Skip to content
Ryedore Logo
Theme
ContactRequest Demo
Industrial equipment and instrumentation

Verify it yourself

Public data. Published results. Signed — so you can re‑run them.

We test our AI on datasets anyone can download, and publish the results — including where we lose.

Every result is cryptographically signed. You don’t need to trust us; you can check.

See the scoreboardRe‑run it in 10 minutes
17.85 RMSE
Remaining life · FD002
signed · re-runnable
0.992 AUC
Anomaly with real labels
signed · re-runnable
0.303 RMSE · 91% coverage
Forecast · calibrated
signed · re-runnable
blocked 0.90 → 0.75
Never‑degrade gate
signed · re-runnable
the scoreboard

Wins and losses, on named public datasets

“Win” means we beat the named published or standard method on that dataset. Where we don’t, we say so.

Remaining useful life — multi‑condition
NASA C‑MAPSS FD002 / FD004 · vs Li 2018 · Chen 2020

Beats both published methods on both subsets. Academic frontier (NAS 2023: 15.96 / 20.0) disclosed.

17.85 RMSE
signed · re-runnable
20.54 RMSE
signed · re-runnable
WIN
Remaining useful life — single condition
NASA C‑MAPSS FD001 · vs DCNN Li 2018 (12.6) · XGBoost

Beats classical baselines; 0.79 off the best published number. Listed as competitive, not a win.

13.39 RMSE
signed · re-runnable
COMPETITIVE
Anomaly — with real failure labels
NASA C‑MAPSS FD003 / FD004 · vs IsolationForest · OC‑SVM · LOF · AE

Decisive with real labels. With pseudo‑labels only, IsolationForest edges us — labels are the lever.

0.992 AUC
signed · re-runnable
0.989 AUC
signed · re-runnable
WIN
Condition classification
UCI Hydraulic (4 targets) · vs RandomForest · XGBoost · LightGBM · CatBoost

We serve the best model family per target; one target (valve) is a data ceiling shared by every method.

0.959 macro-F1
signed · re-runnable
WIN
Sensor forecasting — multi‑condition
NASA C‑MAPSS FD002 / FD004 · vs PatchTST · N‑BEATS · DLinear (3‑seed)

A single deployable model beats the ensembles; its 90 % interval covers 91 %. Zero‑shot vs general foundation models: behind — disclosed.

0.303 RMSE · 91% coverage
signed · re-runnable
WIN

Baselines are published academic methods and standard open‑source models (reproducible). We do not benchmark or name closed commercial products.

DatasetTaskTest setSize stated by
NASA C-MAPSS FD002Remaining life (multi-condition)259 test enginesdataset documentation
NASA C-MAPSS FD004Remaining life (multi-condition, multi-fault)248 test enginesdataset documentation
NASA C-MAPSS FD003Anomaly with real labels100 test enginesdataset documentation
UCI HydraulicCondition classification552 test cycles × 4 targetssigned evidence (test_n)
C-MAPSS FD002 sensorsCalibrated forecastingmulti-condition test splitsigned evidence

Every headline number carries its denominator — a score without a test-set size is not a result. Bundle signed 2026-09-01; signature re-verified 2026-09-01 (21/21 reports, key 2bc036d1).

seen, not read

Our number next to the published ones

The same results as the scoreboard, drawn — green is ours (registry, signed), amber the best published research figure, grey the standard or published baseline. Where the amber bar is shorter, we are still behind, and we say so.

Remaining life · variable conditions (FD002)
win
NASA C‑MAPSS FD002 · RMSE ↓ · signed, re‑runnable
Ryedore17.85Chen 202021.24Li 201822.36best published (NAS)15.96
Shorter bar is better. Green = Ryedore (registry, signed); amber = best published research result; grey = standard or published baseline.
Remaining life · variable conditions (FD004)
win
NASA C‑MAPSS FD004 · RMSE ↓ · signed, re‑runnable
Ryedore20.54Chen 202023.53Li 201823.31best published (NAS)20
Shorter bar is better. Green = Ryedore (registry, signed); amber = best published research result; grey = standard or published baseline.
Remaining life · single condition (FD001)
competitive · disclosed
NASA C‑MAPSS FD001 · RMSE ↓ · signed, re‑runnable
Ryedore13.39best published (DCNN)12.6
Shorter bar is better. Green = Ryedore (registry, signed); amber = best published research result; grey = standard or published baseline.
Anomaly detection · real failure labels (FD003)
win
NASA C‑MAPSS FD003 · AUC ↑ · signed, re‑runnable
Ryedore0.992standard method (IsolationForest)0.961
Longer bar is better. Green = Ryedore (registry, signed); amber = best published research result; grey = standard or published baseline.

Comparison figures are quoted from the named publications for disclosure; the report lists every source. Open the report →

how to verify

Three levels — no trust required

File integrity with a hash. Authenticity with the RSA signature and the public key in the pack. Reproduction on the public data — or on your own data, yourself, with the open harness.

1public dataset2our model3result4signed report5you re‑run it2bc036d1sha256sum -c SHA256SUMS → RSA‑PSS verify → reproduce on NASA C‑MAPSS / UCI Hydraulic · wins AND losses published
The buyer is inside the loop: the same public data and method give the same number — or the signature breaks.
1 · Integrity

Download the evidence pack, run sha256sum -c SHA256SUMS — every file matches its recorded hash.

2 · Authenticity

Bundle and every per‑task report are RSA‑PSS/SHA‑256 signed under the public key in the pack (fingerprint 2bc036d1). verify.sh is included.

3 · Reproduce

The datasets are public. The benchmark harness is Apache‑2.0, isolated and read‑only. Ask for it, or run it on your own data yourself — see Run the audit.

swipe → to see the whole diagram

Signed evidence and reports
evidence pack · 2026‑09‑01

Download the proof, not a promise

signed · re-runnableWhat we don’t disclose: weights, architecture internals, customer data.
demonstrated, not claimed

It can’t quietly get worse — here is the log

A candidate that would have dropped category accuracy from 0.90 to 0.75 was blocked. A candidate that improved on every task was admitted. The mechanism is drawn on How it works.

blocked 0.90 → 0.75
From the demonstration
signed · re-runnable

A model version that would have dropped category accuracy from 0.90 to 0.75 was blocked from serving; a version that improved on every task was admitted. A promoted model can never be worse on any task.

Evidence log · no_degradation_demo
candidatechange on a taskdecision
Acategory accuracy 0.90 → 0.75BLOCKED
Bremaining‑life error 14 → 16.5BLOCKED
Cimproved on every taskADMITTED
GATE 1
No regression
GATE 2
Serving budget
GATE 3
Calibration
GATE 4
Distribution shift
GATE 5
Fairness
GATE 6
Data quality

Named quality gates every version must pass — not an unexplained count of “checkpoints”.

the honest frontier

Where we are still behind — and we say so with the numbers

Three places the best published research is still ahead of us, drawn to scale. Shorter bar is better; amber is the frontier; green is ours. And one more, without a chart: in zero‑shot forecasting, general foundation models lead when nothing is trained in‑domain — in‑domain, ours wins and is calibrated.

Single‑condition remaining life
behind · disclosed
NASA C‑MAPSS FD001 · RMSE ↓
Ryedore13.39best published (DCNN)12.6
0.79 behind. We beat the classical baselines; the published deep method still leads.
Variable‑condition remaining life
behind · disclosed
NASA C‑MAPSS FD002 · RMSE ↓
Ryedore17.85best published (NAS)15.96
We beat the widely cited published methods; the remaining gap to the academic NAS frontier is disclosed here rather than promised away.
Variable‑condition remaining life
behind · disclosed
NASA C‑MAPSS FD004 · RMSE ↓
Ryedore20.54best published (NAS)20
0.54 behind the published NAS frontier — disclosed, not footnoted.
stated, not discovered later

Where it does not work yet

The conditions under which we lose, decline to compete, or refuse to act — from the same signed record as the wins.

Too little labelled history

When a task has only a few hundred labelled cycles, model choice stops mattering — our signed classification results show the ceiling is set by data volume, not by the model. We say so instead of promising accuracy the data cannot support.

General-purpose forecasting

On data far from industrial equipment, large generic time-series models beat ours — disclosed in the signed report. Our edge is in-domain: the machines and sensors the model was built for.

Transfer to unseen domains

Zero-shot transfer to a sector with no overlapping data is not something we claim — our signed domain-adaptation runs did not produce wins, so cross-domain claims stay off this site.

Life-critical sectors act manually

Aerospace, defence and nuclear deployments run with a manual floor: the AI analyses and proposes, and every action requires a human — permanently, not just at first.

Outcomes are scenarios until attributed

Customer outcome figures on this site are worked scenarios and labelled so. They become attributed results only when customers put their name to them.

A version can be blocked for data quality

A candidate whose improvement rests on too little or too lopsided data is refused promotion even when its score looks better — the gate log records the refusal.

The losses behind these statements are in the signed bundle above — and drawn to scale here.

questions sceptics ask

Four answers

What does “signed” actually mean here?
Every report and the whole bundle carry an RSA‑PSS/SHA‑256 signature under a public key shipped with them. Change one number and the signature breaks — so what you download is what we measured.
Why publish losses?
Because a scoreboard with only wins is marketing. We show where the published research frontier is still ahead of us, drawn to scale, so you can weigh the wins.
Are the baselines fair?
They are published academic methods and standard open‑source models, run through the same open harness, multi‑seed, on public datasets. We do not benchmark or name closed commercial products.
Can we run this on our own data?
Yes — the same harness runs on your machines against your dataset; nothing is sent to Ryedore. The audit page walks through it in four steps.
where next

From proof to the rest of the site

what the round builds here

Where verification goes next

Two programs that turn this page’s discipline into a product; each stays labelled roadmap until its own entry appears in the log below. roadmap · not shipped

program 3

Certified serving — a model version ships with a certificate, not a changelog

today · measured
Named promotion gates wired and enforced on the path (no‑regression on every task, serving‑latency budget, calibration, out‑of‑distribution, fairness, data quality); a staged swap with instant rollback; a signed audit chain; the blocked / admitted log above.
next
Out‑of‑distribution and fairness checks computed for every prediction task; conformal risk‑control guarantees per operating regime; a signed promotion attestation exportable to regulators.
proof point
An attestation per promotion in this public log; regulator‑ready evidence for the compliance packs.
program 10

Verification as a product

today · measured
Twenty‑one signed reports across tasks and headline results, a re‑run script, the bundle and public key on this page.
next
A third‑party‑reproducible certification service for industrial AI models — ours first, later others’ — with regulators and insurers as users.
proof point
External parties re‑running the bundle; the first certified third‑party model.

Funded by the growth round — what the round builds, all ten programs →

Run it on your data — yourself.

The signed bundle and the open harness are yours to run; nobody from Ryedore touches your data.

Run the audit yourselfRequest a demo