Decisions that someone can refuse.
Governance, at a small lab, is mostly about who can say no. Three independent councils sit between the work and the rest of the world. This page describes who they are, what they review, and how the calls get recorded.
Five rules that apply to every council.
Independent of the executive team.
The councils that review safety and integrity do not report to the CEO. Their members cannot be removed for the substance of their votes.
External by default.
Most seats on each council are held by people outside Blankline. They are paid for their time and hold no equity.
Written and signed.
Every consequential decision produces a short written record: what was reviewed, what was decided, who voted which way, and the reasoning.
Published where it affects the public.
Decisions that change a public product, a public claim, or a public publication are made public, with the date and the reason.
Reviewed on a schedule we cannot skip.
The charter, the protocols, and council composition are reviewed twice a year. The dates are on the calendar before the meeting that sets the agenda.
Three groups. Different scope. Same independence.
Splitting these reviews across three groups is on purpose. A council that reviews everything ends up reviewing nothing well. Each one owns one kind of question and is good at it.
Four steps from idea to public.
A release does not skip a step because someone is in a hurry. The steps are written down so a hurry can be pushed back on with a sentence, not an argument.
Read the safety protocolsSpec and risk write-up.
The team proposing the release writes a one-page summary: what is shipping, what could go wrong, what evidence they have that it is safe enough.
Internal safety review.
Safety engineering attacks the release against the protocol scope: jailbreaks, tool-call abuse, data leakage, output integrity, runtime escapes. They sign or refuse.
Council review.
The Governance and Advisory Council reviews the write-up and the safety result. They can ship, ask for changes, or decline.
Decision is recorded.
The decision, the reasoning, and the votes are filed. If the release affects a public product, the decision is published on the next review cycle.
Find a gap between this and how we behave?
Governance documents that do not match the work are worse than no documents. If you see a gap, tell us so we can fix one or the other.