Experiments
We publish what we learn — including what failed.
Every experiment follows the same shape: hypothesis, setup, result, learning. Authority comes from proof, not from promotional volume.
HYPOTHESIS/SETUP/RESULT/LEARNING
The format / worked example
FERMI.EXPERIMENT 0047
ILLUSTRATIVE · NOT A PUBLISHED RUNDEVICE FC-EDGE-01/REV A.04/THERMAL_STRESS
HYPOTHESIS
Can an AI agent diagnose a hardware fault using telemetry alone?
SETUP
MCU · current sensor · temperature probe · test bench. 40-minute load profile, 1 kS/s on four channels, no human in the loop after start.
RESULT
Thermal rise began at 22 minutes and repeated under high load. The agent localised it to the power stage at 87.4% confidence and proposed more copper, a thermal via array and a lower switching frequency.
LEARNING
Telemetry was enough to name the subsystem — but not the root cause. The next run adds a thermal camera, because “where” and “why” are different questions.
This is the shape every write-up takes, walked through end to end. It is a worked example of the format — the first published run replaces it, traces and all.
Archive
The notebook is open.
Nothing published yet. Every run lands here in the same four fields, so results stay comparable across months, benches and boards — and so a failed experiment is as readable as a successful one.
FIELD 01
Hypothesis
FIELD 02
Setup
FIELD 03
Result
FIELD 04
Learning — what changed in the engineering system
The field that makes the archive worth keeping. Without it, a run is only a log file.
Method
Show the measurement, the failure and the surprise.
An experiment that only confirms what we already believed is not worth publishing. We end every write-up with what changed in the engineering system.
RUN 0047 / PROBE FC-P4 · 4 CHANNELS · 1 kS/s