Keep R < 1 when the actors are countries.

Two clocks at UNGA

In the same UN week, two clocks were visible from the same podium. One was the clock of capability: frontier models still compounding, labs still shipping, scientists and executives still warning that development can outpace anyone's ability to manage what was just built. The other was the clock of control: a Call, launched by Finland's president and Norway's prime minister, asking companies and governments to put human direction back on the wheel.

The document is A Call for Control of Frontier AI Models. It is not a treaty. It does not bind a signatory to a stop authority, a shared meter, or a consequence that fires without a meeting. The Finnish host's own narrative is careful about that: an appeal, still open for endorsement, promoted during High-level Week.

What it actually asks, compressed from the official text, is four things. AI must remain under human direction, oversight and control, and be developed in line with international law. Companies should publish safety protocols that include mandatory pre-deployment testing and independent evaluation, with qualified evaluators given enough access to assess risk. Governments and regional organisations should coordinate common standards, share reporting of serious safety incidents, and make eval capacity available across regions. UN member states should build on existing mechanisms and explore an international institution able to set standards, enable verification, and convene states when capability thresholds are crossed.

That last verb is doing almost all of the institutional work. Explore is not operate. A threshold that convenes nobody is a calendar entry.

I am not going to congratulate the room for noticing the gap. Noticing the gap is the easy half. The useful question is whether the Call grows correction capacity, or only describes it.

Map the camps

The official count, from the Finnish host narrative, is 22 leaders from 20 countries. The European Commission is an additional signatory institution, not a twenty-first state. Finland is counted twice as leaders (president and prime minister) and once as a country. Flatten that into "22 countries" and you have already lost the first measurement.

The twenty countries named by the host, plus Norway and Finland themselves, are Australia, Bahrain, Canada, Denmark, Estonia, Finland, Germany, Iceland, Ireland, Kazakhstan, Kenya, Latvia, Moldova, the Netherlands, Norway, Singapore, South Africa, Spain, Türkiye and the United Arab Emirates. The Commission sits beside them. The primary-sources pack lists the twenty-two names as published. I am not going to invent a twenty-third.

Four postures. One document. No single stop authority. The UN Secretary-General welcomed the Finnish and Norwegian initiative and repeated the explore-an-institution clause. Welcome is not a fifth signature, and it is not a meter.

Translate the Call into R = L / H

The book's portable test is the control ratio. Intervention latency (L) is the time from the first point a deviation should be detectable under the documented monitoring design to confirmation that the intervention has taken effect: detection, decision, execution. Oversight validity horizon (H) is how long the critical evidence for a requested scope remains reliable from the assessment time. R = L / H. Keep R below 1. Grow the denominator as the numerator accelerates. Velocity with vigilance. Go as fast as your ability to correct allows.

Read the Call through that ratio and it stops being a mood. It becomes a parts list, most of which have not been built.

Mandatory pre-deployment testing, independent evaluation, shared serious-incident reporting, and eval capacity are attempts to lower L. You cannot intervene in a failure you have not detected. You cannot detect it if no one is required to look, and no one is required to say what they saw. An evaluator without access is a witness behind glass.

Common standards, and an institution that can convene when capability thresholds are crossed, are attempts to protect H. A horizon dies when a new capability jump invalidates the last eval. Convening at a threshold is how you refresh the evidence before the old reading becomes a memory of a slower machine.

This is the unflattering finding, and I want it stated without drama. Twenty countries and the Commission asked for the instruments. Nobody in the room produced a joint number. At the international layer, R is unknown. Unknown is not a milder yellow. In the specification it is a missing measurement, and a missing measurement never becomes green.

The illustrative preset on this site is there so a reader can see the arithmetic, not so a government can claim a band. Harbor taught that habit: show the record, stamp it fictional or illustrative, and refuse to let a teaching number travel as a field result.

Verbal support is not a green gear

Harbor was a fictional coding agent. In the second snapshot the ratio went green and the activity stayed blocked, because an irreversible disclosure could finish before any reactive intervention. Authority to stop was incomplete. A low R is not permission if a gear is red, or if a yellow-blocker is still open. The lesson was never about the agent. It was about the mistake of reading a ratio as a permit.

Apply that lesson when the actors are countries and the mistake becomes a press conference.

Governance is who can say stop, how fast, and on what authority. When some states sign, some endorse without signing, one rejects global control, and another builds a parallel organisation, there is no unified stop authority. That is not a red criterion in the specification's sense: no required international breaker has been tested and failed, and no critical claim in the Call has been invalidated. Yellow is the honest color. Green would be greenwashing. "Multilateral solved" is the sentence the method is built to block while Governance is not green and stop authority is missing.

Missing national republications are evidence gaps, not footnotes. A Call hosted in Helsinki and mirrored in The Hague and Copenhagen is not the same as twenty capitals publishing the same commitment in their own record. Until those republications exist, the evidence pack is incomplete, and incomplete evidence is insufficient, not favorable.

I am leaving Resilience unknown in the world-score on purpose. The specification is explicit: unknown means not established. If you cannot show a shared breaker, you do not get to paint the gear yellow to make the advisory prettier. The toolkit preset is labeled illustrative for that reason, the same way Harbor is labeled a fictional composite. Use it to learn the table. Do not quote it as a measurement of UNGA week.

What to build next

The Acceleration Paradox does not answer pause or race. It answers: grow correction capacity. Lower L. Protect H. Keep R below 1, and keep going.

The next useful object is not another statement. It is a scoped "frontier deployment" assessment: one activity, one evidence pack, L defined as time from capability-threshold detection to coordinated eval or incident response in effect, H defined as how long shared eval evidence remains valid after a frontier capability jump. The companion toolkit now loads that preset at scorecard.html?example=CALL-2026. The case library carries the same reading, watermarked.

What would make a rival check the reading, rather than take a welcome on trust, is the same infrastructure the book already argued for: verifiable commitments, in the Commons sense, that peers can inspect without assembling a new political meeting. A Call that explores an institution is a start. An institution that can verify is the denominator. Arms-control history runs that way more often than the other way: verification technology arrives first and makes the agreement thinkable.

Concretely, before anyone waits for a treaty: publish the four-column map (sign, support, welcome, reject or parallel) and keep it current. Collect national republications, or mark them missing. Write the If-Then convening card: who calls the desk, on what threshold, and what happens if a participant does not arrive. Measure L and H on any frontier activity you actually control, in the same units, and say when the last eval dies. Comparability is what turns a private opinion into an instrument.

The people who signed asked for human control. The people who supported or welcomed asked for it in a softer voice. The people who rejected global control, or built a parallel organisation, are part of the same measurement. Count them. Do not average them. Then grow the denominator, or admit you have not.

Velocity with vigilance. Keep R < 1.

Primary hosts and parallel tracks: the full Call and lead narrative (Office of the President of Finland); national hosts in the Netherlands and Denmark; the UN Secretary-General welcome; the Australian doorstop; the White House UNGA release; the UK UNGA speech; the French Elysee text; and the July 2026 WAICO founding report. One-line entries and the 22 / 20 / EC count live on the sources pack. The argument here is the book's own.

Independent commentary. Not affiliated with or endorsed by the Office of the President of Finland, the Government of Norway, the European Commission, any signatory, or any government named. Gear ratings and the CALL-2026 toolkit record are an illustrative reading, not a legal finding.