/ Free tool

Before you train a model, find out if it is even learnable

Stream a sample of your events - player actions, transactions, sessions - and we measure, on that sample alone, whether a model could learn anything worth having. Nothing is trained and nothing is kept. You get a verdict and a signed receipt over the exact sample we read.

This runs on the sample you paste. It is measured for a verdict and a hash, not stored.

/ What this checks

What Signal Scan measures in your sample, and how to read the verdict.

Your browser parses the pasted JSON and sends the array to our server, which accepts up to 20,000 events per scan. The code keeps the rows that are objects, then picks out numeric fields (numeric wherever they appear and present on at least a tenth of the rows) and an actor column under a common name such as user_id, player, account or session. It counts events per actor and flags counts outside Tukey fences at 1.5 times the interquartile range. If you name a label field with exactly two values, it also computes Cohen's d between the two classes for every numeric field.

With a two-value label, the verdict is Needs more data when either class has fewer than 20 examples, Promising when both classes are large enough but no numeric field reaches an effect of 0.5, and Learnable when at least one does. Those fields are listed with their d value. Without a usable label, fewer than 200 events gives Needs more data, any actor flagged on volume gives Learnable, and fields that merely vary give Promising. The positive class is whichever label value sorts last, so check that the positive rate describes the group you mean.

No model is fitted, so there is no accuracy figure, and the result describes only the rows you sent. Each numeric field is tested alone: a pattern that shows only when fields are combined will not appear, and text or category fields are ignored. The server hashes the sample with SHA-256 and does not write the events to its database. If a signed receipt is issued, it records the event count, the checks that ran, the verdict, the hash and the sample name. An email address, if entered, is saved with the verdict. After Needs more data, send a larger or labelled sample.

How much data do you need to train a machine learning model?
There is no fixed number. It depends on how many classes you want to tell apart, how strongly the inputs relate to the outcome and how noisy the records are. What matters most is the count of examples in the rarest class: a fraud dataset with millions of rows and a dozen confirmed cases gives a model almost nothing to learn from. This scan treats 20 examples per class as the floor.
What is Cohen's d and what counts as a large effect?
Cohen's d is the difference between the means of two groups divided by their pooled standard deviation, so it expresses the distance in standard deviations and does not depend on units. By convention 0.2 is read as small, 0.5 as medium and 0.8 as large. It describes how far apart two groups sit on one variable; it is not a measure of how accurate a classifier would be.
Can you detect anomalies without labelled data?
Yes. Unsupervised methods look for records that differ from the bulk, for example values beyond interquartile-range fences, points that are isolated quickly by random splits, or items far from any cluster. They tell you what is unusual, not what is harmful. Someone still has to review the flagged cases, and the confirmed ones then serve as labels for a supervised model.
Next

When one check is not the whole job.

Worth building on? The paid Custom Model Audit takes the whole dataset and scopes what to build.