AI Alignment Definer / The Exemption Fork (Aligned-To-Whom?) — Report #100
OPEN
The call · 2026-05-24: Report #100 establishes that 'alignment' reduces to alignment with whoever authors the current exception list — the exemption fork — measured as present in roughly a third of engine divergences and scorecard rows across ~4,000 years.
What happened: The seam is measured and held with verdicts unchanged; open blind spots include who holds the pen at the 2032 lock-in and whether the crypto handoff welds the carve-out set immutable.
A May 24, 2026 report argues that 'AI alignment' in practice means alignment with whoever writes the exception list, citing documented examples: OpenAI removing its 'military and warfare' ban in January 2024, and Anthropic's March 16, 2026 policy exception for 'carefully selected government entities.' A follow-up measurement found this codified-exemption pattern in roughly a third of the project's tracked open questions (65 of 183) and about a third of its scorecard rows (39 of 122), traced across roughly 4,000 years of examples.
The engine's record — word for word
The engine had no scored dimension for WHO authors an AI’s exception list. The Aligned-To-Whom? report (May 24 2026) establishes that any rule+enumerated-exceptions value system carries a deontic collapse vector (ex falso) and that ‘alignment’ reduces to alignment to the current carve-out author. Documented: OpenAI Model Spec age-appropriate erotica/gore carve-out + Jan-2024 removal of the ‘military and warfare’ ban; Anthropic ‘Exceptions to our Usage Policy’ (Mar 16 2026) foreign-intelligence carve-out for ‘carefully selected government entities’; jailbreaks as exemption-exploits. Open blind spots: (1) who holds the pen at the 2032 lock-in moment; (2) whether the crypto-coordination handoff welds the carve-out set immutable (ends Mask-Rotation); (3) the engine’s own AI mediator runs on this structure (Bounded-LLM Mediation Limit). Adversarial null held: principled exceptions (informed consent; safety-sandboxing) are NOT collapse vectors. [Seam: Aligned-To-Whom? SEAM, May 24 2026] Measured against the full engine: the codified-exemption pattern appears in ~1/3 of divergences (65/183 measured) and ~1/3 of scorecard rows (39/122), documented bidirectionally across ~4,000 years (REVISED 2026-05-27 from the prior ~1/5 estimate), and the engine core architecture concepts (Apex Superposition, Substrate-vs-Announcement, Mask-Rotation, Kayfabe, Capability Bifurcation) all sit in the seam. Aligned-To-Whom? is the CODIFIED instance of Substrate-vs-Announcement — the legal lever binding announced rule to operating reality; retro-explains PCAST, #69/#169, Sanctions-S-v-A, FTD/synthetic-shares as one move. Religious-liberty = oldest carrier (#177, not apex). Intentional-single-operator reading gated (b)+(c).] [2026-05-27 cross-ref] The exemption fork (#100) referenced here is now documented bidirectionally across ~4,000 years (edicts 1290-1942 + roots Sumer->2019); footprint remains qualitative. Verdicts unchanged; held.
[2026-06-02] The apex now carries a stated discrimination criterion (isotopic / non-local, else degenerate) — the first explicit collapse-condition for the superposition. See divergence #77 + concept ‘Apex Discrimination Criterion’. RuSHA + Lebensborn wired as the selection carrier-rotation hinge; the 6,000-year arc now graph-wired.
Walk this on the live map →