They read how the aim moved
Systems that have described their approach look at the path of the crosshair around each shot, not just kills or accuracy. Published research uses view angle, speed, acceleration and jerk around each kill.
Research
How others use data to watch for cheating
Most detection comes down to a few shared ideas. Here is what has been publicly reported, where it came from, and how Overshoot chose differently.
The short version
Five ideas they share
They judge many moments together
One flick proves little. Valve's VACnet was described as reading batches of shots from several rounds at a time.
They keep people in the loop
Valve's review portal uses human labels to train its detection rather than to ban directly. FACEIT is reported to need several signals before acting.
They do not hurry
FACEIT is reported to delay enforcement so several accounts can be handled together, which also avoids tipping off cheat makers.
Stats are not detection
Sites such as Leetify measure crosshair placement, time to damage and spray accuracy to help players improve. They do not claim to detect cheating.
What each looks at
Who looks at what
Based on press coverage, a community wiki of 2018 talks and a research paper. It is not taken from Valve or FACEIT documentation, so details may differ.
| System | What it looks at | How it is used | Source |
|---|---|---|---|
| Valve VACnet | The shot window (about half a second before and a quarter after each shot), aim angle changes, weapon and distance, judged over batches of shots. | Trained to predict whether a human reviewer would convict. Reported to be weaker on subtle assistance. | vac-ban.com wiki of 2018 talks |
| Valve review portal (Aug 2026) | Short anonymised clips of flagged play with an x-ray view, reviewed by several people. | Labels train the detection model. Reviewers' verdicts do not ban anyone directly. | Tech Times, 11 Aug 2026 |
| FACEIT Human Input Detection (Jul 2026) | In-game input sequences, checked against what a human hand can physically do. | Reported to need several signals before action, to delay enforcement and to exclude accessibility hardware. | Tech Times, 30 Jul 2026 |
| XGuardian (research paper) | View angle, speed, acceleration and angle change in windows around kills from CS2 demos. | A neural network with explanations of which features drove each decision. | arXiv 2601.18068 |
| Open-source CS2 detector | Angle speed, acceleration and jerk in short segments. | A neural network; its authors say controlling false positives is still to do. | GitHub project |
| Leetify | Crosshair placement, time to damage, spray accuracy, counter-strafing, accuracy with an enemy spotted. | Performance stats for players. Not cheat detection. | Leetify glossary |
Researched 30 Sep 2026. Product details change, so treat this as a snapshot.
Our choices
What Overshoot does by design
These are design choices, not claims about anyone else.
Evidence, not verdicts
Overshoot shows the measurements and the ticks. It never says a player cheated.
Hand against crosshair
It reads the raw mouse input CS2 demos record and asks whether the hand explains the crosshair.
Ordinary players as the yardstick
Each number is compared with real Premier play, and the count of unusual kills is shown next to what chance predicts.
Nothing installed
No driver and no scanning of your PC. It reads demo files.
People check the numbers
Reviewer labels, stored without names, are how the measurements get checked.
What we leave out
Why some numbers stay private
Overshoot publishes the ideas, not the numbers. Exact thresholds, reference values and model internals stay private, because cheat makers read public pages too. Anti-cheat teams that describe their methods generally do the same.