Integration Examples¶

Purpose¶

The source tree now has a first-party scaffold for optional integration examples under examples/integrations/. These examples are for checkpoint-edit toolchains that want to attach InvarLock regression evidence after producing an edited subject checkpoint.

The integration surface lives in the source tree and keeps optional target dependencies out of the runtime package and core install.

Current Scope¶

Surface	Status
Shared evidence wording	Present under `examples/integrations/_shared/evidence-scope.md`.
Expected artifact checklist	Present under `examples/integrations/_shared/expected-artifacts.md`.
Shared compare wrapper	Present under `examples/integrations/_shared/run_invarlock_compare.sh`.
Public end-to-end handoff	Present under `examples/integrations/public_e2e/`.
CI and registry evidence	Present under `examples/integrations/ci_registry/`.
Target-specific examples	Added one target at a time after backend compatibility is validated.

Browse the integration scaffold in the repository: https://github.com/invarlock/invarlock/tree/staging/next/examples/integrations

Target Example Readiness¶

Target examples should declare one of these labels in their README:

Label	Meaning
`runnable`	Commands are expected to generate `evaluation.report.json`, `verify.json`, and `evaluation.html` in the documented environment.
`reference-pattern`	Commands show how to attach existing reports to CI or registry surfaces without producing a new subject checkpoint.
`exploratory-host`	Commands run with `--execution-mode host --assurance off` for local dependency setup and backend investigation.
`compatibility-investigation`	The external artifact cannot yet be loaded or verified through the documented InvarLock path; the README records the blocker.

Preflight¶

Before adding or publishing a target example:

invarlock doctor
invarlock advanced plugins list --json

For optional backend targets, also check the Python module:

python -c "import importlib.util; print(importlib.util.find_spec('gptqmodel') is not None)"

Replace gptqmodel with the backend module for the target example.

Shared Workflow¶

For a baseline and subject that are already loadable through an InvarLock adapter:

examples/integrations/_shared/run_invarlock_compare.sh \
  --baseline sshleifer/tiny-gpt2 \
  --subject ./models/tiny-gpt2-subject \
  --report-out ./reports/integration-smoke \
  --allow-network

The generated local output should include:

Artifact	Role
`evaluation.report.json`	Canonical verifier input.
`verify.json`	Machine-readable verifier result.
`evaluation.html`	Human-readable report.
`runtime.manifest.json`	Runtime provenance for strict container-backed runs.
`run_command.txt`	Wrapper invocation plus evaluate, verify, and render commands.

Generated reports, models, runs, HTML, and artifacts under examples/integrations/** are ignored by git.

CI and Registry Handoff¶

For a source-only public walkthrough, run the script under examples/integrations/public_e2e/. It copies the checked-in external-edit public report into a local output directory, verifies that local copy, renders HTML, exports MLflow tags, writes a model-card block, writes a release-review packet, and writes a CI summary Markdown file.

Use invarlock report export when the report already exists and you need to attach the evidence to systems users already operate:

invarlock report export \
  --evaluation-report reports/eval/evaluation.report.json \
  --format mlflow-tags \
  --policy-profile ci \
  --verify-result reports/eval/invarlock-verify.json \
  --output reports/eval/mlflow-tags.json

invarlock report export \
  --evaluation-report reports/eval/evaluation.report.json \
  --format model-card-md \
  --report-url https://example.test/evaluation.report.json \
  --evidence-url https://example.test/evidence.zip \
  --verify-result reports/eval/invarlock-verify.json \
  --output reports/eval/model-card-invarlock.md

invarlock report export \
  --evaluation-report reports/eval/evaluation.report.json \
  --format release-review-md \
  --policy-profile release \
  --verify-result reports/eval/invarlock-verify.json \
  --output reports/eval/release-review.md

For GitHub Actions, the repository ships a composite action at .github/actions/invarlock-report-gate. It verifies the evaluation report, renders HTML, writes MLflow tags, writes a release-review packet, appends a PR summary, and uploads the evidence as a workflow artifact.

Important action inputs include report, profile, assurance, runtime-provenance, warning-policy, verify-output, html-output, mlflow-tags-output, and review-output. The generated exports consume the same verify-output JSON so their pass/fail status reflects the verifier result item for the same evaluation report when the action has run verification.

This integration layer is intentionally small. It exports InvarLock evidence for CI and downstream handoff into model cards, registries, MLflow, Hugging Face Hub, Databricks, SageMaker, Vertex, or an organization's approval workflow.

Public Evidence Anchors¶

Use the shipped public evidence as the credibility floor while target examples are being built:

invarlock verify --profile release --assurance strict \
  public_evidence/published_basis/gpt2/evaluation.report.json

invarlock verify --profile release --assurance strict \
  public_evidence/byoe_examples/lora_merge_byoe/evaluation.report.json

invarlock verify --profile release --assurance strict \
  public_evidence/byoe_examples/fine_tune_byoe/evaluation.report.json

invarlock verify --profile release --assurance strict \
  public_evidence/real_runs/tiny_gpt2_quant_rtn/evaluation.report.json

invarlock verify --profile release --assurance strict \
  public_evidence/real_runs/tiny_gpt2_external_magnitude_prune/evaluation.report.json

See the Public Evidence Walkthrough for the full artifact taxonomy.