Warning systems, inspection cameras, vibration alarms and medical monitors are tuned once, for the conditions they were built in. When those conditions drift, they keep running and keep reporting the same confidence. The Overseer works alongside the method you already have and tells you when its output should no longer be trusted, without changing the method itself.
Founder: Edward Matte — Axiometrics, Inc.
My background is graphic arts and prepress, not engineering. Axiometrics grew out of carrying patterns from that work into signal processing, and from one observation that holds in every industry I have looked at: the costly failures are not the ones a system reports, but the ones it does not know it is making.
Everything on this page was tested under a plan written and handed to an outside academic reviewer before the test ran, and is reported whether it passed or not. The failures are listed next to the successes.
An asset added to an existing method, not a replacement for it.
It observes the conditions an existing method is operating in, continuously, alongside it.
It learns from the outcome records an operation already keeps: what actually happened afterwards.
It identifies the conditions under which the existing method tends to be right, and the ones under which it tends to be wrong.
Replacing or recalibrating a working method is expensive, and in regulated settings it means validating it all over again. The Overseer leaves the method and its software untouched. What it adds is knowledge of when to rely on the method and when to look twice.
Any operation that runs an established method and keeps a record of what actually happened afterwards. Each domain is marked by where it stands.
Pre-registered test against National Weather Service warnings and what actually followed them.
Pre-registered test on fire department records, a domain unrelated to weather.
Inspection and process settings that drift with stock, ink, press and humidity.
Vibration and condition monitoring, where alarms tuned on one machine misread another.
Monitors whose reliability changes with patient movement, electrode contact and noise.
If you keep records of what your system said and what really happened, it is a candidate.
“Next candidate” and “Proposed” are domains under consideration, not results.
Every figure comes from a test whose plan was fixed, hashed and held by an outside academic reviewer before it ran. Caveats sit with each figure, not below the fold.
Both tests were registered before they ran and passed under their own pre-stated criteria. Weather and fire share no data, no physics and no equipment, which is why they were chosen as the first pair.
Full write-ups with figures will be posted here. Two domains is a start, not a demonstration of generality, and neither test yet involves a paying operation.
SNR-X chooses, moment by moment, which treatment to apply to a noisy signal and how much of it, including leaving the signal alone. It learns offline and needs no clean reference while running. It is where the discipline on this page was developed.
ST error falls from 147.14 to 123.95 µV and T-wave amplitude error from 318.05 to 261.49 µV, both in 15 of 15 patients, both p = 0.00003 under Holm at a family-wise error rate of 0.05. One-sided exact Wilcoxon, patient-level.
p = 0.00003 is the floor of the exact test at n = 15, so it reflects perfect consistency of direction rather than a large effect; the residual 124 µV error still exceeds the ±100 µV magnitude clinicians care about; and this is error reduction in clinical units, not a demonstrated change in clinical decisions.
The registration originally tested whether SNR-X changes the ±100 µV ST call. A variance pilot on five non-benchmark records showed 53.7% of beats in the clean signal already cross that threshold, so on this database it is not a clinical decision. The endpoint was withdrawn in the registration rather than reinterpreted afterwards.
p = 0.0022, 13 of 15 patients, MIT-BIH Arrhythmia with recorded artifact, leave-one-patient-out. Against six tuned external baselines: +1.447 dB, p = 0.0022, 14 of 15 patients.
15 of 15 held-out blocks, p = 1.2 × 10⁻¹⁴, capturing 85% of the available headroom. EEGdenoiseNet.
The ECG and EEG figures reproduce bit-exactly by rebuilding from the raw public corpora, not by re-reading stored results.
These are here for the same reason the results above are. A figure is worth something only when the failures sit next to it.
Published heartbeat detectors already exceed 99% accuracy on this data, so there was little room for any method to improve them. The null stands as reported. Improving a signal is not the same as improving what someone does with it.
Two attempts to let SNR-X judge its own output while running failed (−0.026 dB, p = 0.28; −0.033 dB, p = 0.40). This is the gap the Overseer was designed to fill: rather than a system grading itself, a separate layer checks it against what actually happened.
An explanation of where SNR-X’s advantage comes from was tested, not supported at n = 15, and withdrawn in writing to the reviewer holding the registration.
If your operation records alerts or decisions and what actually happened afterward, those records may be enough to test the Overseer without changing the system you already use.
Next step: let’s see whether your historical records can support a pilot.The Overseer now needs to move from tested design toward a working product. Axiometrics is looking for engineering expertise to help lead that build.
Next step: talk about the engineering challenge and product path.Start with the evidence, including the failures. Axiometrics reports null results and withdrawn claims alongside successful tests so the record can be judged as a whole.
Next step: review the test record and the path to a first operating pilot.Axiometrics, Inc. is a Massachusetts C-corporation. If you have historical outcome records, engineering expertise, or a research setting where the idea can be tested honestly, I would like to hear from you.