SENTINEL / E4

Root-cause discrimination for UAV telemetry

measured2026-08-15
tests215 offline
rests on3 flights
licenceApache-2.0

The first alarm is not the fault

A drone fault where the true cause is reported one second after the symptom it caused — so the free rule “the first alarm is the cause” is wrong by construction, and anything more expensive has to earn its cost.

WATCH

The fault, in twelve seconds

The compass is what breaks. Watch which alarm arrives first — and which one the free rule grabs. Open it full-size.

0.00 What the free rule scores on this fault. Ten-plus runs, two model families, every configuration tested, no exceptions. A test the cheap answer passes measures nothing.
READ

Three depths, same evidence

PageFor
Why the first warning lies No jargon. The argument, a single diagram, and the limits. Start here if you do not work on drones or evaluation.
The first alarm is not the fault The general write-up: the mechanism, what two named models did with it, and every claim that was retracted.
The discriminating pair Method, measurements and limits. Experimental design, detector state, integrity model, and the standing of each claim.
The one-second gap A 12-second loop. The gap is a duration, and this spends it rather than describing it.
USE

Check it yourself

hardwarenot needed
api keynot needed
time~5 minutes

The flights are captured and committed, so the headline result reproduces offline on your own machine — no drone, no simulator, no account.

git clone https://github.com/RahulRajelli/sentinel-realtime
git clone https://github.com/RahulRajelli/ardupilot-log-analyzer
pip install -e sentinel-realtime && pip install -e ardupilot-log-analyzer

cd sentinel-realtime
pytest -q                                 # 215 passed
python scripts/e4_report.py --bundles bundles \
  --verdicts results/crossmodel/gpt2_run1.json --only compass_offset

Full walkthrough with expected output at every step: QUICKSTART.md. Building a test like this in your own domain: METHOD.md — nothing in it is about drones.

3 Flights. Everything here about AI judgement rests on them. The fault itself is robust and I will defend it. What it implies about models is not established — stated here rather than left for you to find.