Classify and document
Risk tier, model card, training-data summary, license and SBOM, copyright policy.
Ethics: A file a peer or a regulator can read.
Theater: A wizard whose output never leaves the laptop.
Compliance
Four jobs. Not one dashboard.
Ethical compliance tools get sold under a single label. Mixing the jobs is how operators buy a dashboard and think they have a duty. A license file is not governance. A vendor stamp is not a notified body.
The test is simple. Does the tool run on machines you control? Does it leave evidence you can hash? Does it fail a release when it should?
Jobs
Pick by job, not by logo
Most mature shops use one tool per row. There is no best tool across rows.
Probe before release
Jailbreak, injection, leakage, dual-use uplift, eval gates in CI.
Ethics: A merge is blocked on a score.
Theater: One scanner run before a blog post.
Govern at runtime
Policy on input, output, and actions before the agent does them.
Ethics: Fail-closed, logged, signed, on machines you control.
Theater: A chat filter the user can strip.
Prove after the fact
Hash-chained logs, traces, incident pack, mapped controls.
Ethics: An auditor can replay a decision.
Theater: A badge that says compliant on a pre-1.0 README.
Shop kit
A defensible minimum
Desk can skip the gateway. Plant adds signed policy packs and an incident drill. Shop is the missing public default.
- Risk note, model card, and license or SBOM scan in the repo.
- A vulnerability suite and a task suite in CI. Fail the build.
- A gateway policy file you can read — allow and deny tools, budgets, topics — in front of the local engine.
- Traces and a hash-chained decision log on disk you own.
That is cheaper than a platform and harder to fake. Names of scanners and rails change. The contract does not. See Operators for desk, shop, and plant.
Limits
What these tools cannot do
- They cannot recall open weights.
- They cannot survive a fine-tune that strips the refusals. Data-side refusal still beats a rail you know will be deleted.
- They cannot replace judgment about dual-use. A scanner that never tests the harm you claim to care about will green-light a capable model.
- They cannot substitute for law. A checklist generator is not counsel.
- They are not ethical if they only run as a vendor SaaS on your prompts.
Heiwig
Duty is not capture.
We will not become a guardrail company. We will not stamp compliant on anyone\u2019s stack. We will publish the operator contract: classify, probe, govern, prove — with fail-the-release as the default, on machines the member controls.
Law is already drawing incomplete lines, including partial open-source exemptions under the EU AI Act that vanish for high-risk uses and systemic-risk models. Official codes of practice and training-data summaries are the paperwork. This page is the craft. Neither is legal advice.