VerifiedIntelligence
Seven Things That Were Already True
Seven pieces this week turn on the same mistake: treating something already true as something still pending. New Zealand's privacy principle has bound every organisation since May, read this week as approaching. A six-agency guidance paper named the control that would have stopped an AI agent's breach forty-eight days before it happened, left as prose, never tabulated. An open-weight model's sovereignty case rests on an unconfirmed licence. Four companies rewrote their licences to reclaim value a hyperscaler had taken. Four dated documents were read as current this week when their own pages name an earlier date, and this project's own prior research did not survive the same check. On this week's evidence, the obligation, the control or the correction already existed before anyone looked; mapping comes first, action second.
This week's Saturday analysis, on an intelligence sweep misdating four already-current documents and one of its own figures, is worth ten minutes before trusting any date.Read the full analysis
The Rule That Has Been Live Since May
IPP3A has governed every indirect collection of personal information since 1 May, yet this week's sweep described it as approaching.
Orbital Regulation: Fifteen Thousand Satellites and the Rule That Bent for a Rival
The FCC gave SpaceX nine years to build fifteen thousand satellites, and gave Amazon Leo a graduated waiver after missing its own deadline by three quarters. New Zealand's regime reaches the ground station, never the signal reaching a handset.
AI Governance: The Number Gets Worse Before It Gets Better
Organisations running or building a governed context layer for their AI agents report more than double the failure rate of those with no plans for one, because measurement finally found faults that were already there.
Open Weights Tell You Where, Not Whose
A Chinese-published model now tops the honesty leaderboard, fully self-hostable. An afternoon spent trying to confirm its actual licence ended with no answer: self-hosting settles where inference happens, never who shaped the training.
Who Holds the Undo?
Six allied cyber agencies described, in prose, the exact control that would have stopped an AI agent's breach of a government portal. Forty-eight days separated the guidance from the breach it had already answered, unbuilt.
The Great Extraction, and the Revolt
Four open-source infrastructure vendors rewrote their own licences between 2018 and 2024 to reclaim value a hyperscaler had taken. Every fork answered within weeks; a board needs to know who can move its own terms.
AI Verification: Four Dates and One of Them Was Ours
Four correctly dated documents, including a rule in force since May, were read as current this week; a fifth correction landed on this project's own research. Checking a date is the cheapest verification step, and almost nobody runs it.
Of the obligations, controls or licences your organisation currently relies on, which one would survive an afternoon spent checking it against its own primary source, and which would not?