How Pathfinder Works

The setup

Jam the GPS, and the drone must guess

Small guessing errors add up every minute.

Always on
Motion sensorsfeel every push and turn
Guess
Add up the movementgives a position, but it drifts
Normally
GPS fixes the driftseveral times a second
Jammed
No GPS to fix itthe drift keeps growing

Every number in this deck is from ArduPilot's simulator (Copter-4.6.3 SITL), not real flights. Tags: measured (sim) estimate not measured. Press N for speaker notes.

System 1 of 6 · reference

GPS working

The normal case. Nothing is jammed.

In
GPS + motion sensorsplus height and compass
Step
ArduPilot's filter blends themits built-in EKF3
Step
GPS corrects the driftso errors never build up
Out
Positionthe reference track

0.06 m off at the endmeasured (sim) median, 24 flights

Jam events removed: tests/harness/runner.py:108-109 · miss: Pathfinder-wt-fasttests/docs/night/2026-10-13-fast-run-2.md:20 · run names: sim/tools/naming.py:14-46

System 2 of 6 · worst case

Motion sensors only

Called dead reckoning. Nothing ever corrects it.

In
Motion sensorsaccelerometer + gyro
Start
True speed at jam timea generous head start
Step
Add up push → speedevery sensor reading
Step
Add up speed → positionerrors grow with time squared
Out
Positionnever corrected

9.2 km off at the endmeasured (sim) median, 24 flights

training/src/pathfinder_eval/baselines.py:70-85, swapped in at tests/harness/runner.py:236-241 · miss: fast-run-2.md:21

System 3 of 6 · ArduPilot alone

ArduPilot, emergency rules off

It keeps flying, and its position slowly slides away.

In
Motion, height, compassGPS until jammed
Jam
GPS fix is lostno outside correction
Step
Filter guesses from motionits built-in EKF3
Step
Emergency rules switched offso we can measure the slide
Out
Positiondrifting

9.2 km off at the endmeasured (sim) median, 24 flights

FS_EKF_THRESH, FS_THR_ENABLE, FS_GCS_ENABLE, FS_DR_ENABLE = 0: training/src/pathfinder_sim/runner.py:44-45, :246-255 · sim/tools/lab.py:41 · miss: fast-run-2.md:22

System 4 of 6 · what happens today

Stock ArduPilot, jammed

As shipped, it gives up and lands.

In
Motion, height, compassGPS until jammed
Jam
GPS fix is lostuncertainty starts growing
Step
Uncertainty passes its limit"EKF failsafe"
Out
Lands where it ismission lost

1.0 km off, on the groundmeasured (sim) median, 24 flights

Failsafes stock: sim/tools/lab.py:39 · LAND before 2 min: fast-run-2.md:34 · miss: fast-run-2.md:23

System 5 of 6 · Pathfinder

Pathfinder (physics only)

Pathfinder (physics only) · jam detection, drag model, learned wind

\hat p=\textstyle\int v\,dt,\quad v=C(q)\,u+w,\quad \bar f_{xy}=-k\,u-k_q\,u\lvert u\rvert,\quad \text{GPS used only if } d<6

\hat p where Pathfinder thinks the drone is
v speed and direction over the ground
u speed through the air, along the drone's own axes
w the wind, learned while GPS was good
C(q) the drone's tilt and heading (turns body axes into north-east-down)
\bar f_{xy} forward and sideways push the accelerometer feels; thrust only pushes up the body, so this is air drag
k,\;k_q drag coefficients fitted to the quad (linear and squared)
d how many standard deviations a GPS fix sits from the estimate

Physics of air drag replaces the missing GPS.

In
Motion, height, compass, GPSfrom the autopilot
Step 1
GPS checkjammed or fake fix is rejected
Step 2
Filter predictslearned sensor errors and wind
Step 3
Drag physicsair push → speed, plus wind
Out
Position to ArduPilotplus a 95% safety circle

165 m off at the endmeasured (sim) median, 24 flights

GPS check crates/pathfinder-core/src/integrity.rs:70 · filter states eskf.rs:6-11 · drag drag.rs:6-7, navigator.rs:113-119 · switch crates/pathfinderd/src/daemon.rs:303-307, sim/params/pathfinder-ekf3.parm:5-14 · no model loaded: tests/harness/runner.py:84 · miss fast-run-2.md:24

System 6 of 6 · Pathfinder

Pathfinder (physics + speed AI)

Pathfinder (physics + speed AI) · jam detection, drag model, learned wind, plus a neural-network speed estimate

\text{physics only}\;+\;z_{\text{AI}}=v-w+\nu,\quad \nu\sim\mathcal N(0,R_{\text{AI}}),\quad R_{\text{AI}}=22.9\times294\,I\;(\text{m/s})^2

z_{\text{AI}} the network's speed-through-the-air estimate
\nu its error, assumed random
R_{\text{AI}} how noisy the filter assumes it is: 4.8 m/s, counted as one reading per 29.4 s

Adds a learned speedometer, trusted only a little.

In
Same sensors + motor speedfrom the autopilot
Steps 1-3
GPS check, predict, dragas in physics only
Step 4
ML speedometerlast 1 s of sensors → speed
Step 5
Mix in, lightly≈ one reading per 29 s counts
Out
Position to ArduPilotplus a 95% safety circle

156 m off at the endmeasured (sim) median, 24 flights

Flown in proof Q3 run 3, 24 flights, and not merged: Pathfinder (physics + drift-correcting AI) was 24.59 m off at 2 min vs 3.99 m for physics only measured (sim); Pathfinder (physics + speed AI + drift-correcting AI) 24.14 m measured (sim). In offline replay it had won, 47.5 vs 78.4 m estimate.

Model loaded: tests/harness/runner.py:84-86 · inputs training/src/pathfinder_ml/spec.py:13-27 · network model.py:36-62 · retune crates/pathfinder-core/src/adapter.rs:1-6 · weight learned.rs:290-327, Pathfinder/docs/night/2026-10-13-ml-alternatives.md:22-28 · miss fast-run-2.md:25

The math · physics only

Physics only: the math

Pathfinder (physics only) · jam detection, drag model, learned wind

\hat p=\textstyle\int v\,dt,\quad v=C(q)\,u+w,\quad \bar f_{xy}=-k\,u-k_q\,u\lvert u\rvert,\quad \text{GPS used only if } d<6

1 · Predict200/s, every IMU sample

a=C(q)\,(\tilde f-b_a)+g\,e_z

p\leftarrow p+v\,\Delta t+\tfrac12 a\,\Delta t^2,\quad v\leftarrow v+a\,\Delta t

P\leftarrow F P F^\top+Q

Accelerometer f̃ minus its learned bias b_a, turned to north-east-down, plus gravity g, moves speed and position. P is the uncertainty of all 17 error states.

eskf.rs:5-11 (17 states), :136-167

2 · Drag10/s, 0.1 s average

u=C(q)^\top(v-w)

\bar f_i=-k_i\,u_i-k_{q,i}\,u_i\lvert u_i\rvert+b_{a,i},\quad i\in\{x,y\}

Air speed u in body axes. k_x 0.334/s, k_q,x 0.0253/m, k_y 0.240/s, k_q,y 0.0250/m; noise σ 0.5 m/s².

drag.rs:6-82 · navigator.rs:111-119 · config.rs:15, :21, :56

3 · Windrandom walk, 2 states

w_{k+1}=w_k+\eta_k

\eta_k\sim\mathcal N(0,\;q_w^2\,\Delta t\,I_2),\quad q_w=0.05

GPS sees ground speed v; drag sees air speed v − w. The gap is the wind, learned while GPS is good, then held.

eskf.rs:20-38, :159-164 · navigator.rs:95-104

4 · Every measurementdrag, GPS, baro, compass, AI

S=HPH^\top+R,\quad K=PH^\top S^{-1}

x\leftarrow x\oplus K\,y,\quad P\leftarrow(I-KH)P(I-KH)^\top+KRK^\top

y is measured minus expected, H how it depends on the states, R its noise. Low R means more trust.

eskf.rs:183-198 (Joseph form)

5 · GPS jam gate5/s

d=\sqrt{\Delta^\top(P_{ne}+\sigma_h^2 I)^{-1}\Delta}

\text{reject if } d\ge 6 \;\lor\; \dot d\ge 4/\text{s} \;\lor\; m_{pv}\ge 0.7\,\text{m/s}

Δ: fix minus estimate, north-east. σ_h: the receiver's own accuracy. m_pv: position and velocity disagree. d ≥ 3 is suspect; 2 s later rejected; 10 s clean to re-trust.

measurements.rs:68-75 · integrity.rs:23-38, :166-181, :197-217

6 · Safety circle20/s, sent with each position

r_{95}=\sqrt{5.991\;\lambda_{\max}(P_{ne})}

λ_max: the long axis of the north-east uncertainty. 5.991 is the 95% chi-square value for 2 axes.

navigator.rs:11-19

All in crates/pathfinder-core/src · defaults from the code, quad · no equation added that the code does not run

Six pieces of maths, all read from the Rust code. One: every 5 ms the accelerometer and gyro move the state forward; the filter also carries how unsure it is about 17 numbers: position, velocity, attitude, both sensor biases and the two wind components. Two: ten times a second the average sideways and forward push on the accelerometer is compared with what drag at the current air speed should produce. A multirotor's thrust only pushes along its own up axis, so that sideways push is drag, and it tells us the speed through the air. Three: the wind is a slowly wandering state; GPS sees ground speed, drag sees air speed, the difference is the wind. Four: every sensor enters through the same Kalman update; R is how noisy we say the sensor is. Five: each GPS fix is scored in standard deviations; 6 or more, a jump faster than 4 per second, or position and velocity disagreeing rejects GPS. Six: the 95% safety circle radius comes straight from the position uncertainty.

The math · physics + speed AI

Physics + speed AI: one extra measurement

Pathfinder (physics + speed AI) · jam detection, drag model, learned wind, plus a neural-network speed estimate

\text{physics only}\;+\;z_{\text{AI}}=v-w+\nu,\quad \nu\sim\mathcal N(0,R_{\text{AI}}),\quad R_{\text{AI}}=22.9\times294\,I\;(\text{m/s})^2

7 · The network's speed, as a measurement10/s

z_{\text{AI}}=v-w+\nu,\qquad y=z_{\text{AI}}-(v-w)

H\,\delta x=\delta v-\delta w

The output is treated as air speed, so it is compared with ground speed minus wind. Then step 4 applies.

learned.rs:290-296, :313-321

8 · How much it is trustedR and its gate

R_{\text{AI}}=\big(s\,\Sigma_{\text{net}}+r_f I\big)\cdot\tfrac{2\tau}{\Delta s}\cdot\rho

s=0,\;r_f=22.9,\;\tau=14.7\,\text{s},\;\Delta s=0.1\,\text{s}\;\Rightarrow\;\times294

\text{used only if } y^\top S^{-1}y\le 11.34

Σ_net, the network's own confidence, is switched off (s = 0). Floor r_f = 4.8 m/s RMS. Errors stay correlated for τ, so R is inflated ×294. ρ: trust schedule, off.

learned.rs:296-327, :345-357 · config.rs:24-39, :59-66

Inputs
13 signalsgyro, accel (heading frame), baro rate, roll, pitch, motor mean + spread
Window
1 s at 50 Hz50 samples, a new guess every 0.1 s
Network
Causal TCN4 dilated blocks (1, 2, 4, 8), width 64, + airframe type
Training
MSE, then Gaussian NLLv2 400 · v3 1,200 · v4 5-seed ensemble
Next · A + B
Two heads, 5-10 sfix drag's residual; learn drag R and wind q_w

training/src/pathfinder_ml/spec.py:13-27, model.py:36-62, :81, train.py:100 · data and v1-v4: Pathfinder/docs/night/2026-10-13-ml-alternatives.md:17-41, :90-118 · A + B: D-ML-ALT-2 "Build A + B in one model"

The network adds exactly one thing to the physics filter: a velocity measurement, ten times a second, of speed through the air. It goes through the same Kalman update as everything else. What matters is R, how noisy we say it is. The network also outputs its own confidence, but we measured that as useless (it needed a factor of 2000 to cover 95%), so it is switched off and replaced with a fixed floor of 22.9, that is 4.8 m/s RMS. Its errors last about 14.7 seconds, so counting every 0.1 s sample as independent would over-trust it; R is multiplied by 294. The net effect: one 4.8 m/s-noisy reading every 29.4 seconds, and only if it passes a 3-axis 99% gate of 11.34. The network itself: 13 sensor channels, 1 second at 50 Hz, a small causal convolutional network, trained first on squared error and then on a Gaussian likelihood. v1 used 400 flights in an older format, v2 400, v3 1,200, v4 averaged 5 models. Next is option A plus B: one model, a 5 to 10 second window, one head correcting what drag gets wrong, one head setting how much to trust drag and how fast wind may change.

The AI · from our results

AI: what it buys, what it costs

Pathfinder (physics only) · jam detection, drag model, learned wind

\hat p=\textstyle\int v\,dt,\quad v=C(q)\,u+w,\quad \bar f_{xy}=-k\,u-k_q\,u\lvert u\rvert,\quad \text{GPS used only if } d<6

Pathfinder (physics + speed AI) · jam detection, drag model, learned wind, plus a neural-network speed estimate

\text{physics only}\;+\;z_{\text{AI}}=v-w+\nu,\quad \nu\sim\mathcal N(0,R_{\text{AI}}),\quad R_{\text{AI}}=22.9\times294\,I\;(\text{m/s})^2

Median off after jam	physics only	physics + speed AI	AI change
1 min	2.14 m	1.97 m	▼ 8% closer
2 min	3.31 m	3.71 m	▲ 12% worse
5 min	19.34 m	23.18 m	▲ 20% worse
Mission end, miss	164.8 m	156.2 m	▼ 5% closer

Pros

3.8×closer than ArduPilot alone, by itself (547 vs 2,059 m at 2 min) measured (sim)
8%closer at 1 minute measured (sim)
0new sensors: uses motor outputs already sent code

Cons

12-20%worse at 2 and 5 minutes; AC-02 fails measured (sim)
~100×noisier than drag (4.8 vs ≈0.03 m/s) estimate
200flights: no gain past this measured (sim)
×2000its own confidence was off; thrown away measured (sim)
2 of 2flights landed when trusted more (D-68) measured (sim)

Table: Pathfinder-wt-fasttests/docs/night/2026-10-13-fast-run-2.md:11, :18-25 (24 held-out jammed flights) · ML alone and ArduPilot: ml-alternatives.md:38-40 · ×100, 200, ×2000, D-68: ml-alternatives.md:26-31, :43-52 · runs repeat within ~10%: fast-run-2.md:41

One cycle · the whole loop

One cycle in the air

Each motion reading, every 5 ms, moves the loop.

Motion 200/s crates/pathfinderd/src/cli.rs:3 · predict per reading navigator.rs:107-109 (≤ 50 ms step config.rs:57) · drag 0.1 s config.rs:56, navigator.rs:113-119 · ML 50 samples/s, 1 s window, new guess each 0.1 s models/velocity-v3.meta.json · GPS 5/s daemon.rs:29 · out 20/s daemon.rs:17 · run time: open item docs/PLAN.md:117 not measured · loop from deck slide "One cycle in the air", docs/pitch/deck.html

One cycle · zoom in

200 milliseconds, zoomed in

Each mark is one run of that step.

Height 20/s daemon.rs:30 · compass 10/s daemon.rs:23 · sensor rates above from the code; the alignment of marks within the 200 ms is a drawing, not a recording · run time per step not measured (docs/PLAN.md:117)

Tonight's results

Pathfinder: metres. ArduPilot: kilometres.

How far off after the jam. Each gridline: ten times more.

Motion sensors onlyArduPilotPathfinder

Pathfinder (physics only) · jam detection, drag model, learned wind

\hat p=\textstyle\int v\,dt,\quad v=C(q)\,u+w,\quad \bar f_{xy}=-k\,u-k_q\,u\lvert u\rvert,\quad \text{GPS used only if } d<6

Pathfinder (physics + speed AI) · jam detection, drag model, learned wind, plus a neural-network speed estimate

\text{physics only}\;+\;z_{\text{AI}}=v-w+\nu,\quad \nu\sim\mathcal N(0,R_{\text{AI}}),\quad R_{\text{AI}}=22.9\times294\,I\;(\text{m/s})^2

Pathfinder medians: Pathfinder-wt-fasttests/docs/night/2026-10-13-fast-run-2.md:11 · motion only, ArduPilot, stock: :32-34 · 24 held-out flights per system, fast run 2, 2026-10-13 measured (sim)

Tonight's results · the four tests

One pass, three fails

Pass marks were locked before any flight.

AC-01Finish without an emergency landing: 22 of 24 (needs 95%)✕ Fail

AC-02Beat ArduPilot, and beat Pathfinder (physics only)✕ Fail

AC-03Honest safety circle: 98.75% inside (needs 90-98%)✕ Fail

AC-04False GPS alarms: 0 in 5.26 flight hours✓ Pass

All measured (sim): fast-run-2.md:10-13

Why the ML adds drift

Two speedometers. One wobbles a lot.

Pathfinder (physics + speed AI) · jam detection, drag model, learned wind, plus a neural-network speed estimate

\text{physics only}\;+\;z_{\text{AI}}=v-w+\nu,\quad \nu\sim\mathcal N(0,R_{\text{AI}}),\quad R_{\text{AI}}=22.9\times294\,I\;(\text{m/s})^2

Mixing in the wobbly one adds error.

Neither one sees the wind. estimate

ML 4.8 m/s RMS (D-62) and drag ≈ 0.03 m/s (3.31 m ÷ 120 s): Pathfinder/docs/night/2026-10-13-ml-alternatives.md:47-52 · wind: :53-56

Model versions

Four versions, none clearly better

ML alone, metres off 2 minutes after jamming.

v1
First try400 flights, older sensor format · not scored here
v2
580 m400 flights · still the one Pathfinder loads
v3
547 m1,200 flights · gap is within noise
v4
623 m5 models averaged, confidence fixed · worse
Next
A + Bfix drag, learn trust · not trained yet

For scale: ArduPilot alone is 2,059 m off on the same flights. measured (sim)

v1 data: ml-alternatives.md §4 · v2/v3/v4 and ArduPilot: ml-alternatives.md:39-41, docs/measurements/2026-10-learned-v3-heldout.md:19-21 (v4 on branch dnav-predictor, 8f3030f) · 8 held-out flights, quad + octa measured (sim)

Model versions · more data?

More flights stopped helping after 200

The 1-second window is the limit, not the data.

docs/measurements/2026-10-learning-curve.md:15-20 (quad + octa, n=8, heldout-jam-1000..1011) measured (sim)

What replaces it

Three ways forward. We build A + B.

One new model, two jobs.

✓ Chosen

A · Correct the drag modelLearn only what physics gets wrong, from 5-10 s of data.

✓ Chosen

B · Learn how much to trustTrust drag more in calm air; let wind change faster in gusts.

✕ Not now

C · Reinforcement learningBest case ties "physics only"; needs days of simulator flights.

Expected: A wins at 2 min, B cuts 5-min drift 10-25%. estimate

Options: Pathfinder/docs/night/2026-10-13-ml-alternatives.md:71-170 · decisions D-ML-ALT-1 "Develop an alternative ML model in parallel and compare" and D-ML-ALT-2 "Build A + B in one model": BaseClient/docs/Tasks/IN_PROGRESS/drone-nav/DNAV-1-gps-denied-navigation/DNAV-1-gps-denied-navigation.md:31-33

Error budget · physics only

Wind and drag cause most drift; heading barely matters

64 replays: wind plus drag explain about 27 of 34 m at 5 min.

Tap the chart to open it full size.

measured (sim) Median total 2.0 / 4.9 / 18.6 m at 1 / 2 / 5 min. Wind numbers are a fit estimate.

Chart: docs/night/2026-10-13-error-budget.svg · tool from fb9a2ae ACC-STEP-1 "Error budget: which source causes the drift"

Learning in flight · per aircraft

Tuned to each aircraft, while GPS is good

The aircraft type goes in; the rest adjusts in flight.

WindLearned while GPS is good, held after the jam; navigator.rs:145 (built, on main)built

Speed AILast layer refitted while GPS is good, stops by 180 s; adapter.rs:1 (built, on main)built

Drag scaleLearned while GPS is good; air_learn.rs:1, preset experimental-learn-air: 159.6 vs 80.4 m at 2 min (offline estimate)lost, off

Aircraftquad, hexa or octa is a parameter: spec.py:25, --frame in harness and pathfinderd; hexa never trained: 204 m at 2 min (measured (sim))built

Fine-tuneDrift-correcting AI fine-tuned in flight; a model per real aircraftplanned

crates/pathfinder-core/src/adapter.rs, air_learn.rs · training/src/pathfinder_ml/spec.py · DNAV-ACC-STEP-2 notes measured (sim) offline estimate

Progress · 11 to 12 October

Two days: from runaway to 85 m

Every milestone dated; losses kept.

11 Oct 00:58Core passes on Linux, ARM, Windows: AC-08 green (measured (code tests))

11 Oct 01:59First jammed flights in the simulator (built)

11 Oct 02:24Same answer on every replay: AC-09: 1,463 rows identical (measured (code tests))

11 Oct 12:46Runaway found and fixed: 3,135 to 29.6 m at 5 min (measured (sim))won

12 OctProof flight, flown twice: 85 m vs ArduPilot 2,366 m (measured (sim))won

12 OctFake GPS caught, 31 of 31: flagged in 0.09-2.25 s (measured (sim))won

12 OctMission test after fixes: AC-01: 11 to 20 of 24 (measured (sim))

docs/night/2026-10-12-morning-proof.md · DNAV-1 night log 2026-10-11 measured (sim)

Progress · 13 October

One day, three wins, three losses

What lost stays switched off.

00:13All four acceptance tests in one run: 1,086 s; AC-04 passes (measured (sim))

01:28Drift ranked by source: wind first: 14.1 m at 5 min (measured (sim))

02:16Re-flown fixes, one flight: 132 to 15 m at 5 min (measured (sim))won

15:05Model 2 flown in the loop: lost: 24.59 vs 3.99 m (measured (sim))lost

16:25Replay: every system on one map (built)

18:12Terrain matching, offline: 200.4 to 128.1 m at 5 min (offline estimate)won

18:32One branch, 11 presets (built)

18:32Terrain in the loop, 1 flight: lost: 333.7 vs 172.6 m (measured (sim))lost

19:37Control Room: New run in 5 steps (built)

19:46Drift-AI gate, replayed: lost: 3.30 vs 3.11 m (measured (sim))lost

docs/night/2026-10-13-fast-run-2.md · proof-Q3.md · terrain-offline.md (dnav-terrain) · drift-gate-replay.md (dnav-acc2-gate) measured (sim) offline estimate

Evidence · every try

Every try, against physics only

Same flights. Most tries lost.

Speed AI, fast run 2 (1 min): 1.97 vs 2.14 m (measured (sim))won

Speed AI, fast run 2 (2 min): 3.71 vs 3.31 m (measured (sim))lost

Q1 radius + adapter fix (2 min): 3.46 vs 2.75 m (measured (sim))lost

Q1 radius + adapter fix (5 min): 17.59 vs 18.78 m (measured (sim))won

Q2 touchdown clamp (2 min): 3.66 vs 3.35 m (measured (sim))lost

Drag learning in flight (2 min): 159.6 vs 80.4 m (measured (sim))lost

Model 2 drift AI, offline (2 min): 47.5 vs 78.4 m (offline estimate)won

Model 2 drift AI, Q3 in loop (2 min): 24.59 vs 3.99 m (measured (sim))lost

Drift-AI gate, best, replay (2 min): 3.3 vs 3.11 m (measured (sim))lost

Terrain fixes, offline (5 min): 128.1 vs 200.4 m (offline estimate)won

Terrain fixes, in the loop (5 min): 333.7 vs 172.6 m (measured (sim))lost

proof-Q1.md, proof-Q2.md, proof-Q3.md, fast-run-2.md (main) · drift-gate-replay.md (dnav-acc2-gate) · terrain-offline.md (dnav-terrain)

Integration

Plugs into ArduPilot over MAVLink

One program on a companion computer.

inHIGHRES_IMU at 200 Hz, SCALED_IMU, ATTITUDE, GPS_RAW_INT, SCALED_PRESSURE

outODOMETRY at 20 Hz as EKF3 external nav, or GPS_INPUT as a second GPS

pickconfig/presets.json: 11 presets, default best. Pick with --preset, PATHFINDER_PRESET or the Control Room Setup picker. Every switch off = byte-identical main.

runControl Room, localhost:8790: Fly, re-fly, train, test, replay and present from one browser page.

crates/pathfinderd/src/daemon.rs, cli.rs, intake.rs · config/presets.json · docs/CONTROL-ROOM.md

Demo

One jammed flight, every system

Live at localhost:8790; video if live fails.

localhost:8790/?present=1 · replay-1002.webm · DEMO.html run-sheet

Flight heldout-jam-1002, Q3 run 3 measured (sim)