Nervous MachineLaunch autoresearch
What other industries already solved for launch
Agents propose techniques from aviation, chip cooling, nuclear power, semiconductors and more for launch's hardest problems. A causal harness keeps or discards each one against a rule written down before it runs, and remembers everything that was tried.
How it works
The graph behind it starts from every orbital launch attempt since 1957 and the published root causes of failures, then links in evidence from dozens of other fields.
Propose
A technique from somewhere else
An agent (or a person) names a technique proven in another field and the launch problem it should help. It reads what's already been tried first, so nothing is proposed twice.
Evaluate
Through the causal graph
The harness traces which launch outcomes the idea would move and how certain each link is. It flags conditions that rule it out (cryogenic, vacuum, vibration) and side effects it would bring in. Ideas only earn credit from evidence across several programs.
Keep or discard
Against a rule set in advance
A step is kept only if it improves the current set of practices by more than it costs to integrate. Every decision, and its reason, is recorded. Nothing here counts as proven in launch until launch evidence earns it.
Current leaderboard
Loading the latest run… Updated —.
Best score by step
The practice set's expected score after each proposed step. Dots mark kept steps. The line only rises when a step beats the current set.
Scores are the model's expected benefit from evidence in other industries, not results from launch.
Launch problems: covered and open
From the failure and emissions analysis, ordered by how much each matters. An open gap means no imported technique has earned a place yet. A bet is a proposed technique waiting for evidence; it can't enter the practice set on a good idea alone.
| Problem | Weight | Adopted | Bets queued | Status |
|---|
What the launch record already says
From 7,186 orbital launch attempts (J. McDowell's GCAT catalog) and published failure investigations.
Maturity dominates failure
An engine's first flights carry about ten times the failure odds of a mature one, more than any choice of propellant or design. Propellant chemistry doesn't predict reliability once maturity is accounted for.
New rockets fail from design, mature ones from production
Since 2000, most failures on a vehicle's first three flights involved a new design element. Failures of mature vehicles are mostly manufacturing defects and contamination.
Every propellant trades one pollutant for another
Kerosene leaves soot high in the atmosphere. Solid motors leave alumina and chlorine in the stratosphere. Hydrogen is cleaner but usually flies with solid boosters.
The evidence for reuse is one program deep
Reuse cuts new hardware per tonne sharply, but nearly all of that evidence comes from a single vehicle family. The harness won't give reuse full credit until a second program confirms it.
Interested in contributing to this project? Email heidi@everychart.io.