A physically-real atmospheric re-entry simulator — Sutton-Graves heating, ablative shield burn-through, plasma comms-blackout, peak-g limits, atmospheric skip. Not hand-written: Qwen3.8-27B (NVFP4) coded it under Nemotron-3.5-Lightning orchestration, both co-resident on one DGX Spark — through a live correction loop.
Entry at 11 km/s, −6.5°. It survived the fire, then came in a hair too shallow and skipped back off the atmosphere.
Live below. Drag velocity, entry angle, ballistic coefficient and shield thickness, then Launch. Every number is simulated.
Runs entirely in your browser — the same single HTML file the model wrote. Open full-screen ↗
The point isn't the sim — it's that the orchestrator handed the whole job to the coder, and we corrected the coder like a teammate.
The task went to the Hermes agent whose main model is Nemotron-3.5-Lightning (:8001). Nemotron decomposed it and delegated the code to the Qwen3.8-27B coder (:8000) — two NVFP4 models co-located on one GB10.
The coder emitted a 32 KB single-file sim — with a missing control-panel and an un-interpolated rgba() template. It ran frozen.
We fed a precise correction back through the orchestrator; the coder edited the file in place and fixed both. Then we calibrated the physics (ablation rate, surface heat capacity so the shell reaches ionization, the re-entry-corridor verdicts) against a headless-browser harness that measured each run.
Throughput read straight off vLLM's Prometheus /metrics on the Spark — dense-27B decode held its measured baseline while the orchestrator fronted the conversation.