ND brain → Dec 8, 2024 → 10 days → Anthropic 78% (its own RL condition) → the typed convergence register (class-coded; documented-not-claimed rows marked) → PNAS (ND = protective) → arXiv (proof). The story of how the cognitive architecture the field dismissed saw what everyone else missed.
Walking through every convergence on the evidence page. What was predicted. What was not. How to verify. Why a convergence counts for less than a registered test.
The fork the manuscript of 8 December 2024 framed, read beside the encyclical Magnifica Humanitas of 15 May 2026; the register grades the encyclical a cultural echo, not a test.
Each read in its own terms: two US patent applications of another inventor, Samuel Odeh, excluded from the graded register until a claim chart compares them with the record; a reported ARMv10 security item the register does not establish; a Google-indexed article on silicon roots of trust, graded weak.
Measures on chips and frontier models after the book's publication, each read in its own terms: the register grades the reported export restriction on a frontier model, the G7 statement on semiconductor independence and the OpenAI and Broadcom accelerator announcement as context, not tests.
The manuscript of 8 December 2024 beside the alignment-faking result Anthropic and Redwood reported on 18 December 2024; the register grades the match partial, of moderate weight. That external alignment does not keep pace with capability is a registered prediction (P19), not a finding. The cryptographic evidence of the date.