Summary

The frontiers part took up the three limits that compute does not lift. Learning runs on a finite stock of human text, and every escape route changes the shape of that problem rather than removing it: synthetic data, reward from a checkable environment, more compute at inference, models that keep learning after deployment. Measurement turned out to be harder still: a task-length horizon reports progress that a saturated benchmark hides, but the quoted horizon is a fifty-percent-reliability number, and the reliability a deployment needs corresponds to a horizon several times shorter. Verification closed the arc: generating a claim is cheap and checking it is not.

Running through all three is the same shift: the binding constraint has moved from production to acceptance. A frontier system can produce candidate answers, programs, proofs, designs, and hypotheses faster than any institution can confirm them. What is scarce is no longer the result but the standing to trust it, and standing is built from proof objects, replications, adversarial review, and the operating record of whoever vouches for the work.

The open question is whether proof, replication, oversight, and operating contracts can scale quickly enough to keep accepted knowledge from becoming the bottleneck. That question is not purely technical, since who pays for verification and whose verification counts are settled in a market rather than a lab. Part XI turns to ecosystem and economics, where the cost of capability, the terms under which it is released, and the structure of the industry selling it become part of the architecture.