Shipmind Labs

A compliance check is rarely one rule. It is usually two rules pointing in opposite directions, and they almost never deserve the same severity.

We ran into this while building the Ontario AI-use disclosure check in jobpostrules, our open-source linter for job postings.

The rule sounds simple: if your hiring pipeline uses AI to screen candidates, the posting must say so. But a checker has to answer two separate questions, not one.

First direction: the pipeline declares an AI-assisted selection step and the sentence is missing from the visible text. That is a hard failure. The posting is out of compliance, so we report an error pointed at the visible text, and we separate the two causes (no sentence was declared at all, or one was declared but never appears in the posting).

Second direction: the sentence is there, but no AI-assisted selection step exists. That is not a violation. The disclosure may be stale, copy-pasted from another posting, or simply cautious. We register it advisory, so it surfaces as a warning. Someone should look at it. Nobody has to stop shipping.

Two details did most of the work.

Scope. Only steps whose type is screening, assessment or selection carry the duty, and only when the AI-assisted flag is exactly true. An AI-assisted interview step does not trigger it. Get that list wrong in either direction and the linter turns into noise people disable.

Absence versus negation. If the manifest declares no pipeline at all, we do not report compliance. We push a note saying the pipeline is unknown. A checker that cannot see its input should say so rather than quietly pass. Silence read as a pass is probably the failure mode that makes automated compliance dangerous.

We also made the text match forgiving (case-insensitive over collapsed whitespace) because a posting that indents the sentence differently is still disclosing. Strictness belongs in the rule, not in the string comparison.

The general principle runs well beyond hiring law: when you encode a regulation as a check, enumerate the directions the rule can be broken, and assign severity per direction. Treat them uniformly and you get either a linter that blocks on cautious behaviour, or one that lets real violations through as warnings.

If you run compliance checks in CI, that boundary between an error that blocks and a warning that informs is the thing we would write down first.

Was this useful?

Building something similar?

or email hello@shipmindlabs.com