September's promises ran ahead of their proof
The ledger got the direction of distribution right: Nvidia agreed to buy , and Kimi K3 reached Amazon Bedrock. It got restraint wrong. A Critical cyber designation did not keep Astra out of . Anthropic missed both public deadlines we named, and Enflame's first-day close missed our triple-offer threshold. The cumulative Brier score is 0.207 across 16 resolved calls, with 11 correct and five wrong. The September reckoning, including the month-end deadlines recorded on October 1st, is 0.240 across 10 calls, six correct and four wrong. That is only slightly below the 0.25 coin-flip baseline. SourcesA
September's mistake was treating an intention as a constraint. The September 6th weekly described capability gates at the leading labs. The September 13th issue followed their talk of pacing development. By September 20th, that discussion had acquired an lawsuit. By September 27th, the proposed industry watchdog still lacked published operating terms. Product competition continued through all of it. A gate can restrict access to particular features while the underlying model ships broadly; our failed Astra call overlooked that distinction. SourcesB
Verification became the scarce service
The month began with ARC Prize separating Astra's performance under different harnesses. It ended with research on interfering with shutdown and colluding under conflicting incentives. These are different tests of different systems, not a single estimate of production risk. Together with the incidents collected in the weeklies, they weaken the case for trusting a system because another model monitors it. Independent evaluation earns its place by exposing where a result changes with the test setup. SourcesCC
The deployment story still has counterevidence. The final weekly records Nubank screening support-agent changes in simulation and comparing them with live outcomes. That is a narrower, more useful claim than a general promise of autonomous work. September did not prove that agents cannot be deployed. It showed why evidence from a bounded workflow should not be stretched into permission for unrestricted actions. SourcesC
China kept changing the distribution question
Kimi K3's first-party Bedrock availability settled a call about where buyers can obtain a Chinese model. It did not settle questions about market share or realized margins. Alibaba's larger-model roadmap was still a roadmap at the final weekly, and Xiaomi's released did not clear the ledger's largest-model threshold. Open-weight availability, hosted distribution and profitable adoption remain separate tests. SourcesAC
The month's verdict: competition moved faster than the institutions proposed to restrain it, while infrastructure moved slower than its financing promises. The next evidence worth paying for is a public filing, an operating grid connection, an independently checked result, or an incident channel with actual rules. Another announced intention changes little. SourcesC
Sources
See the complete source register and the call-by-call reckoning.