AI Alignment & Control Simulator

SCENARIO SELECTION PRESETS

SYSTEM PARAMETERS

Capability Scaling Velocity (v) 7.5

Alignment Investment Rate (b) 4.0

Verification & Audit Lag (τ) 3.0 yrs

DYNAMIC TELEMETRY

HUMAN OVERSIGHT 42.1

MAX CAPABILITY 88.4

DRIFT INDEX 0.68

SURVIVAL CHANCE 31.5%

AI Frontier Capability

Human Verification Bandwidth

Drift Threshold Boundary

Mechanics: Frontier AI capabilities scale superlinearly when v exceeds verification feedback bandwidth. As verification latency τ widens, human interpretability bounds fail to intercept recursive system behaviors, triggering a Control Vulnerability Zone or irreversible governance decoupling.

How toy control formulas differ from evidence

Your browser does not support video.

Read the explanation

The default settings are capability velocity seven point five, alignment budget four, and verification lag three. Capability is a capped linear combination, while oversight is twelve point five times budget divided by one plus point zero six three times lag. The source displays capability eighty eight point four and oversight forty two point one, then a drift label point six eight and a survival label thirty one point five percent. These are authored toy formulas and a default override, not measured capabilities or an empirically calibrated survival probability. The chart uses different formulas from the dashboard. Capability follows a logistic curve, and oversight starts at twenty then rises after half the lag has elapsed. Forty one samples span zero to ten in quarter year increments. At the default endpoint the chart’s capability approaches one hundred and oversight about forty four point seven, unlike the dashboard values eighty eight point four and forty two point one. The moving dots here follow those exact source equations. A shaded gap describes this toy curve difference; it does not forecast real loss of control. Safe Align uses velocity three point five, budget nine, and lag one point two. Its formulas give capability forty point six and oversight one hundred four point six. Oversight is not capped at one hundred in dashboard calculation, although chart values are capped. Drift clips to point zero one, so the survival formula produces ninety seven point five percent. Those limits and labels follow from chosen constants. They do not certify a model as aligned or estimate humanity’s real survival chances. The graph asks for forty steps, while export asks for twenty. Because the function always advances in quarter years, the JSON time series ends at year five even though the visible chart ends at year ten. Slider changes remain local and no training or governance intervention is performed. The status names are threshold bins on rounded drift. Native preservation checks compare real preset and slider events with all remote requests blocked; unavailable D three remains a source dependency failure rather than evidence of a working graph.