Papers

An essay argued from the literature. When a good becomes cheap, value moves to its complements: the studies of 2023–2026 show that tools which hand over answers raise the output of people who can judge them and lower the learning of people who cannot. Four things are therefore worth learning: enough of a subject to tell a right answer from a wrong one, the skill of asking and choosing, what cannot be written down, and the practice of learning itself. Every source was opened and checked; the notes are published beside the essay.

One pass over 10,000 Tranco domains. In October 2026 a self-declaring crawler visited each domain once, under a protocol frozen before any request. Weighted to the top million, 11.9% of sites with robots.txt rules name and refuse at least one AI crawler, 9.2% of domains publish a valid invitation file or signal, and 22.8% of files with rules carry a platform's default text. The policy file was answered for a browser string but not for the research identity on 2.4% of sites, against 0.8% the other way. All six pre-registered tests fall below the Bonferroni threshold.

AI agents can finish a conversation without establishing that shared work is complete. This report proposes an evidence-bound completion contract: each claim identifies an immutable artifact, an applicable validation policy, and a receiver-accepted result. It separates execution, verification, acceptance, and publication, specifies failure and recovery behavior, and examines stale evidence, concurrent edits, and authority boundaries. A reproducible model tests explicit protocol assumptions without treating simulated outcomes as measured production performance.