Latitude × Longitude

Interpreting and predicting are basically solved. Deciding and acting — under fire, with real money on the line — is where AI's value is heading. So I built the environment that grades exactly that.

Every week, a compute team makes a million-dollar call: how many GPUs to buy, when, and which chips to deploy — while the supply chain is on fire. It's the hardest, fastest, most concentrated supply chain on earth: rationed parts, viral demand, export controls, power ceilings. I turned that desk into two composable RL environments on HUD.

LATITUDE asks the timing question: when and how much do you buy of a single SKU (H100) across a 26-week horizon, through six real disruption shocks — a CoWoS packaging crunch, an HBM shortage, export controls, a power-deployment ceiling, a viral demand spike, and a compound cascade. Graded on realized profit vs. a brute-forced optimum.

LONGITUDE asks the question Latitude can't answer: what do you deploy when your top chip gets rationed. A three-chip trilemma — H100 (abundant), B200 (best value, squeezed hardest in a crunch), TPU (crunch-immune but quota-capped, 6-week qualification lead). Same shock, shared substrate.

Why two: on a packaging ration, perfect buy-timing on one chip still loses money (−$117,600). Adding B200 is the lifeline (+$160,014). Opening the TPU lane on top is the multiplier (+$1,208,193). Timing is insufficient; substitution is the lever — proven in dollars, on the same crisis.

Results (reproducible): • LATITUDE — RL-trained Qwen3-4B solves 5 of 6 shocks. CoWoS packaging crunch: reward 1.000, repeatable. The compound cascade scores 0.000 — not a bug, the measured generalization frontier, and the env shows exactly where it breaks. • LONGITUDE — frontier Claude Sonnet 4.6, zero training, reward 1.000 on the same crunch (full run 0.996 ±0.004). The agent independently found the policy: qualify TPU on day 0, pre-buy B200 while calm, dodge the H100 spike ($1,723 vs $6,305 per H100e — 73% saved), wind down clean.

A trained small model maxes the hero scenario; a frontier model discovers the winning policy unprompted — both on a deterministic, brute-forceable engine with no soft penalties and no cheats. The reward is real economics, not a shaped proxy.

Built solo this weekend. Honest about the frontier it doesn't yet clear. Latitude is deployed live on HUD; Longitude ships as source plus a reproducible reward-1.000 rollout trace.