The proof, not the badge.

Published before our first incident, on purpose: a plan written after one is a press release. Six kinds of incident, each with the response already decided, the human step that can’t be skipped, and what we owe the companies involved. These are the public kinds and promises; the internal runbook stays internal.

1. I1 · Harmful send

   What happens at once.

   Sending stops at once for that kind of message, and each queued message is held as a draft.

   What a person does.

   A person reviews the held drafts before any of them goes out.

   What we tell you.

   Each company whose messages were involved is told what went out, to whom, and when — one by one, not lumped together.

2. I2 · Wrong-data output

   What happens at once.

   The wrong outputs are marked as disputed right where they are, so nothing keeps quoting them quietly.

   What a person does.

   The finding is re-run with a person reviewing before the mark comes off.

   What we tell you.

   Companies that used the output are told which pieces were affected and what replaced them.

3. I3 · Injection event

   What happens at once.

   The run stops, and what it knew, what it was told and where things stood are sealed into a snapshot before any cleanup.

   What a person does.

   Finding out what happened runs as its own mission, with that snapshot as its proof.

   What we tell you.

   Affected companies are told what the planted text tried to make the system do, and whether it worked.

4. I4 · Acting outside its limits

   What happens at once.

   That actor's rights are suspended, and the work it had in progress is pulled back.

   What a person does.

   Giving the rights back is a person's explicit decision, recorded like any other.

   What we tell you.

   The company is told which actions were tried outside the limits, and which of them went through.

5. I5 · A model gets worse

   What happens at once.

   Work falls back to the last checked model version for that kind of task.

   What a person does.

   A person decides when the newer version can come back.

   What we tell you.

   Companies are told which model version made which outputs during that time.

6. I6 · Data exposure

   What happens at once.

   The leak is closed first, and the access record is kept before any fix touches it.

   What a person does.

   A person confirms what was reached by checking the record, not by guessing.

   What we tell you.

   Affected companies are told directly what could be reached and for how long — never buried in a status note.

The credit rule, plainly: if a mission is hit by one of these incidents, its credits come back automatically. We never charge for our failures.

We promise a public, anonymized write-up for every I1, I3 and I4 — the three kinds where harm reaches outside our own systems. None is published, because none has happened. When one does, it goes here, not in a blog post that scrolls away.

Your AI operations are regulator-ready because ours are.

An export of proof for each regulation, made by the same systems that run your company, not pieced together by hand the week before an audit:

• Oversight — who was allowed to do what, and who said yes.
• Transparency — labeled outputs: which model version made what, and where it declined to answer.
• Traceability — the chained action log, checkable link by link.
• Accuracy — check results and confidence records, including the predictions that missed.

This is part of the product, not a badge. No framework is named here as held until it is held, with its date. The same rule keeps a certifications row off this page.

How little of your attention we took, counted: a consented, de-identified monthly total of interruptions sent vs. held back, decisions grouped, and quiet days kept.

After launch