The Overseer sits above a site’s existing alarm system without changing it. It records each alarm, the operating conditions when it fired and what was later found, and learns in which conditions those alarms can be trusted. It advises; it never overrules the existing system, and a person always decides.
Founder: Edward Matte
Most sites already have a method for raising alarms, and it works well in some conditions and less well in others. The Overseer is not meant to replace that method. It is meant to sit beside it and keep score, so that the people using it know when to trust it and when to look twice.
Its method has been tested once, on public weather-warning data, under a plan that was fixed and acknowledged by an outside academic reviewer before any outcome was opened. The result was positive and modest, and it says nothing yet about factory or machine alarms. That needs a site’s own records, and that is the next step.
I came to this from graphic arts and prepress rather than from engineering. That is why this page is written the way it is: every claim is tested against a prediction written down first, and the failures are reported beside the successes. The same rules were applied to SNR-X, the company’s earlier work, described further down this page.
It keeps score on an existing alarm system — recording each alarm, the conditions it fired in and what was found — and advises in which conditions those alarms can be trusted, without changing the system or overruling the people who use it.
One registered test so far, on public data. The figure and its bounds are stated together.
Its method has now been tested once on real data. No factory dataset of the size needed is yet in hand, so the first test used a large public record where every alarm already has an outcome: 355,464 U.S. National Weather Service tornado and severe thunderstorm warnings, 2010 to 2025, from 116 forecast offices, each marked as confirmed or not by storm reports.
On forecast offices never used in fitting, four operating conditions set in advance predicted which warnings were confirmed better than the warnings’ own details did: a 1.21% improvement in Brier score against a minimum of 1% fixed in advance, ahead of all 1,000 shuffled controls (p = 0.001). The method also recovered, unaided, the published finding that tornado warnings at night are confirmed less often. Fitted on 2010 to 2019 and scored on 2020 to 2025, the improvement was 1.35%.
Pre-registered. The test plan was fixed, hashed with SHA-256 and acknowledged by an outside academic reviewer on 18 October 2026, before any outcome was opened, and run once.
What it does not show: anything about factory or machine alarms, which still need a site’s own data, or that forecasters are worse in any condition. A warning is confirmed only if someone reports the storm, so the result reflects reporting as well as forecast skill. The effect is modest and real; it is evidence about the method on weather-warning data only.
An earlier offline pilot, not registered, found that a site learning only from its own delayed outcomes gained nothing. It was stopped before registration, and no figure from it is claimed as a result.
The Overseer is not built as a product. In the design, raw data would stay on the customer’s site, only a fixed, pre-defined set of information would ever be shared, and the customer would see everything sent.
The next test needs a site’s own records: alarms that were inspected, with what was found written down. Anything further claimed for the Overseer will be reported under the same pre-registration rules as everything else on this page, pass or fail.
Earlier work
SNR-X chooses, window by window, which treatment to apply to a signal and how much of it to apply — including the option of leaving the signal alone — in settings where the best treatment changes faster than any fixed setting can track and no clean reference is available at runtime.
SNR-X came first and is no longer the company’s main work. Its results are kept here in full because they were produced under the same rules the Overseer is now held to.
Start with the eight failures below. The interesting problem is the one that is still open: scoring an output at runtime with no reference.
What I want: if you have attacked reference-free quality estimation before, I want to hear what failed. Eight of my own attempts have.Start with the ST and T-wave errors, then read the bounds under them. Single lead, MIT-BIH, error in microvolts — not a change in clinical decisions.
What I want: a measurement task where a 20 µV error reduction changes what someone does. I have not found one. If you know of one, that is the most useful thing you could tell me.Start with the beat-detection null. Everything else on this page is easier to believe once you have seen which result is missing from it.
What I want: I am not raising. The gap is a demonstrated decision change, not a larger decibel figure — so the useful conversation is which decision would have to move before this is worth anything.I built SNR-X for domains where the signal matters but the conditions refuse to sit still. A fixed filter has to be chosen once, for a range of conditions it will then meet in whatever order they arrive. On the benchmark below, the best single fixed filter in the same candidate bank scores +0.059 dB; choosing per window scores +1.456 dB. That comparison now carries a qualification, stated with the figure below.
Every figure below comes from a frozen build and re-verifies from the public corpora. Some of them also ran as pre-registered branches, hashed and held by an outside reviewer before any result existed — where that is true it is stated with the figure, and where it is not true it is stated too.
The most important result is the one that did not move: downstream beat detection. It is the reason the commercial question is still open, and it is the first thing I would want a sceptical reader to see.
I came to this from graphic arts and prepress rather than from engineering, and I have been at it for three and a half years. That is also why the rest of this page looks the way it does. I cannot ask anyone to take my word for a result, so every figure here is frozen and re-runnable, the failures sit next to the successes, and the predictions were held by someone outside the company before the runs.
It decides, window by window, which of a bank of candidate treatments to apply to a noisy signal — or none — trained once offline against clean references, and needing neither clean reference nor labels at runtime.
The distinction that matters is between what happens once, offline, and what happens on every window at runtime.
SNR-X is trained once per domain, offline, against clean reference recordings. At runtime it works from the noisy signal alone — no clean reference, no labels, and no score of its own output. The cost of bringing it to a new domain is therefore one clean-reference corpus, obtained once: not a second sensor, not a reference channel, not anything that has to exist in the field.
A September 2026 read-through of the frozen build confirmed that boundary holds in the shipped code. The runtime path reads the noisy window and nothing else; the clean reference enters the training target alone.
How the selection itself is made is not described on this page.
Nothing in that description is specific to weak signals, or to medicine. It is per-instance algorithm selection under changing conditions, and the domain enters only through the candidate bank and the training corpus.
A fixed filter is one choice, made in advance, for conditions that do not hold still.
On the ECG benchmark below, the best single fixed filter in the same candidate bank scores +0.059 dB. Choosing per window scores +1.456 dB, against the same reference and the same folds. That gap is the whole claim, and it is measured rather than argued.
A caution belongs next to it. How the selector’s behaviour varies across noise regimes has been tested and only partly supported: behaviour does differ by regime, but the stronger claim — that this is demonstrably where the margin comes from — failed its own pre-registered test and has been withdrawn. The margin is established. The explanation for it is not.
A second caution, found in September 2026 and not yet settled. The benchmark scores each output against a reference recording that carries its own baseline offset, and counts removing that offset as damage. Scored with the offset removed, a plain fixed baseline-removal filter does better than SNR-X overall on the same ECG data, although SNR-X still does better on clean signals. That comparison was not pre-registered, so it is a lead rather than a result. Until a registered branch tests it, I do not describe the ECG figure as beating fixed filtering in general. Details below.
A second use has now been tested. The same selector, given a published denoising network as one more option, did better on ECG than the network alone, including on clean signals. That is the role I now think matters most: deciding when, and how strongly, to apply a method someone already uses. It holds on one domain so far, over two published networks of similar strength, with the qualifications stated with the result. On wrist heart rate, pilots in front of a strong published method found no worthwhile gain.
Every figure comes from a frozen build evaluated under a protocol fixed before the run. Caveats are stated with each, not below the fold.
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.
Pre-registered. The registration was hashed with SHA-256 and held by an outside academic reviewer before the branch ran, on 12 September 2026.
Bounds worth keeping attached: 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. That comparator was the best fixed filter under the benchmark’s original scoring, which the September scoring check showed to be close to doing nothing; the comparison against a stronger one follows.
Re-checked against a stronger comparator, 17 September 2026. A second pre-registered branch, held by the same outside reviewer before it ran, repeated the comparison against the strongest fixed filter for these measurements: a band-pass filter that also removes baseline, chosen by a fixed rule on five records outside the benchmark. SNR-X still has the lower error: ST by 15.2 µV (95% CI +8.1 to +22.3; 13 of 15 patients; p = 0.0006) and T-wave amplitude by 16.8 µV (95% CI +4.1 to +29.6; 12 of 15; p = 0.018). Both tests passed under Holm.
The margins are smaller than those above, and the advantage is not uniform. It comes almost entirely from electrode-motion artifact and from the noisiest input level. Under baseline wander and muscle noise, the fixed filter matches SNR-X on ST error and does better on T-wave amplitude. On clean input, neither improves on the untouched signal.
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 from the Noise Stress Test Database, leave-one-patient-out. The best fixed filter in the same bank scores +0.059 dB on its own. That filter is picked with hindsight across all patients, which favours it; this is disclosed rather than corrected.
Against six tuned external baselines outside the shipped bank, tuned per fold with access to the clean reference: +1.447 dB, p = 0.0022, 14 of 15 patients.
Frozen, not held outside. The +1.456 dB figure predates any registration. The comparison against external baselines ran as branch B-02, before outside review of registrations began at branch B-03. In both cases the protocol was fixed in code before the run and the build is frozen and re-runnable, but no outside party held the analysis plan in advance. That is a weaker guarantee than an outside-held registration, and it is labelled that way deliberately.
Qualified, September 2026. The +1.456 dB comparison is scored against the reference as recorded, baseline offset included. Scored with that offset removed, a plain fixed baseline-removal filter does better overall than the frozen build; see the scoring check below. Whether the same applies to the external-baseline comparison has not yet been checked.
Scored the benchmark’s original way, the frozen build loses 0.786 dB on clean ECG (+12 dB input SNR), worse than doing nothing. Scored with each window’s baseline offset removed from both the reference and the output, the same choices gain +0.447 dB (95% CI +0.079 to +0.789), and 14 of 15 patients are within 0.25 dB of doing nothing or better. Both pre-registered tests passed: at least 97% of the loss is explained by the offset at the lower 95% bound, and one-sided exact Wilcoxon p = 0.0017.
Pre-registered. Branch B-07 was hashed and held by the outside reviewer before it ran, on 16 September 2026. The published −0.786 dB figure is unchanged; this explains it.
The same check turned up something less welcome, outside the registration. With the offset removed, a plain fixed baseline-removal filter scores +3.76 dB overall against +2.84 dB for the frozen build, and does better in 13 of 15 patients; the frozen build still does better on clean ECG. This was not pre-registered, so it is a lead, not a result. It does mean the +1.456 dB figure above rests partly on a scoring convention that handicaps that filter. A later registered branch (18 September 2026) retrained the selector under the corrected scoring as part of a wider test. Retrained, it scored above that filter by 0.55 dB (95% CI +0.25 to +0.81). That comparison was reported as an effect size inside the branch, not as one of its registered tests, so it narrows the caution without closing it. The EEG figure is not affected: the clean EEG recordings carry almost no offset.
A published ECG denoising network (DeepFilter), trained per fold as published, was added to the selector’s options. On ECG the selector did better than the network alone by 0.80 dB overall (95% CI +0.46 to +1.22; 14 of 15 patients; p = 0.00009) and by 1.15 dB on clean input at +12 dB (95% CI +0.42 to +2.05; 12 of 15; p = 0.004). An improved version of the selector, registered at the same time, stayed no worse than doing nothing on clean and lightly noisy ECG within a 0.25 dB margin (p = 0.005 and p = 0.00003), and cut the share of those inputs it made more than 1 dB worse from 19% to 7.5% (13 of 15; p = 0.0002). ST error fell from 136 µV with the network alone to 112 µV with the selector, reported as an effect size rather than a registered test.
Pre-registered. Branches B-09 and B-10 were hashed and held by the outside reviewer before any network was trained on benchmark data, on 18 September 2026. Both pilots had predicted that the ECG tests would fail. No published figure changes as a result.
Bounds worth keeping attached. Most of the advantage over the network comes from baseline wander; under muscle noise and mixed noise it is not distinguishable from zero. The network was trained on the second channel of the noise recordings, never the channel used in testing, while the selector trains on other patients’ cases built from the test channel, a difference stated in advance as favouring the selector. The comparison uses the corrected scoring above; scored the benchmark’s original way, the selector does worse than the network. It is one simple published network; the repeat over a second network is described below. The 112 µV ST error is still above the ±100 µV magnitude clinicians care about. On EEG both branches passed formally, but the network there, trained as registered, did worse than doing nothing, so beating it shows little. A registered repeat on 19 September 2026 trained that network to convergence, and it still did worse than doing nothing (−0.76 dB overall). The EEG question therefore stays open and no EEG result is claimed.
Repeated over a second published network, 21 September 2026. Branch B-12, held by the outside reviewer before any network was trained, repeated both ECG tests with the strongest model in the same published family (DeepFilter’s Multibranch LANLD). Both passed again: the selector did better than that network alone by 0.66 dB overall (13 of 15 patients; p = 0.0004) and by 1.26 dB on clean input at +12 dB (14 of 15; p = 0.0002), and the improved selector cut the share of clean and lightly noisy inputs it made more than 1 dB worse from 19.5% to 6.7%. How to read it was fixed in advance: the second network did not prove measurably stronger than the first (p = 0.0535), so this is a replication over a network of similar strength, not evidence over a stronger one. The same bounds apply: corrected scoring only, advantage mainly from baseline wander, and on clean ECG leaving the signal alone still gives the smallest ST error.
15 of 15 held-out blocks, p = 1.2 × 10⁻¹⁴, capturing 85% of the oracle headroom. EEGdenoiseNet, 15 contiguous blocks in recording order. The comparison is against the best fixed filter in the bank, picked with hindsight across all blocks, not against ICA or a published EEG method. The noisy recordings are synthetic mixtures of clean EEG and recorded artifact, as the dataset provides them, and the dataset carries no subject identifiers, so held-out blocks cannot be confirmed to come from different people.
The two domains differ in sampling rate, morphology and dominant artifact sources, and the candidate bank and the policy were retrained from scratch for each. The same architecture was positive in both, which is evidence about the mechanism rather than about any one signal type.
The obvious objection, stated here rather than left for a reader to raise: both are surface biopotentials, recorded through similar electrodes at similar scales, so this is not transfer across physics. What it supports is narrower — that the selection mechanism is not tied to one signal’s morphology or artifact structure. A third signal, wrist heart rate from an optical sensor, has since passed one registered test; that result is below, and it is narrower than it may sound.
Pre-registered. This is the figure as corrected by a pre-registered branch on 11 September 2026, which fixed the block split. The registration went to the outside reviewer before that run. The earlier figure stays in the record as what was obtained under the original split.
PPG-DaLiA: optical pulse (PPG) recorded at the wrist during everyday activities, with heart rate from a chest ECG as the truth. A separate bank of treatments for this signal and a selector trained once on five people were applied, unchanged, to ten people held back. Mean heart-rate error: raw signal 15.70 bpm, best single treatment 13.56, SNR-X 10.29, perfect hindsight choice 6.90. SNR-X was lower than the best single treatment by 3.25 bpm (95% CI 1.70 to 4.81; 9 of 10 people; one-sided exact Wilcoxon p = 0.0029), capturing 61% of the hindsight headroom.
Pre-registered. Branch B-15 was hashed and held by the outside reviewer before it ran, on 22 September 2026. It was built beside the frozen build, which it does not touch, so it moves none of the ECG or EEG figures above. Two of its three registered predictions were wrong, both in the favourable direction.
Bounds worth keeping attached. The heart-rate method SNR-X sat in front of was deliberately simple and does not carry its estimate from one window to the next, so part of the room exists because that method is weak. In front of a strong published method, two pilots found no worthwhile gain. One of the ten people did worse than with the best single treatment, and two did worse than with the raw signal. SNR-X still does harm on clean signals: sitting still, its error was 4.90 bpm against 2.34 for the raw signal. The sensor is optical rather than electrical, but this is still a signal from the body, so it narrows the question of transfer across physics without settling it. The dataset’s own download page limits it to non-commercial scientific use.
Re-checked on 14 September 2026. The B-06 analysis reproduces the published clinical figures from the shipped row data, and the measurement script, re-run unmodified against MIT-BIH, regenerates those rows bit-exactly on a three-record spot check. The EEG grid has since been rebuilt from the raw EEGdenoiseNet corpus, on 18 September 2026, and matched its checked-in grid exactly.
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 R-peak detectors already exceed 99% F1 on this data, so there was little headroom for any method to capture. The null stands as reported. It is the reason the clinical measurement study above was run at all, and the reason the commercial question is still open: improving a signal is not the same as improving what someone does with it.
Two attempts at a label-free confidence signal, each written down before it ran, returned −0.026 dB (p = 0.28) and −0.033 dB (p = 0.40). Both predate outside review of registrations, which began at branch B-03. Six further attempts preceded them: eight failures, none of which produced a usable score.
Branch B-13, held by the outside reviewer before it ran on 21 September 2026, gave the selector one more option: replace each heartbeat with the typical shape of the beats around it. The selector did not improve overall (−0.08 dB; 7 of 15 patients; p = 0.87) and did worse on clean ECG (−0.30 dB; 3 of 15). A small reduction in ST error (113 to 110 µV; p = 0.013) passed its own registered test, but the reading fixed in advance says the option is not added to any claim when the primary test fails, so it is not. The pilot had predicted a pass.
The first new domain tried was HF radio: decoding FT8 amateur-radio transmissions from real recordings, where the decoder’s own error check says whether each message came through, so no clean reference is needed. A pilot on 21 September 2026, 237 recorded cycles from three receivers, found that on these recordings none of the candidate treatments improved on the recording as received, and that most of the extra messages gained by decoding many versions of each recording came from simply trying again rather than from the treatments. The branch was stopped before registration and no study data were collected.
This was a pilot, not a registered test: three receivers, about 20 minutes each, one afternoon, and none of the recordings carried the kinds of interference the treatments were built for. It says nothing about the ECG or EEG results.
Bearing vibration (the Paderborn dataset) was checked the next day, on 22 September 2026. Existing methods leave room there, but a pilot found that the room lies in the fault classifier and its training data rather than in interference a treatment could remove. It was stopped before registration. The lesson now governs which domains are tried: room left by existing methods is not enough on its own.
Two more domains were screened at the desk in September 2026 and set aside before any pilot: outdoor air-quality sensors, where the room lay in refitting the correction equation rather than in interference, and motion-sensor orientation under magnetic disturbance, where a freely available newer method already recovers most of what the common one loses. Neither reached a pilot, so neither is reported as a result.
The wrist heart-rate result was obtained in front of a deliberately simple method, so the obvious next question was whether anything remains in front of a strong one. On a second, newer wrist dataset (GalaxyPPG), a published method (BeliefPPG, ETH Zurich) was fixed as the host after two weaker candidates failed a strength check on the same data. A pilot on five people found that even a perfect hindsight choice left only a small margin, below the line set in advance, and almost all of it in running, jogging, walking and typing. A second pilot, proposed after the first and declared as such, let SNR-X act only on the minority of windows flagged as difficult, without using labels to flag them. It left the easy windows unharmed but gained almost nothing on the difficult ones. Both were stopped before registration on 24 September 2026. The remaining nineteen people in that dataset have not been opened and are kept for any future registered test.
These were pilots on five people, not registered tests, and no figure from them is claimed as a result. What they do settle is how the wrist result above is read: as a result in front of a simple method only.
Selector behaviour does differ by noise regime (p = 0.0001). The stronger claim — that the advantage in one stratum is demonstrably larger than in another — was not supported at n = 15 and has been withdrawn, in writing, to the reviewer holding the registration.
The failures above are not where this stopped. Each of them has a line of work, though no registered branch is running at this moment.
Each of these runs as a registered branch under the same rules as everything further up this page: the prediction is fixed and held outside the company first, the result is reported second, pass or fail. That is why the failures above exist in public at all — they are what the process produces when it is followed.
At runtime there is no clean reference, so the system cannot tell whether the treatment it chose helped. Offline supervision sidesteps this for the base capability, and the frozen build needs no runtime score to work. But abstention, and any supervisory layer above it, do need one.
Every route I am aware of comes down to finding a second, partly independent view of the same underlying truth. Which routes those are, and which are live, is not set out here. Where no such second view exists, abstention has no basis I currently know of.
These are selection criteria, not a claim of results beyond ECG, EEG and the one wrist heart-rate result above. One structural limit is worth stating plainly: where inputs are already clean, candidate outputs converge and there is little for any selector to capture. The architecture needs variable-quality inputs with real headroom.
Axiometrics, Inc. is a Massachusetts C-corporation working on the Overseer and SNR-X. Enquiries from clinical, industrial and research groups are welcome, as are conversations with engineers interested in the runtime scoring problem above.