Minecraft anticheat guide

Checking FairPlay AIM Model Retrains Before Going Live

What to verify after a FairPlay AIM model retrain: backups, staging tests, false-positive review, and the publication rules in the engineering notes.

Updated 2026-11-07 · ~5 min read

fairplay aim modelanticheat ml retrainminecraft aim detection false positivesfairplay model backupanticheat model staging test

Short answer: what a retrain can change

Per FairPlay engineering notes, AIM model retraining changed how legitimate examples are used, how candidates are calibrated, and when a model is published. A retrained model can be published even when it misses the false-positive target. Back up the current model and test the candidate on staging before it goes live.

What changed in training and calibration

The notes list several changes. Each one affects which model replaces the current one, so read them as a list of decisions rather than accuracy promises.

The notes stress that these are method changes. Any statement about accuracy on a given server depends on your traffic, your settings and your staff review, and none of those were measured in the notes.

Hard legitimate examples from the training set are replayed against the incumbent model, and the highest-scoring ones are added back. Held-out calibration groups are not used.
Legitimate examples above a tail-risk cutoff receive extra weight during optimization.
Candidate selection favors lower full-session legitimate false-positive rates when cheat coverage is equal.
Short-window qualification was added beside the rolling decision used before.
Identical current-schema models can be reused without a second strict requalification.

When a candidate is published

The notes replace the old pass-or-stay rule with a fallback sequence. A strict winner publishes immediately. When none passes, one bounded rescue attempt runs, and then the best calibrated candidate publishes in recovery mode regardless of the false-positive target.

Publication outcomes described in the notes
Situation	Outcome described in the notes
A candidate passes the strict criteria	Publishes immediately
No strict winner, but a calibrated candidate exists	Publishes in recovery mode after one rescue attempt
Optimization produces no replacement and the incumbent works	Incumbent is retained
No calibrated candidate, or the optimizer fails	No new model is produced

Limits the notes call out

The notes are direct about what the changes do not cover. Read these limits before deciding a candidate is safe.

The method is a training change, not a whitelist. A window in the training set is not guaranteed to classify as legitimate after training.
The short-window check assumes default single-window settings. Servers with that behavior disabled or customized may see a model rejected that would have been acceptable for their configuration.
Each AIM model is evaluated alone. A combined multi-model profile is not qualified as an ensemble.
A high false-positive model can be published when no better calibrated candidate exists.

Pre-go-live checklist for server owners

Run these steps on a staging copy of your server. Keep the current model until the candidate has passed your own review. Record a dated entry for every model you run, so a later change in alert volume can be traced to a specific model.

Back up the current AIM model file and note its version and date.
Retrain or load the candidate on staging, using the same configuration as production.
Replay sessions of known legitimate play, including high-ping and high-CPS players, and record every alert.
Have staff review the alerts, then compare the false-positive count with the incumbent.
If your short-window or other detection settings differ from the defaults, validate the candidate under those exact settings.
Restore the backup if alert volume rises on live traffic.

Back up the current AIM model before replacing it (replace the placeholders)

cp <path-to-aim-model-file> <backup-dir>/aim-model-$(date +%F).bak

How to read alerts after a retrain

Compare alert history over a fixed period before and after the change, using the same staff review process. Count alerts by check and by player, and separate the alerts staff confirmed from the ones they dismissed.

If a model published in recovery mode produces a noisy stream, treat that as a reason to restore the backup, not as proof the model is wrong. Recovery mode means no candidate met the strict criteria, so noise is the expected outcome until a better candidate is trained.

What has not been verified

The notes describe training overhead as complexity estimates, not benchmarked wall-clock ratios. Training may take modestly longer because more examples are included.

The smoke tests in the notes ran in an isolated harness without Maven, and no live-server test is recorded for these changes. The notes also say older tournament tests were updated for the new fallback publication contract, so earlier expectations about publication no longer apply. Treat accuracy and alert volume on your own server as pending until your staging review is complete.

FAQ

Does a retrain make FairPlay more accurate?

The notes do not claim an accuracy gain. They describe calibration that prefers lower legitimate false-positive rates at equal cheat coverage, and that still needs staging evidence.

Will my current model be deleted if retraining fails?

The notes say existing deployed models are not deleted when the optimizer fails. Keep your own backup anyway.

Is legitimate-player training data a whitelist?

No. The notes state that it is a training-method change and that a training window is not guaranteed to be classified as legitimate.

What does recovery mode mean for my server?

It means no candidate passed the strict criteria, so the best calibrated candidate was published anyway. Expect more review work on alerts until a later retrain passes.

Why would FairPlay publish a model that misses its false-positive target?

The fallback sequence publishes the best calibrated candidate in recovery mode when no strict winner exists, so a server is not left without a replacement. That is why staging review matters.