.