invarlock 0.5.1

Edit‑agnostic robustness evaluation reports for weight edits (InvarLock framework)

Edit‑agnostic robustness reports for weight edits

Catch silent quality regressions from quantization, pruning, and weight edits before they ship.

Quantizing, pruning, or otherwise editing a model’s weights can silently degrade quality. InvarLock compares an edited subject checkpoint against a fixed baseline with paired evaluation windows, enforces the canonical guard chain (invariants → spectral → RMT → variance → invariants), and produces a machine-readable evaluation report you can gate in CI.

Why InvarLock?

• Quality gates for weight edits: catch regressions before deployment.
• Statistical guarantees: paired primary metrics with confidence intervals.
• Auditable evidence: deterministic pairing metadata + policy digests in evaluation.report.json.
• CI/CD-friendly: stable exit codes, --json outputs, and portable “proof packs”.
• Offline-first: network is disabled by default; enable downloads per command.

Who is this for?

• ML engineers shipping quantized/pruned checkpoints.
• MLOps teams building CI quality gates and reviewable artifacts.
• Researchers validating compression/edit methods with reproducible, paired eval.

How it works

┌───────────────────────┐     ┌────────────────────────────────────────────┐
│ Baseline (checkpoint) │────►│                                            │
└───────────────────────┘     │  invarlock evaluate                        │
                              │  ├─► Paired windows (deterministic)        │
┌───────────────────────┐     │  ├─► GuardChain pipeline                   │
│ Subject  (checkpoint) │────►│  │   └─► invariants → spectral → RMT → VE  │
└───────────────────────┘     │  └─► Emit: evaluation.report.json          │
                              │                                            │
                              └────────────────────────────────────────────┘
                                                     │
                                     ┌───────────────┴───────────────┐
                                     ▼                               ▼
                                 ✅ PASS                          ❌ FAIL
                                 (ship)                          (rollback)

Quick start

Colab (CPU-friendly):

The secure-default CLI path runs model-loading commands inside the runtime container and expects an OCI container engine such as docker or podman. In a repo checkout, build the local runtime image once with make runtime-image; InvarLock automatically prefers invarlock-runtime:local when it is present. Trusted local workflows can opt into host execution explicitly with --mode local on invarlock evaluate, but the attested verification step below expects container execution. The quickstart block below assumes a repo checkout; do not skip make runtime-image if you want the attested container path.

# Repo-checkout quickstart for the attested container path
# HF adapter stack (torch/transformers)
pip install "invarlock[hf]"

# Required in a repo checkout for the attested path; do not skip this step.
make runtime-image

# Version + report schema (when available)
invarlock --version

# Compare baseline vs subject (downloads require explicit network enable)
# Secure-default execution uses the runtime container and writes
# reports/eval/runtime.manifest.json next to evaluation.report.json.
INVARLOCK_ALLOW_NETWORK=1 invarlock evaluate \
  --baseline gpt2 \
  --subject  distilgpt2 \
  --adapter auto \
  --profile dev \
  --report-out reports/eval \
  --quiet

# Validate the attested evaluation report
test -f reports/eval/runtime.manifest.json
invarlock verify --json reports/eval/evaluation.report.json

# Render HTML for sharing
invarlock report html -i reports/eval/evaluation.report.json -o reports/eval/evaluation.html

If you pass a directory to invarlock report, it must contain canonical report.json or evaluation.report.json; other report-like filenames are not auto-selected.

Example output (abridged; counts vary by profile/config):

INVARLOCK v<version> · EVALUATE
Baseline: gpt2 -> Subject: gpt2 · Profile: dev
Status: PASS · Gates: <passed>/<total> passed
Primary metric ratio: <ratio>
Output: reports/eval/evaluation.report.json
Attestation: reports/eval/runtime.manifest.json

Command Surface

• Core workflow: invarlock evaluate → invarlock verify → invarlock report html.
• Advanced workflows live under invarlock advanced ....
• Trusted host execution for the core evaluate path uses --mode local.
• Optional adapter/backend installs use normal Python extras such as pip install "invarlock[hf]" rather than CLI install commands.

Proof packs (portable evidence bundles)

Proof packs bundle reports + verification metadata into a distributable artifact.

• Guide: https://github.com/invarlock/invarlock/blob/main/docs/user-guide/proof-packs.md
• Verify from an installed wheel: invarlock advanced proof-pack verify <dir> --strict
• Repo harness alternative: scripts/proof_packs/verify_pack.sh --pack <dir> --strict

Note: configs/ and most scripts/ remain repo resources and are not shipped in wheels. Installed wheels include the public contracts and the invarlock advanced proof-pack verify verifier.

Installation

# Minimal CLI (no torch/transformers)
pip install invarlock

# HF workflows (torch/transformers)
pip install "invarlock[hf]"

Optional extras: invarlock[gpu], invarlock[awq,gptq]. Full setup: https://github.com/invarlock/invarlock/blob/main/docs/user-guide/getting-started.md.

Documentation

• Quickstart: https://github.com/invarlock/invarlock/blob/main/docs/user-guide/quickstart.md
• Compare & evaluate (BYOE): https://github.com/invarlock/invarlock/blob/main/docs/user-guide/compare-and-evaluate.md
• Reading a report: https://github.com/invarlock/invarlock/blob/main/docs/user-guide/reading-report.md
• CLI reference: https://github.com/invarlock/invarlock/blob/main/docs/reference/cli.md
• Assurance case: https://github.com/invarlock/invarlock/blob/main/docs/assurance/00-assurance-case.md
• Threat model: https://github.com/invarlock/invarlock/blob/main/docs/security/threat-model.md

Community

• Questions/ideas: https://github.com/invarlock/invarlock/discussions
• Bug reports: https://github.com/invarlock/invarlock/issues
• Contact: mailto:[e-mail]

Citation

If you use InvarLock in scientific work, please cite it (canonical metadata is in CITATION.cff):

@software{invarlock,
  title  = {InvarLock: Edit-agnostic robustness evaluation reports for weight edits},
  author = {{InvarLock}},
  url    = {https://github.com/invarlock/invarlock},
}

Limitations

• InvarLock evaluates an edited model relative to a baseline under a specific configuration; results are not “global” guarantees.
• Not a content-safety/alignment tool.
• Native Windows is not supported (use WSL2 or Linux).

Support matrix

Platform	Status	Notes
Python 3.12+	✅ Required
Linux	✅ Full	Primary dev target
macOS (Intel/M-series)	✅ Full	MPS supported (default on Apple Silicon)
Windows	❌ Not supported	Use WSL2 or a Linux container if required
CUDA	✅ Recommended	For larger models
CPU	✅ Fallback	Slower but functional

Project status

InvarLock is pre‑1.0. Until 1.0, minor releases may include breaking changes. See CHANGELOG.md.

For guidance on where to ask questions, how to report bugs, and what to expect in terms of response times, see SUPPORT.md.

Contributing

• Contributing guide: https://github.com/invarlock/invarlock/blob/main/CONTRIBUTING.md
• Fast local checks (repo clone):
  • make targets auto-select Python 3.12+, preferring an active 3.12 env, python3.12, then the Conda env invarlock-py312 when present.
  • make dev-install
  • make test
  • make lint
  • make docs-live

License

Apache-2.0 — see LICENSE.