Horos Engineering · proof of code · structural control

Run the precursor detector yourself

The element of the structural rail — what it is built from, what is gated, and the three records it stands on.

The same governed detector, on a different substrate. This is the structural precursor test running in your browser against NASA C-MAPSS turbofan run-to-failure data and FEMTO/PRONOSTIA bearing run-to-failure data — both public, both third-party, both run until the unit actually failed. Pick a unit. Move the gates. Watch the lead time and what it costs.

200 engines across two independent C-MAPSS subsets and six run-to-failure bearings. Constants were declared in code before the results they govern, and there was no per-dataset tuning — the flight-cycle substrate and the vibration substrate carry different constants because they are different substrates, and both sets were fixed in advance. The figures below are recomputed in this page from the raw series, not quoted.

The instrument #

Across the whole fleet #

The declared row is highlighted. Narrow the gate and detection rises while the baseline starts alarming on healthy units. Widen it and the alarms stop along with some of the detections. That trade is the instrument.

What an adversary does to it #

From the archived run, on the flight-cycle substrate. A crude replay is caught almost every time. An adaptive one is caught most of the time. A statistically faithful overwrite forgery evades a telemetry-only observer in the majority of trials — that is the named boundary where the adversarial ladder closes, and it is the reason the architecture carries an identity channel that does not depend on the telemetry at all.

Reproduction check #

What this is not #

Detector: causal rolling median, baseline fixed at [run-in, run-in+B), gate at median + k×MAD of that baseline, onset declared after P consecutive exceedances. Data: NASA C-MAPSS turbofan degradation and FEMTO/PRONOSTIA bearing datasets, archived and hashed; the record cites both by SHA-256.