The 10% AI Extinction Number Does Not Measure "AI": It Mixes Two Different Objects
Research monograph v3.8-EN — admission kernel and applied non-identifiability of unconditioned expert AI-risk aggregates
doi:10.5281/zenodo.22902611 · published in Zenodo, CC BY 4.0
What this article is. A research monograph on applied non-identifiability: why an unconditioned expert aggregate such as a single “AI extinction” percentage mixes two different objects, and what an admission-kernel framing changes about the question. The Zenodo PDF (v3.8-EN, math-rendered) is the version of record.
What it is not. It is not a technical note of the ElarionX CPMS Cryogenic Referent Registry (Notes 000–004). It is not a Kerr / spacetime paper. It does not claim an AGI production timeline, and it does not deposit proposer internals as part of this site page. The imprint is ElarionX, not ElarionX CPMS.
On the title. A survey can report that experts put roughly ten percent on extinction from “AI.” That figure does not tell you which object they averaged. Class A and Class B improvements are different control problems. Mixing them produces a number that sounds like it measures AI and does not.
Abstract
Unconditioned expert aggregates for AI-catastrophe risk collapse distinct objects into one percentage. This monograph separates Class A (improvement that does not require a live human admission gate) from Class B (improvement that does), and treats the admission decision as its own kernel rather than as a silent average inside “AI.”
The core track is applied non-identifiability of those aggregates (Theorems A–D in the PDF). Secondary tracks develop an admission-kernel language, finite CMDP witnesses, a Part IV toy, a Part V specialization-vs-general toy tagged as an engineering hypothesis, and a proposer/admitter role split with an invitation for admitter-side help only.
Read the version of record
The citable, math-rendered PDF is on Zenodo:
- Version DOI: doi:10.5281/zenodo.22902611
- Concept DOI (all versions): doi:10.5281/zenodo.22902151
- ORCID: 0009-0004-4994-1554
This page is the public entry on elarionx.com. Where site text and the deposit ever differ, the Zenodo deposit governs.
What the PDF contains
- Non-identifiability pillar — why a single unconditioned extinction percentage does not identify “AI.”
- Admission kernel — representation-invariant language for the human gate.
- Class B framing — live human veto and where risk then sits (including gate capture).
- Finite witnesses / toys — including Part V specialization under fixed budget (engineering hypothesis).
- Role split — proposer ≠ admitter; §9 invitation is for admitter help only.
Cite
Elizondo Arias, L. J. (2026). The 10% AI Extinction Number Does Not Measure “AI”: It Mixes Two Different Objects (v3.8-EN). Zenodo. https://doi.org/10.5281/zenodo.22902611
Version of record. The citable version of this article is the Zenodo deposit, doi:10.5281/zenodo.22902611. The text on this page is the same version; where they ever differ, the deposit governs. To cite the article across all future versions rather than this one, use the concept identifier doi:10.5281/zenodo.22902151.
Found an error? Corrections are wanted and will be credited. Where a correction changes a conclusion, the change is recorded as a change rather than edited away.