James Oosterhouse / Live demo
TRIAGE: Flight Edition
Ingenuity Flight 6, the first in-flight anomaly on another planet. Fifty-four seconds in, a dropped navigation image sent bad timestamps through the pipeline and the helicopter started rocking up to 20 degrees. You can watch it happen over the real Jezero Crater landscape, follow the charts underneath, and read what a language model makes of the flags it gets handed. Every number on this page is computed rather than generated, and the model only writes the prose around them.
Data: the public Ingenuity (Mars 2020 helicopter) flight log (72 flights, used for the fleet envelope statistics), compiled from NASA status reports via Wikipedia's List of Ingenuity flights. Terrain: the real Jezero Crater landscape from the USGS Mars 2020 CTX DEM and orthomosaic (NASA/JPL-Caltech/USGS, public domain), a 10 km window centered on Octavia E. Butler Landing. The flight path and channel curves are reconstructed from the recorded summary: deterministic, anchored to the recorded numbers, dramatized where the record describes the event, and always labeled as reconstruction. Code and README on GitHub.
Drag to orbit · scroll to zoom
How the narrative layer is evaluated (golden set, LLM-as-judge, traces)
The golden set is 22 labeled cases from the public flight record. Exact match checks the predicted anomaly category against the label; an LLM-as-judge pass (Claude, cached run) scores each narrative 1 to 5 for category correctness, grounding, and calibration. The match rate is high because the deterministic layer is doing most of the work here, which is the whole idea. The judge score is the number worth reading. Click a row for its full trace: inputs, flags, exact prompts, output, reference, verdict.
| Flight | Expected | Predicted | Match | Judge |
|---|
View the raw flight data as a table
What this demo shows
The internal TRIAGE was built at the JPL hackathon to disposition Mars Sample Return helicopter anomalies, with a flight deck that replays reconstructed flights over real Jezero Crater terrain. This public rebuild keeps that architecture on open data and tells one story with it: Flight 6, the sol 91 navigation anomaly that Ingenuity survived. The deterministic layer computes fleet statistics and envelope flags entirely client-side from all 72 public flights, the replay runs over the same USGS Jezero terrain data the original used, the reconstruction is seeded and anchored to the recorded numbers with the event dramatized where the record describes it, and a fenced language model turns the flags into an example narrative through a server-side API route (the key never reaches the browser). The rule I carried over from the original is that every statistic gets computed, the model only writes the prose around it, and the person reviewing makes the call.