🔬 Math Verification Framework

Beyond RTP: The Constraint
Checklist That GLI Actually Tests

A compute-based verification framework for Chain Reaction e-pull-tabs. Seven constraints that separate a math model that passes certification from one that survives production.

"AI generates. Humans specify. An independent third party verifies."

MadLab Research · TwinGaming.AI · BrazLotto

The RTP Trap

RTP is the most overused number in iGaming. It tells you the long-run return to the player — and nothing else. 96.5% can be achieved a thousand different ways, and 990 of them are commercially unviable.

"An LLM can generate a slot math model in thirty seconds. The RTP will converge. And the game will still destroy the operator."

— MadLab Math Research, TwinGaming.AI

🎯 The Problem

Traditional math verification asks one question: "Does the RTP converge to target?" This is necessary but catastrophically insufficient.

Two games at identical 85% RTP (MN max) can produce entirely different player experiences — one retains players, the other hemorrhages them. The difference lives in the constraints surrounding the RTP, not in the RTP itself.

🔬 The Solution

A constraint checklist that verifies seven parameters beyond RTP — each targeting a specific failure mode that will either kill the game commercially or fail GLI-14 certification.

Instead of running an arbitrary number of rounds, we define the required precision for each constraint and let the verification engine determine the necessary volume. Convergence, not volume.

⚖️ Why It Matters for MN

Minnesota GCB Rule 7864.0235 mandates max 85% RTP and Prize-Based Closing (PBC). GLI-14 v2.2 requires full math model disclosure including prize distribution, not just aggregate RTP.

Our framework goes beyond minimum requirements — making the math submission so thorough that GLI has nothing left to question.

Core Architecture: Server as Auditor

The compute-based model inverts the traditional client-server relationship. Instead of the server dictating outcomes (Oracle), the server confirms that the computed outcome is correct (Auditor).

1. Seed Generation

Server generates cryptographic seed, commits SHA-256 hash before play

2. Client Computation

Deterministic engine runs full game simulation with seed — same JS on client + server

3. Result Report

Client reports outcome to server with full input log

4. Server Verification

Server re-runs identical simulation — match = valid, mismatch = flagged

5. Independent Audit

Any third party (GLI, regulator) can verify any historical session from seed + code

"The industry measures effort. Verification should measure convergence."

— MadLab Verification Methodology

The 7 Constraints Beyond RTP

Each constraint targets a specific failure mode. Missing any one can kill the game commercially or fail certification — even with perfect RTP.

Business Survival

💰 1. WinCap — Maximum Payout

Maximum payout per single round, expressed as a multiple of the bet. A single uncapped hit can bankrupt the operator's monthly margin.

Industry Range 5,000x — 25,000x
Chain Reaction Target 2,000x
MN Regulatory Cap $599 per deal

⚠️ An AI can generate a paytable where symbol combinations yield 500,000x. Mathematically valid. Commercially lethal.

Math Fingerprint

📊 2. Win Tier Distribution

How total RTP distributes across win magnitude tiers. This is the game's fingerprint — two games at identical RTP with different tier distributions are completely different products.

Base wins (1x–50x) ~76% of RTP
Mid wins (50x–500x) ~16% of RTP
Top wins (500x+) ~2.3% of RTP

Formula: frequency = 1 / (RTP_total × RTP_share / avg_win)

Player Experience

💀 3. Dead Spin Rate

Percentage of rounds returning zero. The heartbeat of player experience. Too high without bonus compensation = player leaves. Too low = math doesn't converge.

High Volatility Reference 83.15%
Comfortable Range (Western) 60% — 85%
Pull-Tab Equivalent % non-winning tickets

Dead spin rate and bonus frequency are coupled — change one without compensating the other and the model breaks.

Coupled Parameter

🎰 4. Bonus Trigger Rate

How frequently the player enters the bonus round. Locked with bonus payout — together they must produce the designed bonus RTP share.

Free Spins (typical) 1 : 80 — 1 : 250
Pick Bonus (typical) 1 : 20 — 1 : 30
Chain Reaction Bonus Per deal structure

An AI treats trigger rate and bonus payout as independent variables. They aren't. Shift one, the other must compensate.

Retention Killer

📉 5. Max Drawdown

Longest sequence of consecutive rounds returning less than 1x bet. Not the average — the tail. This kills retention.

// Expected max losing streak max_drawdown ≈ log(n) / log(1/p) // Gumbel distribution parameters location: μ = ln(n(1−p)) / ln(1/p) scale: β = 1 / ln(1/p) // 99.9% confidence bound max_99.9 = μ + 6.91 × β // Example: p=0.86, n=10M // → ~140 consecutive losing rounds // → At $5 bet = $700 drained

No AI will flag this. The tail doesn't move the mean. But the player who hits it never comes back.

GLI Requirement

🔀 6. Mode RTP Separation

Each play mode or denomination is a separate product with its own RTP and volatility profile. Regulators require transparency per mode.

Chain Reaction Denominations $0.50–$5.00
Each denomination Independent RTP verification
MN GCB Requirement Full disclosure per mode

Collapsing all modes into a single model = certification failure + regulatory action.

Methodology

🎯 7. Convergence-Based Verification

The industry's traditional approach: "we ran 100 million rounds." This is a ritual, not a method. 100M rounds may or may not be sufficient depending on volatility and required precision.

The inversion: Define required precision first (e.g., RTP within ±0.1% of target), then let the verification engine determine the necessary volume. Scale from 1M → 10M → 100M until every constraint converges.

Traditional approach "We ran N rounds" (ritual)
Convergence approach All constraints within ±0.1% (method)
Key principle If ANY constraint hasn't converged, verification isn't complete

Convergence Engine

Define precision targets. Measure convergence. Scale volume until every constraint passes. This is what separates a math submission that GLI accepts on first review from one that bounces.

📐 Convergence Protocol

Step 1: Define Precision Targets

RTP ±0.1% | Dead Spin ±0.5% | Bonus Trigger ±5% | WinCap verified | Max Drawdown at 99.9% CI

Step 2: Initial Run (1M rounds per denomination)

5 denominations × 1M = 5M total rounds. Measure all 7 constraints.

Step 3: Check Convergence

For each constraint: is the confidence interval within the target precision?

Step 4: Scale as Needed

Unconverged constraints → 10x volume → recheck → repeat until all pass

Step 5: Generate Verification Report

Full constraint matrix with convergence proof per denomination → GLI-14 submission artifact

📊 Convergence Status (Simulation Target)

Constraint Status
RTP (±0.1%)
✓ CONVERGED
WinCap
✓ VERIFIED
Dead Spin (±0.5%)
✓ CONVERGED
Bonus Trigger (±5%)
✓ CONVERGED
Win Tier Dist.
SCALING 10M
Max Drawdown
SCALING 10M
Mode Separation
✓ 5 MODES

* This is the target verification structure. Actual values will populate from Chain Reaction math engine simulation runs.

"100 million rounds is a ritual, not a method. Define precision first — let the math determine volume."

— MadLab Convergence Principle

Win Tier Distribution Analysis

Reference distribution from a production high-volatility game (WinCap 2,000x, 10M rounds). This is the format our Chain Reaction math submission should follow.

📊 Reference: Production High-Volatility Slot — 10M Round Snapshot

Win Tier Count Frequency Avg Win RTP % RTP Share
0× (dead spin) 8,314,911 83.15%
>0 – 1× 286,581 1 : 35 0.5× 1.50%
1.5%
1× – 5× 876,267 1 : 11 2.4× 20.81%
20.8%
5× – 10× 299,010 1 : 33 6.7× 20.08%
20.1%
10× – 20× 150,037 1 : 67 13.2× 19.76%
19.8%
20× – 50× 54,906 1 : 182 28.8× 15.79%
15.8%
50× – 100× 12,723 1 : 786 65.3× 8.31%
8.3%
100× – 500× 5,284 1 : 1,893 154.4× 8.16%
8.2%
500× – 1,000× 243 1 : 41,152 723.1× 1.76%
1.8%
1,000×+ 38 1 : 263,158 1,525.6× 0.58%
0.6%
Max observed: 1,653×  |  WinCap: 2,000×  |  Tiers 1×–50×: 76% of RTP  |  Tail 100×+: <11%

🎯 Why This Matters

Verification must check not just the RTP total, but the RTP contribution of every tier — and the resulting frequencies — against the specification.

If AI accidentally loads 10% of total RTP into the 1000x+ tier instead of 0.58%, mega-wins become frequent enough to destabilize operator margin.

If it puts 0.05% there, the top end is a ghost — the marketing claim becomes legally questionable.

🔢 The Closing Formula

Given any two of three values — frequency, average win, and RTP share — the third follows directly:

frequency = 1 / (RTP_total × RTP_share / avg_win) // Example: 1000x+ tier // RTP_total = 96.75%, RTP_share = 0.58% // avg_win = 1525.6× freq = 1 / (0.9675 × 0.0058 / 1525.6) freq = 1 / 0.00000368 ≈ 1 : 271,739

Chain Reaction Application

How the 7-constraint framework applies to our e-pull-tab game under MN GCB regulation.

🎮 Game Parameters

Game Type Electronic Pull-Tab
Denominations $0.50, $1, $2, $3, $5
Max RTP (MN) 85%
Closing Method Prize-Based Closing (PBC)
Regulation MN GCB 7864.0235
Certification GLI-14 v2.2

🔄 Pull-Tab vs Slot Mapping

Pull-tabs have a fundamentally different structure from slots but the constraint framework still applies:

Dead Spin → Non-winning tickets % of deal
WinCap → Max prize per ticket $599 MN limit
Bonus → Chain Reaction feature In-game mechanic
Tier Dist → Prize schedule Pre-determined deal
Max Drawdown → Longest loss streak Within deal

PBC-Specific Constraints

Prize-Based Closing adds a unique layer to verification that doesn't exist in online slots:

  • 🔸 Every deal has a known, finite prize pool — RTP is deterministic per deal, not statistical
  • 🔸 Deal must close when all top prizes are awarded (PBC trigger)
  • 🔸 Remaining tickets after PBC must be refunded at face value
  • 🔸 Player experience of the "last tickets" before PBC closure is the critical UX moment

Our Verification Advantage

What makes our math submission stronger than the minimum:

  • Per-denomination verification — each of 5 price points verified independently
  • Win tier distribution table — full breakdown, not just aggregate RTP
  • Convergence proof — confidence intervals for every constraint, not arbitrary round count
  • Max drawdown analysis — Gumbel distribution bounds at 99.9% CI
  • PBC simulation — verifying deal closure behavior under all scenarios

"Who writes the spec? That's not computation. That's fifteen years of watching games succeed and fail in production."

— MadLab Research on AI vs. game math expertise

GLI-14 v2.2 Submission Checklist

Enhanced submission framework combining standard GLI-14 requirements with the 7-constraint verification methodology. Goal: first-pass approval.

📋 Math Model Submission Package

  • 1️⃣ Game Rules & Mechanics Description
    Complete chain reaction mechanic documentation, symbol definitions, win conditions, deal structure
  • 2️⃣ Prize Schedule per Denomination
    5 separate prize tables ($0.50, $1, $2, $3, $5) with ticket counts, prize values, and expected RTP
  • 3️⃣ Win Tier Distribution Analysis
    Per-denomination tier tables showing count, frequency, avg win, RTP share — the game's fingerprint
  • 4️⃣ RTP Convergence Proof
    Confidence intervals at ±0.1% per denomination, with simulation volume justified by convergence, not ritual
  • 5️⃣ WinCap Verification
    Maximum single-ticket prize within $599 MN limit, verified across all denominations and deal sizes
  • 6️⃣ Dead Ticket Rate & Draw Analysis
    % non-winning tickets per deal, max drawdown (consecutive losses) at 99.9% Gumbel CI
  • 7️⃣ Prize-Based Closing (PBC) Simulation
    Deal closure behavior verified: when top prizes awarded → close triggered → remaining tickets refunded
  • 8️⃣ Chain Reaction Feature Analysis
    Bonus mechanic contribution to total RTP, trigger frequency, avg bonus payout — coupled parameter validation
  • 9️⃣ RNG / Shuffling Algorithm Documentation
    Ticket shuffling method, seed generation (crypto-grade), determinism proof for audit trail
  • 🔟 Full Constraint Convergence Matrix
    7 constraints × 5 denominations = 35 convergence proofs. All green = submission ready.

🏛️ MN GCB Specifics

Rule 7864.0235 Pull-tab game requirements
Rule 7861.0285 Electronic device standards
Max RTP 85%
Max single prize $599
PBC required Yes — mandatory

🔬 GLI-14 v2.2 Focus Areas

Math model accuracy RTP within ±0.1%
Prize distribution Full tier disclosure
RNG quality FIPS 140-2 or equivalent
Deal integrity All prizes awarded
Audit trail Full session reproducibility

"The certification gap closes by construction, not by process. The engine is the specification."

— MadLab Architecture Principle

MadLab Math Verification Framework

Chain Reaction × Diamond Game Enterprise × BrazLotto
MN GCB / GLI-14 Verification Framework
Built by MadLab Research · TwinGaming.AI · April 2026