Quick Answer: Most $DOM airdrop hunters optimize for volume. The actual scoring weights activity quality 3:1 over raw transaction count. I scraped 2,400 wallets from the DOM browser game beta and found that 4 inputs explain 67% of variance in final airdrop tier — and "days active" beats "total clicks" by a factor of 4.2.
TL;DR: DOM's airdrop scoring isn't a black box. It's a gradient-boosted model with known feature importances. If you're grinding the browser game, here are the 4 inputs that actually move the needle, with real numbers from live data.
"I Farmed 847 Clicks and Got Tier 3. My Friend Did 12 Sessions and Hit Tier 1."
That DM I got last week sums up everything wrong with airdrop meta. We're trained to optimize for visible metrics — transactions, volume, NFTs minted. DOM's scoring model doesn't care.
The DOM browser game launched in closed beta 11 weeks ago. It's a simple clicker: you manage a virtual node, stake points, upgrade hardware. The $DOM token airdrop was announced at week 6. Suddenly, 40,000 wallets started "playing."
I built a scraper. Tracked 2,400 wallets with known airdrop tiers (leaked from a frontend API bug, patched 72 hours later). Ran the features through XGBoost with SHAP values. Here's what actually predicts your tier.
The 4 Scoring Inputs, Ranked by Feature Importance
| Input | Weight | What It Actually Measures | Easy to Fake? |
|---|---|---|---|
| Days Active (streak-weighted) | 34% | Habit formation, not bursts | Hard |
| Session Depth (actions per session) | 28% | Real engagement vs. botting | Medium |
| Social Proof (referrals + verified) | 22% | Organic growth, KYC-adjacent | Hard |
| Economic Commitment (stakes locked) | 16% | Skin in the game | Easy (but costly) |
Raw transaction count? 3% importance. Below "browser fingerprint entropy." The team actively downweighted it after Sybil farmers flooded week 2.
Input 1: Days Active (The Compounding Killer)
DOM uses a modified Elo-style streak formula. Day 1 = 1 point. Day 2 consecutive = 2.5 points. Day 7 = 12 points. Miss a day, reset to 1.
The math is brutal. Two wallets:
- Wallet A: 30 days, all consecutive = 1,847 points
- Wallet B: 60 days, two 15-day streaks = 620 points
Wallet A wins despite half the calendar span. I verified this with 340 wallets in my sample. Correlation between streak score and final tier: r=0.71.
The anti-pattern: "I'll grind hard for 3 days then rest." DOM punishes this harder than any airdrop I've modeled. More on retention engineering — the same behavioral loops apply to product onboarding.
Input 2: Session Depth (Where Bots Die)
Actions-per-session distribution separates humans from automation. Real users show log-normal patterns: 8-15 actions, tail to 40. Bots cluster at round numbers — exactly 10, exactly 20 — or show uniform randomness.
DOM's model uses a Gini coefficient of action timing within sessions. Human clicks are bursty: 3 rapid, pause, 2 rapid, long pause. Bots are metronomic or perfectly random.
| Metric | Human Wallets (n=1,800) | Bot Wallets (n=600) |
|---|---|---|
| Avg actions/session | 12.3 | 11.7 |
| Timing Gini | 0.34 | 0.18 |
| Final tier | 2.1 | 4.2 (worse = higher number) |
The 600 "bot" wallets I flagged? All Tier 4 or unranked. They had more total actions than most humans. Didn't matter.
I spent 3 hours trying to script realistic timing patterns. Gave up. The feature engineering is too specific.
Input 3: Social Proof (The KYC Bypass)
DOM doesn't do full KYC. It does social graph verification — Twitter account age >180 days, mutual follows with 2+ existing verified users, and referral depth.
Key insight: referral quality > referral quantity. A chain of 3 verified users beats 50 unverified signups. The model uses PageRank on the referral graph, not in-degree.
One wallet in my sample: 3 referrals, all Tier 1 themselves. Scored higher than a wallet with 340 raw referrals, 98% unverified.
This is where DOM's scoring gets interesting. They're not measuring "how many people you brought." They're measuring "which community you belong to." It's a credentialing system disguised as growth hacking.
Input 4: Economic Commitment (The Expensive Signal)
Staking in DOM's browser game locks points for 7-90 days. Longer locks = higher weight. But here's the catch: early unstake penalties are 100% for the first 48 hours, then decay linearly.
The model uses "effective lock duration" — time-weighted by penalty paid. A 90-day lock unstaked at day 10 scores lower than a 30-day lock held to maturity.
| Stake Behavior | Effective Score | Final Tier (avg) |
|---|---|---|
| 90-day, held full | 90 | 1.2 |
| 90-day, unstaked day 15 | 12 | 3.1 |
| 30-day, held full | 30 | 1.8 |
| No stake | 0 | 3.6 |
This is classic costly signaling. The team wants to filter for users who believe in the project, not mercenary capital. It works. Correlation between effective lock and 30-day post-airdrop retention: r=0.54.
What I Got Wrong (And What DOM's Team Won't Confirm)
I don't have ground truth on model architecture. My XGBoost achieved 67% accuracy on tier prediction — decent, but the real model likely uses embeddings I can't reconstruct. The leaked API only exposed final scores, not intermediate features.
Honest limitation: My bot detection is heuristic-based, not ground-truthed. I flagged wallets by timing Gini <0.22 and session count >50 in 7 days. False positive rate unknown. Could be 10%. Could be 30%.
Also: DOM may have reweighted features after week 8. My data is weeks 6-9 only. The current model (week 11) could differ materially.
The Retention Connection
Here's why this matters beyond airdrop farming. DOM's scoring model is their retention model. The same features that predict airdrop tier predict 30-day post-claim activity.
Wallets in Tier 1-2: 73% still active 30 days post-airdrop. Tier 3-4: 31%. The airdrop isn't a marketing cost. It's a filtering mechanism for the user base they actually want.
This is where most GameFi projects fail. They airdrop to everyone, then wonder why their token dumps and their DAU collapses. DOM's approach — expensive to game, rewards genuine behavior — inverts the incentive.
Building the Same Analysis Infrastructure (For Actual Products)
If you're running a product with retention metrics, you need the same capability: behavioral feature extraction, model training, inference at scale. But with actual privacy guarantees.
Most teams ship user behavior data to OpenAI or Anthropic for analysis. Your retention model's training data — which users churned, which converted — sits on US servers, potentially in training sets.
We built something different. [VoltageGPU's Confidential Agent Platform](https://voltagegpu.com/for-developers-api) runs retention models inside Intel TDX enclaves. The data stays encrypted in RAM. Even we can't access it.
from openai import OpenAI
client = OpenAI(
base_url="https://api.voltagegpu.com/v1/confidential",
api_key="vgpu_YOUR_KEY"
)
response = client.chat.completions.create(
model="financial-analyst",
messages=[{
"role": "user",
"content": "Analyze this retention cohort data for churn signals..."
}]
)
print(response.choices[0].message.content)
Same OpenAI SDK. Hardware attestation that your data ran sealed. Compare vs. Azure's $14/hr DIY setup — we're at $3.75/hr for H100 TDX, ready in minutes not months.
Honest limitation: TDX adds 3-7% latency overhead vs. non-encrypted inference. For real-time retention scoring, that