The beta is closed at 930 companies
Goliath SI is in limited deployment with 930 companies already running on the system. We are not adding new organisations while the operating environment settles and while the models, Cerebrum and the Sling fleet are being fine tuned. Existing deployments continue without change.
930 companies, and no 931st for now
There are 930 companies on Goliath SI today. That figure is a ceiling rather than a milestone. The beta was opened to organisations that already had the operational discipline to run an autonomous fleet against their own systems, and it grew to 930 before we stopped admitting new ones.
Closing the cohort was a decision, not a capacity limit. We have the compute to run a larger beta. What we do not yet have is confidence that we could expand the deployment surface at the same time as the environment around frontier deployment is moving, without one of those two things getting less attention than it needs. Growth is the easier of the two to defer.
Nothing changes for the 930. Tethers hold, Slings stay seeded against the same objectives, Cerebrum keeps adjudicating every release, and the connectivity work continues against the estates those companies already run. The freeze applies to admission, not to the systems in production.
Three conditions, all of them outside our control
Accelerator supply is allocated quarters ahead, and the terms attached to it have become materially more specific about where frontier class hardware runs, who operates it and what it is used for. A beta that expands faster than its compute commitments are confirmed is a beta that degrades for everybody in it. We would rather hold admission than write a cheque against silicon that has not been scheduled.
Aviation is the second condition. A meaningful share of our deployed work sits near regulated airspace operations, where expectations around automated decision support are tightening and the standard of evidence for any system that touches an operational decision is rising. We agree with that direction, and we are not going to widen a deployment into it before our own evidentiary record meets the bar we expect to be held to.
The third is legislative. Frontier AI oversight is being worked through in hearings, and the questions being asked, about provenance, about what an autonomous system is permitted to do without a human in the path, about who answers when it is wrong, are the same questions our architecture was built to answer. Answering them in a hearing room with a record, rather than with a roadmap, is worth more to us than the next hundred customers.
None of the three is a crisis and none of them is a surprise. Taken together they make expansion the wrong use of this quarter.
What a closed cohort is being spent on
Models first. The Cortex federates a frontier model estate and runs above it, and the routing layer that decides which model sees which work is being retrained against the traffic the 930 actually generate rather than against a benchmark suite. Where a specialised model is cheaper and better on a narrow class of work, the router should already know that before the request arrives. Fine tuning here is mostly about removing the cases where it does not.
Cerebrum second. The guardrail intelligence is being tuned in the direction that costs us the most: a higher hold rate on thin provenance. Evaluators that drift toward agreement with the Cortex are rotated out, and the rotation schedule has been tightened. We are also shortening the distance between an objection and the evidence that resolves it, because a hold that takes a day to clear is a hold that teaches operators to route around the gate.
Slings third. The fleet runs ahead of the request, which means a Sling that is wrong is wrong before anyone asked it anything. Work in this cycle is on the intent horizon, how far ahead a Sling is allowed to work before it checks in, on tether behaviour under partial connectivity, and on the dead man interlock that halts a Sling that misses check in rather than warning about it. Synthesised adapters are being hardened against the estates where schema reconciliation currently escalates more often than it should.
The three are tuned together, not in isolation. A faster Cortex with an unchanged guardrail is a worse system, and a longer intent horizon without a harder tether is a liability. The cohort stays frozen until the three move as one.
What changes for the 930
Nothing you have to act on. Fine tuning lands continuously and under the same release interlock as everything else: three evaluators, three lenses, one credible objection to hold. Behaviour changes that would alter an output you depend on are flagged against your tether before they ship, not after.
The holding rate on your own work may rise slightly during this cycle. That is deliberate. A held artifact comes back with the objection attached and the evidence that would clear it, and your operators can answer it or abandon the work. We would rather you see the hold than see the output and find the fault later.
Expansion inside an existing deployment is not affected by the freeze. More Slings, more estates, more seats within an organisation already on the system all continue through the normal path.
There is no date, and we will not invent one
The beta reopens when the compute commitments are confirmed, when our evidentiary record stands up to the standard we expect aviation adjacent work to be held to, and when the three fine tuning tracks have converged. We are not going to put a month against that, because any month we named would be a guess dressed as a commitment.
What we will do is say so here when it changes. This page is the record of the cohort size and the reason it is held, and it gets updated when either of those moves.