Integrity is the product.
There is no house RNG — results come off a real server. The only thing that matters is the integrity of the matchup. Every duel runs a forensic gauntlet: demos parsed, behavior scored, risk fused into a posterior, and a verdict issued — auto-cleared, held, or routed to a human. The hard engineering is the moat.
Hover a signal to preview its effect on the posterior · illustrative forensics
Every match runs the gauntlet
Event-driven, async, post-match. Each stage is its own bounded context, wired only through the bus — no shared source. A demo enters; a verdict leaves.
Demo parse
Fetch the server-recorded .dem, parse to structured events + columnar features.
Feature extraction
Aim kinematics, flick distributions, pre-fire timing, recoil control, snap geometry.
Cheat scoring
Statistical + ML likelihood per vector: aimbot, triggerbot, wallhack.
Risk fusion
Fuse scores with account history + the account-link graph into a posterior.
Enforcement verdict
Auto-clear, hold winnings + restrict, or route to the human review queue.
Posterior below the review band. Payout released instantly.
High posterior. Winnings held, account restricted pending verdict.
Routed to the C13 console with a full evidence package.
A wrongful ban is catastrophic — so we never auto-ban on one match
Auto-ban fires only on a posterior ≥ 0.92 sustained across ≥ 2 matches. Everything in the band below routes to a human with the full evidence package. The model is tuned to a false-positive rate, not raw accuracy.
every duel, every server
of flags are true cheats
wrongful flags per clean match
the rest go to humans
The auto-ban gate is posterior ≥ 0.92 AND ≥ 2 matches. A single high-scoring match never bans — it can only hold winnings and open a review. The cost of a wrongful ban is treated as far higher than a missed catch.
Auto-ban only fires when the signal holds across independent matches. One game can single-handedly hold & review — never ban.
The mass sits near 0 — clean play. The review and ban bands are a thin tail.
Four attackers, one load-bearing defense
Each adversary needs a distinct detector — but trust tiers gated by stake is the unifying control that blunts all four at once, and doubles as the AML mechanism.
Smurfs / alts
Manufacturing a low rating to farm money off weaker opponents.
RD-gated rating + device fingerprint + tiered stake ceilings.
Boosters / ringers
A stronger player logging into someone else’s account.
Behavioral baselining + Steam-tenure signals + trust tiers.
Collusion / chip-dumping
Linked accounts arranging to move money — the laundering vector.
Account-link graph + directional-flow detection + dual-control clawbacks.
Cheaters
Aimbot / triggerbot / wallhack running on the server.
Server-side demo forensics + ML scoring + held-payout on high posterior.
Trust tiers gated by stake
A higher stake requires a higher tier — which requires more identity verification, longer tenure, and a cleaner demo-forensic record. One mechanism raises the cost of every attack at once, and satisfies AML obligations. Build it early; it is load-bearing.
Stake ceilings illustrative · money is a per-region feature flag