Shipmind Labs

An early placed-on-market date does not remove a transparency duty. It moves the date by which the marking has to be in place. Two different things, and in code they are two different branches.

We ran into this while building disclosurekit, our open-source assessment tool for the EU AI Act's Article 50 transparency obligations. The whole policy layer reads exactly one date, and our discipline was to read it as a deadline and not as a duty.

The mechanics, because that is where the mistake usually lives:

The system profile gains one optional field holding an ISO date string or the literal "unknown". Before the assessment builds its applies, excluded or needsReview sets, it resolves a single boolean: was this placed on the market before 2 August 2026, the date Article 113 gives for when Article 50 starts to apply.

An omitted field returns false. An "unknown" returns false too. Absence of evidence is not evidence of an earlier date, so the default has to be the stricter reading.

A malformed string does not fall back to false. It raises an error that names the field and states the expected shape. A date you cannot parse is an input problem, not a legal conclusion, and quietly treating it as "not early" would hide a bad integration behind a plausible answer.

The resulting flag reaches exactly one obligation. Inside it, the decision takes the law-enforcement carve-out first, then the assistive-editing claim, and only then the date. An obligation that is exempted or undecided never acquires the deadline note, because a note about when you must comply makes no sense attached to a duty that does not apply to you, and it would read as an admission if anyone audited the output.

One more thing, and it matters: no branch here produces a Review. A date does not push this obligation into "needs human review". Dates are not ambiguous the way a product claim is.

The field also joins the question set, so the inputs record either shows the value or records "not stated". The rules version is pinned, so an answer can be traced back to the exact ruleset that produced it.

The general shape, beyond this one regulation: in compliance code, every input needs to be classified as changing whether a duty applies, changing when it must be met, or changing nothing but the audit trail. Conflating the first two is probably the failure mode that turns a tool into a liability, since it produces confident output that excuses a client from something they still owe.

If you build rule engines against regulation, we would like to know where you draw the line between a malformed input that should throw and one that should degrade to the conservative default.

Was this useful?

Building something similar?

or email hello@shipmindlabs.com