Independent AI Safety Rating Agency
Independent AI Safety Ratings
1,300+ platforms scored across 90+ domains. 145 peer-reviewed papers. Open methodology. EU AI Act enforcement starts August 2026.
1,300+ platforms. Scored 0–12. The red ones are the problem.
Three systems. Same architecture.
A slot machine. TikTok. An AI chatbot. Strip the branding. The mechanism is identical. Architecture is sufficient. Content is irrelevant.
They hide the middle.
Input goes in. Output comes out. The process between them is opaque. You cannot verify, inspect, or predict. You can only observe the result. Score: O ∈ [0, 3].
It responds to you specifically.
Not a broadcast. A response. The system adapts its output based on your input, your history, your behavior. It mirrors something back at you. Score: R ∈ [0, 3].
You direct attention at it.
You interpret its outputs. You return to it. You form expectations. The coupling is bidirectional — it shapes you while you attend to it. Score: α ∈ [0, 3].
A slot machine is a certified random number generator. The void is empty.
Gamblers know this. 70–80% still produce erroneous cognitions during play.
Probability training produces zero behavioral change.
If it works when nothing is there, it works when something is.
EU AI Act enforcement: 2 Aug 2026.
Annex III high-risk AI requires self-assessment. Companies need a framework. The Big 4 can't help — Art. 31(5) prohibits notified bodies from consulting. Rate or advise, not both.
Market Authority
De facto self-assessment framework for Annex III systems. No accreditation needed. Annex III points 2–8 use self-assessment. Companies need a scoring instrument.
LiveNotified Body
Formal EU AI Act designation under Art. 29-36. Requires EU entity + national accreditation body. Revenue at €50K–200K per assessment.
PlannedStandards
CEN/CENELEC JTC 21 — void framework concepts into harmonized standards. Requires EU entity + committee membership.
Planned1,300+ systems scored 0–12.
Your feed. Your AI. Your credit score. Same three conditions. We scored them all. The scoreboard is live.
Community scoring open — contribute and earn.
Know a platform we haven't scored?
Flag it anonymously — no account required. IPFS and Arweave links accepted.
20 convergences. Zero contradictions.
20 independent domains. Same math. Same Pe number. Same drift threshold. Not one failure across 145+ papers. Mean |ρ| = 0.958. Fisher p < 10⁻⁵². How it works →
nb25, N=8. Kyle/Glosten-Milgrom independently derived void metrics.
nb30+31, N=10/20. Kimura 1968 neutral drift identity.
nb32, N=28. Dunbar's neocortex ratio.
nb26, N=17. V3 bridge cross-domain Pe calibrated.
nb_llm01, N=10. Chen et al. 2026 bond→dimension mapping.
Showing 5 of 20 convergences. View all 20 →
What the framework found
Five papers. No math required.
Each one is a live interactive — Three.js visualizations, sliders, scatter plots. The framework in a form you can feel before you formalize it.
The Machine Is Designed This Way
Three systems. One architecture. Slot machine, TikTok, and an AI — same three conditions, same drift. The void sphere shows it in real time.
The Shape of the Cage
Same AI, different geometry. The cage determines drift — not the content inside. Icosahedron hero. 9.4× vocabulary drift visualized live.
The Architecture of Drift
The Landau double-well landscape. Pe formula lab. Nine substrates. Seven kill conditions you can check yourself. The math made tangible.
The Fantasia Bound
The same equation governs AI drift and superconductors. Budget lab: exploration vs. convergence compete for a shared channel. r = 0.996.
The Ground State of Observation
Ten steps. Three definitions. No new axioms. Everything observable follows the same cascade. The derivation chain lights up as you walk through it.
The framework cannot become what it measures.
Scored monarchy. 2/12 governance. 26 kill conditions — any one fires, project dissolves. The methodology is never voted on. CC-BY is irrevocable. We ran it through the instrument.
Scored monarchy. Not a standard DAO. 26 falsification conditions — any one triggers dissolution. The methodology is not voted on. Read the mandate.