TDF SENTINEL
NAVIGATOR'S LOG R&D
MISSION COMPLETE — CREW SAFELY HOME

We taught an AI what smooth cruising looks like.
Then we asked: when does the ride get rough?

NASA's Orion spacecraft flew to the Moon and back. We fed its real trajectory data into a detector that learns the pattern of calm, steady flight. Before the mission, we locked a prediction. Now the mission is over — here's what the detector found, and what it missed.

PREDICTION LOCKED APRIL 5, 2026 AT 01:37 UTC · FLYBY APRIL 6 · SPLASHDOWN APRIL 10, 2026
ORION'S FLIGHT PATH

Where Orion Went

This is the real flight path, plotted from 427 position measurements published by NASA's Jet Propulsion Laboratory. Hover anywhere on the path to see distance and speed at that point. The crew set a new distance record of 252,756 miles (406,771 km) — surpassing Apollo 13's record by 4,111 miles. Splashdown: April 10, 2026, 00:07 UTC.

DAY 0.0
0 km
from Earth

Earth is at center. The dashed ring is the Moon's orbit (~384,400 km out). Orion's initial orbits around Earth are too small to see at this scale. The small loop near Earth is the crew practicing manual spacecraft handling before heading to the Moon. Data source: JPL Horizons, 30-minute intervals.

COMM BLACKOUT
--
Orion passes behind the Moon
DISTANCE RECORD
--
Apollo 13: 400,171 km (248,655 mi)
MISSION TIME
--:--:--
days : hrs : min
EARTH DISTANCE
---
kilometers
SPEED
---
km/s
MISSION PHASE
--
--
SPACE ENVIRONMENT

Was the Crew Protected?

Earth's magnetic field creates an invisible shield called the magnetosphere, extending roughly 70,000 km into space. Beyond that boundary, the crew is directly exposed to the solar wind and any radiation the Sun throws at them. The crew spent most of the mission well outside the magnetosphere. The translucent blue bubble on the trajectory map shows the protected zone they left behind.

Calculating...
Determining Orion's position relative to the magnetosphere

Live Solar Conditions

Real-time data from NOAA's space weather satellites. These numbers directly affect the crew's radiation exposure outside Earth's magnetic shield.

SOLAR WIND SPEED
--
km/s
WIND DENSITY
--
protons/cm³
Kp INDEX
--
geomagnetic activity
PROTON FLUX
--
>10 MeV particles

Data: NOAA Space Weather Prediction Center (DSCOVR satellite at L1). Updates automatically.

THE PREDICTION

How Unfamiliar Does Each Phase Look to the AI?

🛣️
Think of it like a driving student. You train someone exclusively on straight, flat highway driving. Then you put them on a mountain switchback road. They'll grip the wheel harder, brake at unexpected times, and feel out of their depth. The more different the road is from what they learned, the more "surprised" they'll be. That's exactly what we're measuring. A value of 1.0 means "this feels like the highway." A value of 2.45 means "this is very different from anything I trained on."

Click any bar below to learn what's happening during that phase and why the AI reacts the way it does.

--

--

--

Cyan bars = phases the AI trained on (the "highway"). Red bars = phases the AI has never seen (the "mountain road"). Taller bars = more surprise.

YOUR TURN

Make Your Own Prediction

Before the results come in, lock in what YOU think will happen. How much will the flyby and reentry surprise the detector? Drag the sliders, then lock your prediction. After splashdown, you'll see how close you were.

LUNAR FLYBY Our prediction: 2.45x
2.45x
EARTH REENTRY Our prediction: 2.87x
2.87x
HOW WE DID THIS

Four Steps from Raw Data to Prediction

The detector has no knowledge of physics, gravity, or the Moon. It only sees numbers changing over time. Here's how it works:

01
Download the flight data
427 position-and-velocity snapshots of Orion from JPL's Horizons system. These are the same navigation vectors NASA uses. Updated every 30 minutes.
02
Measure the shape of the motion
A geometric encoder looks at sliding windows of the trajectory and computes 8 numbers that describe the "shape" of the motion: how curved, how complex, how spread out, how fast-changing. These work on any system, not just spacecraft.
03
Teach the AI what "normal" looks like
A neural network trains on Days 0 through 2 (the outbound coast). It learns the pattern: "when you see this shape, the next shape will look like this." That's its entire model of normal flight.
04
Ask it about the rest of the trip
For each later phase, the AI tries to predict the next shape. When it's wrong, that means the flight dynamics have changed. The bigger the prediction error compared to an honest out-of-sample baseline, the more unfamiliar the phase is.
THE RESULT

We Were Wrong. That Was the Point.

The detector found 3.51x surprise at the lunar approach — above our 6.0x GENUINE threshold? No. The initial SOI crossing test returned 0.80x BELOW. The model trained on deep-space coast dynamics generalized smoothly right through the lunar sphere of influence.

That’s not a failure of the tool. That’s the paradigm boundary revealing itself. Smooth Newtonian gravity gradients don’t produce precursor geometry — the detector has no shape change to catch. Contrast this with Parker Solar Probe (25.6x GENUINE at the Alfvén surface, a genuine plasma phase boundary) and Voyager 2 (12x GENUINE at the Jupiter gravity assist, an energy-exchanging encounter).

Artemis II sits exactly where it belongs: the paradigm boundary case. The result that looked like a failure was actually the most useful data point in the set. Smooth gravity → BELOW. Energy-exchanging gravity assist → GENUINE. Plasma environment boundary → GENUINE.

PREDICTED (FLYBY)
2.45x
TDF SENTINEL, Apr 5
PREDICTED (REENTRY)
2.87x
TDF SENTINEL, Apr 5
ACTUAL (SOI TEST)
0.80x
BELOW — Apr 10
MISSION STATS — ACTUAL
SplashdownApril 10, 2026 — 00:07 UTC Distance record252,756 mi / 406,771 km Lunar closest approach4,067 miles above surface Total distance traveled694,481 miles Re-entry speed24,661 mph (Mach 33) CrewWiseman, Glover, Koch, Hansen