Fighting Game · Guides & deep dives

V2 Anti-Cheat Classifier Suite Model Card

The V2 anti-cheat classifier suite prioritizes fair-play review for suspicious competitive activity.

7sections1 minread

On this page

Purpose#

The V2 anti-cheat classifier suite prioritizes fair-play review for suspicious competitive activity. The suite covers smurf, win-trading, coordinated-throw, and geographic-anomaly classifiers.

Packages#

  • Training loop and event-bus dataset contract: @nous/training.
  • Runtime model metadata and Model Card hosting: @nous/safety.
  • AI-system-of-record and conformity audit export: @themis/accountability.
  • V2 bridge: @v2/nous-anti-cheat-classifiers.
  • Labeled event transport: @oshun/event-bus.
  • Appeal route for classifier-assisted sanctions: @v2/themis-dispute-resolution.

EU AI Act Classification#

Risk class: Limited risk.

Disclosure text: V2 uses fair-play classifiers to prioritize manual review of suspicious competitive activity. Classifiers do not issue automatic discipline.

Inputs#

  • Match completion telemetry and replay evidence.
  • Ranked point deltas and repeated opponent graph signals.
  • Player skill-profile changes and cross-account similarity features.
  • Tournament completion and bracket relationship features.
  • Session starts, account geo changes, and region entitlement checks.

Event-Bus Training Data#

V2 publishes labeled examples on @oshun/event-bus topics including v2.match.completed, v2.player.skill-profile.updated, v2.rank.points.changed, v2.tournament.event.completed, v2.replay.evidence.attached, v2.player.session.started, v2.account.geo.changed, and v2.entitlement.region.checked.

@nous/training requires clean and positive labels for every classifier before a training plan can be accepted.

Player Controls#

Ranked integrity signals do not offer opt-out. Players receive a transparency notice where required by region and can appeal classifier-assisted sanctions through @v2/themis-dispute-resolution.

Safety And Evaluation#

  • Rollback isolation: no live classifier RPC is reachable from rollback frames.
  • Human review: classifier output can prioritize review queues but cannot issue automatic discipline.
  • Appeal coverage: all classifier-assisted sanctions keep an appeal route.
  • Fallback: if classifier training or Model Card metadata is unavailable, V2 uses deterministic rules and manual review queues only.