A complaint that a regulator can act on is not the same as an angry post
Open-submission incident databases fail in three predictable ways: they expose the reporter and the host to defamation, they convict a provider before anyone has checked, and they produce something no authority can actually open a case on. This page sets out the design we think answers all three — and marks plainly which part of it exists today.
Measurement, not certification. Board unreachable from this browser — read it yourself at /api/gspc
Three ways a public incident database goes wrong
The first is legal: a database that publishes an allegation about a named company the moment it arrives puts both the reporter and the host in front of a defamation claim, and the reporter usually has the least protection of anyone involved. The second is procedural: naming a provider before anyone has checked convicts them by audience, and it contaminates the neutrality of whoever is meant to measure them afterwards. The third is practical: an authority needs evidence it can open a case on, and an anonymous post is not that.
- PainDefamation exposure lands on the reporter and the host, not the platform that caused it
- PainPublishing first destroys the neutrality of the measurement that should follow
- PainRaw reports are not in a shape an authority can act on
- Only hereThe fix is not more moderation — it is a different pipeline shape
Four states, and only the first one runs today
A report moves through four states. SIGNAL is intake: someone describes behaviour they saw. SHIELDED holds the allegation in escrow so it is not public while it is unverified. CHALLENGED gives the named provider a fixed reply window before anything is published — the subject hears it from us before they read it. ACTIONABLE assembles what survives into a complaint package a market surveillance authority can open. Today, only SIGNAL exists. The other three are design, and we would rather say that than let the diagram imply otherwise.
- You getSIGNAL — the public intake. This runs.
- You getSHIELDED — allegation held, not published. Design.
- You getCHALLENGED — the provider gets a reply window first. Design.
- You getACTIONABLE — assembled into a complaint package. Design.
The subject hears it from us before they read it
Every structural safeguard that keeps an independent body credible while it publishes uncomfortable findings comes back to the same thing: notice before publication, and a real chance to reply. It is the difference between an instrument and a pillory. It is also the difference between a finding an authority can use and one a lawyer can dismantle. We are not claiming this creates any legal protection for us — that is a matter for a court, on the facts, in a jurisdiction. We are saying it is the right way to behave, and that we would rather be slow than be a rumour mill.
- Only hereNothing about a named system is published before its provider has been notified
- You getA correction is a new record; the original is never quietly deleted
- PainWe make no claim about the legal effect of any of this — that is for a court
Article 85 gives the right. A compiler would make it usable.
Article 85 of the EU AI Act gives any natural or legal person the right to lodge a complaint with a market surveillance authority. The right already exists; what is missing is that most people cannot assemble a complaint an authority can open — the provision it engages, the system and version, what was observed, when, and what evidence supports it. A compiler that produces that package is a document generator, not a grant of standing, and this page does not pretend otherwise.
- You getThe right to complain is statutory and already exists
- PainA compiler assembles a document; it confers nothing
- Only hereMeasurement stays separate: we produce evidence, an authority decides
What this page does not claim
We publish the limits with the results. Everything below is something a reader could reasonably assume from a page like this one — and each is something we cannot presently evidence, so we say so rather than let the assumption stand.
- We do not claim this loop is built. State one — the public intake — exists. Escrow, the reply window and the complaint compiler do not: searching this repository for escrow, right-of-reply, 72-hour and Article 85 returns no implementation.
- We do not enforce anything, and we are not a route to enforcement. We measure and publish; a market surveillance authority decides what to do. The boundary is on our homepage and this page does not move it.
- We do not make anything legally binding, and we do not confer legal standing. Article 85 confers the right to complain; a compiler would only assemble the document.
- We do not claim that holding an allegation in escrow creates qualified privilege or any other legal protection. That is decided by a court, on the facts, and differs by jurisdiction. We describe the behaviour, not its legal effect.
- We do not assert what a market surveillance authority can or cannot act on. We say only that an anonymous report is not usually in a shape that supports a case.
- We do not publish a public incident register today. The register at /watchdog/report is deliberately empty, because an empty register is honest and a populated illustrative one is not.
- We do not describe this as trustless. Nothing here removes the need to trust someone; it makes what we did checkable, which is a smaller and more defensible claim.
Coverage on this page is never typed by hand. Board unreachable from this browser — read it yourself at /api/gspc Corrections to anything we have published live in the refutation ledger — append-only, never a silent edit.