NASA asked for a model. The obstacle was never the model.
The brief asked for fully autonomous, real-time launch operations: a model performing under a strict set of circumstances. Aphelion and Boxcar started somewhere else, and that decision surfaced the real problem early.
At 22 minutes of round-trip signal delay, a human cannot be in the moment. With this, they are still in the loop.
From a model brief to confidence machinery
1 · The brief
A NASA SBIR program for launch autonomy. The deliverable, on paper: a model that performs under strict circumstances, in environments where humans must stay in the loop but cannot be in the moment.
2 · The contrarian start
Instead of starting with the model, the team started at the human touchpoints: what ground control would see, question, and have to trust.
3 · The discovery
Within weeks the real obstacle was clear. Not whether the model could perform, but whether the people responsible could believe it. Confidence, not capability, was the bottleneck.
4 · The machinery
A dashboard where operators simulate the model's calls, inspect its reasoning, and feed corrections back for continuous learning. Confidence you can rehearse before you have to rely on it. Edge inference on the vehicle; terrestrial cloud for what does not have to ship. Intent and policy travel together.
5 · The unplanned payoff
The confidence machinery turned out to be exactly what terrestrial launch operations need today. What was built for Mars-distance autonomy became valuable on the pad this year, and mission profiles that could not be designed before became designable.
What was true at the launch pad is true in every company adopting AI.
Capability is arriving faster than confidence. The teams that win are not the ones with the best model; they are the ones who can rehearse, inspect, and inherit the reasoning. Boxcar is that machinery for the product lifecycle.
Your version of this story starts with one workflow
Five weeks, explicit autonomy limits, measured value, and every artifact stays with you.