Build the controls
White noise, pink 1/f noise, correlated noise and a daily rhythm: ordinary backgrounds that can already fool a detector.
THE LABORATORY / LOOKING IS NOT ENOUGH
Can you see
what has changed?
A perturbation may be hiding in a signal. Or your intuition may be playing tricks. Answer first, then discover what the detectors saw.
01 / THE BLIND CHALLENGE
Your eyes. Then the algorithms.
Where does the added perturbation begin? Click the plot, or use the slider with your keyboard. You can also answer that no perturbation was injected. One answer reveals the result.
The challenge begins when you are ready.
| Detector | Perturbation found | Delay | Alerts outside perturbation |
|---|
02 / THE MACHINE TAKES THE TEST TOO
What Trinity finds.
What it misses.
Generated signals, hidden changes and validation seeds kept separate from training. The current 1/fβ slope is compared with three simple methods on the same trials.
Connecting to the laboratory…
The joint false-alarm rate will be measured on the validation controls.
| Detector | Found / trials | Median delay | Min. / max. delay |
|---|
Delay includes successful detections only. Missed trials remain in the denominator. A small number of seeds gives an exploratory estimate.
False starts in signals without a break
| Background | Detector | Alerts | Eligible hours | Scaled to 30 days |
|---|
THIRD-PARTY ARCHIVES / ANOTHER TEST
Methods meet other signals.
A fixed set of annotated Numenta Anomaly Benchmark (NAB) archives, separate from Trinity's measurements. The β64 prototype uses 64 consecutive samples: it is distinct from the production detector and the synthetic trials above.
Loading the archive comparison…
| Method | Detected / annotated windows | False-alarm episodes | False-alarm / normal points | Point precision | Point recall | Median / p90 delay | Scored / test points |
|---|
Episodes and points are shown together so that alert duration remains visible. Delays count samples from annotated window boundaries, not physical anomaly onsets. Missed detections remain in denominators; no point adjustment is applied.
Sources, future partition and protocol
The training prefix fixes thresholds before testing. Original order is retained; some clocks are irregular or nonmonotone. β64 therefore uses sample indices, without a physical frequency band in hertz.
This exploratory subset is neither the official NAB score nor a universal method ranking. Data comes from a third-party repository under the MIT license.
03 / AN EXPERIMENT YOU CAN REPEAT
The rules are fixed before the test.
Hide a change
Level, variance, periodicity or spectral texture, at several amplitudes and durations. Detectors do not receive the injection time.
Keep the trials separate
Training controls set comparator thresholds. Other seeds then evaluate sensitivity, delay and false alarms.
Keep the misses
Code and parameter fingerprints precede validation. Checkpoints allow resumption. No threshold is retuned to improve the resulting score.
The laboratory measures tools on these synthetic scenarios. Observatory investigations keep their real data, dates and registry separate.