Empirical safety boundary simulator modeling self-improving recursive AI velocities, alignment lead times, and catastrophic breach likelihoods over a 10-year horizon.
Compound capability acceleration factor per year during autonomous training loops.
Pre-deployment evaluation and red-teaming runway prior to capability release.
Dedicated alignment and interpretability budget proportional to capability compute.
| Intervention Framework | Enforcement Mechanism | Simulated Risk Reduction | Target Lead Time Shift | Efficacy Rating |
|---|
Simulation models the existential risk dynamic referenced in The Verge reporting on X, detailing the departure of researcher Jacob Coxon from Anthropic and subsequent statements by Alignment Science Lead Evan Hubinger asserting a >10% likelihood of catastrophic outcome if autonomous self-improvement loops bypass verifiable alignment bounds.
This calculator assigns illustrative risk and timing, not validated extinction probabilities. Its general default arithmetic multiplies growth by the recursion boost, subtracts alignment times point nine five times the lead factor, and gets a deficit of one point one zero two five. Multiplying by point one eight and rounding gives nineteen point eight percent. Its timing expression gives four point eight years. But an exact default branch replaces those results with twelve point five percent and four point two years. These are source fixtures, not empirical validation or verified researcher estimates. The chart is generated with eleven yearly points. Capability compounds a growth rate whose recursion boost increases by point zero eight each year. Alignment compounds a constant factor of one point six at the defaults. By year ten, these recurrences yield about one hundred forty four point eight eight capability units and one hundred nine point nine five alignment units. The plotted values are arbitrary scales. They do not measure intelligence, proven safety, or a real autonomous system, and the recurrence is separate from the headline terminal-risk equation. The breach curve computes the positive gap between those two scales, divides it by gap plus fifteen, and applies a timing gate. At the defaults, the gap stays zero through year seven. Rounded chart risk then becomes thirteen percent at year eight, forty three percent at year nine, and seventy percent at year ten. The headline terminal-risk card still says twelve point five percent. Those are different formulas, so the chart endpoint does not validate the terminal card. Neither result has been calibrated to actual event frequencies. The button labeled run Monte Carlo with one thousand runs calls the same update function with a boolean flag. That flag changes the status to say one thousand iterations converged. It does not run a loop of random trials, estimate uncertainty, or test convergence. The one thousand run count is metadata. Policy and intervention names are assigned threshold labels, not demonstrated legal requirements or evidence-based recommendations. Export saves this deterministic scenario and its source context. Our local native checks verify selected controls and canvas preservation, not the truth of the catastrophe estimates.