Integrity / Methodology
How the integrity check works
integrity-v1. Every stored report records the model version that produced it.What it is, and what it isn't
The integrity check is an anomaly detector. It shows how unusual a player's measured stats are compared with similar players, and how consistently. It is not a cheat detector. We don't have a verified, labelled set of cheaters and non-cheaters to train or test against, so the index can't be a probability: a score of 80 does not mean an 80% chance of anything. Skilled players, smurfs and players on a great streak can all score high.
Facts vs. assessment
- • Facts from Steam: account creation date, VAC and game bans, CS2 playtime, shown exactly as Steam reports them.
- • Facts from matches: stats from FACEIT's API and parsed demos.
- • RoundCtrl's assessment: the anomaly index and everything in the statistical comparison.
- • Missing or private: shown as "Private / Unavailable", never as zero and never as a negative sign.
Who players are compared with
Peers are real players in RoundCtrl's public match data with at least 10 matches from the same source (FACEIT and demo stats aren't mixed). For FACEIT, peers are narrowed to similar FACEIT levels (1–3, 4–6, 7–8, 9, 10); for demos, to similar Premier ratings. If a bracket has fewer than 30 players, all players from that source are used instead, and the report says so. Peer levels are their current level, not their level at the time of each match. Reference values are recalculated every 12 hours.
What's measured
- • Headshot %, needs at least 100 kills
- • K/D per match, needs at least 10 matches
- • Damage per round, needs at least 10 matches
- • Opening duel win %, needs at least 30 opening duels
- • Clutch win %, needs at least 20 clutches
- • Win rate, needs at least 10 matches
- • Utility damage per round, needs at least 10 matches
- • Sudden improvement: K/D and headshot % over the last 10 matches vs. everything before, once a player has 30+ matches. It only counts if the change is statistically strong and the new level is itself unusual. Simply getting better is never flagged.
A player needs at least 10 matches and 100 kills from one source before any index is shown.
How the index is built
- For each metric we compute the player's value and a 95% interval for it (Wilson intervals for percentages, a standard error for averages).
- A metric only counts if the low end of that interval is above the peer 95th percentile. Its strength rises from 0 to 0.5 up to the peer 99th percentile, and to 1 beyond it. Small samples have wide intervals, so they can't produce strong signals.
- Metrics are grouped into areas, and each area takes its strongest metric. Each area has a maximum weight:
- Aim: up to 40
- Fragging: up to 35
- Opening duels: up to 25
- Sudden improvement: up to 25
- Clutches: up to 15
- Results: up to 15
- Utility: up to 10
Areas combine as index = 100 × (1 − Π(1 − weight × strength)). No weight is above 0.4, so one area alone can never go past "Some unusual patterns". Reaching the higher bands takes several independent areas being unusual at once.
- • 0–24: No significant measured anomalies
- • 25–49: Some unusual patterns
- • 50–74: Multiple anomalies warrant review
- • 75–100: Strong statistical anomalies warrant detailed review
These bands are provisional. They haven't been calibrated against verified outcomes and will be revised as data grows.
Never used against anyone
Account age, inventory value, profile privacy, FACEIT presence, nationality and Steam ban history are shown for context and have no effect on the index. Being new, private or having a cheap inventory is not evidence of anything. VIP status doesn't affect anyone's score either.
Not analysed yet
- • Aim and visibility: crosshair movement, target acquisition and line of sight need per-tick view angles and map geometry. RoundCtrl's demo parser doesn't extract these yet, so no score is based on them.
- • Inventory value: needs a licensed item-pricing source. We won't estimate prices without one.
- • Premier peer groups: need more parsed demos. Until then, demo comparisons usually say there isn't enough reference data.
The "reviewable evidence" list (kills through smoke, wallbangs, kills while flashed, no-scopes) comes straight from parsed demos. It's for people to look at, with common legitimate explanations, and isn't part of the index.
Corrections
If a report shows something factually wrong (an account date, a ban, a linked FACEIT account), sign in and use "Report incorrect information" on that report. We check every report and fix verified mistakes.