2,800+models tracked

31benchmarks

Capability Scaling plots benchmark scores against training FLOP with log-linear and sigmoid fits. Training Evidence runs cross-signal plausibility analysis (chip counts, power draw, cost, supply chain consistency). Autonomy Monitor tracks METR time horizons.

These tools connect compute governance to capability outcomes. How much compute produces what level of capability? Are company claims about training runs plausible given known infrastructure? Which systems approach safety-relevant autonomy thresholds?

Capability Scaling

Benchmark scores vs. training compute across 2,800+ models. Log-linear fits, saturation detection, and compute-efficiency frontiers.31 benchmarks

Training Evidence

Cross-signal claim plausibility: does the claimed compute match available chips, power, cost, and supply chain data?

Autonomy Monitor

METR autonomous capability time horizons plotted against training compute. Safety-relevant threshold tracking.

Capabilities

Capabilities