CQ-GESUS · Transparency Framework
We hold a CS2 prediction system to the same standard as academic research.
We publish what we know, what we don't, and where we might be wrong. Every statistic comes from a demo we parsed ourselves. Every prediction is frozen the moment it's made and attributed to the exact model that made it. Every paper bet is settled in public against the real result — wins and losses alike.
119,957
Matches in DB
48,468
Demo-driven matches
165
Frozen predictions
927u
Shadow bankroll
Core Transparency Assets
📄
Technical Whitepaper v1.0
Full methodology: architecture, parser-first data foundation, CQE/CQR/Elo ratings, the GESUS model family, validation, and risk. Quant-level depth, no marketing.
Read the whitepaper →
🔮
Live Predictions
Every upcoming and settled match prediction — frozen probability, the model & version that made it, and the top-3 feature drivers behind the call.
View live predictions →
📊
Live Shadow-Bet Ledger
The public scorecard: bankroll, ROI, hit-rate, open exposure and expected value, plus every paper bet broken down by tier and by price source.
View the ledger →
⚠️
Why CQ-GESUS Might Fail
Eight documented failure modes — roster churn, meta shifts, Tier-3 noise, market-beats-model, parse gaps, overfitting, format variance, small sample — with honest mitigations.
Read the failure analysis →
🗄
Data Governance
The parser-first rule: only validated, v4-parsed demo data drives any metric. Scraped numbers are isolated and never feed a rating or a model. Live gold-table coverage.
See the data foundation →
🏆
CQE / CQR Ratings
The player and team ratings built on the same parsed data — opponent-adjusted, Bayesian-shrunk, eligibility-gated. The benchmark the models also consume.
Browse the rankings →
What We Publish and What We Don't — The Tiered Model
We follow a tiered disclosure model: enough to let anyone independently evaluate the quality and honesty of the methodology, while protecting the implementation details that constitute the core IP. Reproducing this system from what we publish would still require building the demo-parsing pipeline and re-deriving the engineering from scratch — which is the point.
| Tier | What we share | Why this level |
|---|---|---|
| Public | System architecture, data sources, gold-table schema and live coverage, parser-first governance rules, feature-category counts, per-tier AUC ranges, the honest ~0.70 ceiling, the full live prediction & shadow-bet ledger, model names/versions, and the immutability invariants. | Sufficient to judge methodology quality and statistical validity independently. |
| Gray box | Model family (gradient-boosted ensembles + calibration), rating design principles (opponent adjustment, Bayesian shrinkage, point-in-time Elo), feature categories, the leakage-free training discipline, and the per-match model-selection logic. | Gives genuine technical depth for evaluation without handing over a build recipe. |
| Vault | Exact feature list and transformations, model hyperparameters and weights, the CQE/CQR/Elo formulas and constants (K-factors, shrinkage priors, component weightings), ensemble blend weights, and the parsing-pipeline internals. | Core IP — the years of pipeline and modelling work that can't be cloned from a page. |
Our commitment to radical transparency
- Every prediction frozen & timestamped at creation — never silently changed
- Every statistic traceable to a parsed demo and a known parser version
- Per-tier results published separately — no hiding behind a blended average
- The live shadow-bet ledger is public: wins, losses, ROI and drawdown alike
- The honest AUC ceiling is stated plainly — we don't claim to beat variance
- No paid product until a statistically significant live track record earns it
- Failure modes documented before they happen, updated when they do
- Methodology described at the family level so it's evaluable, not a recipe