Risk Engine

The Raphael Sentinel Risk Engine is the core decision-making system that converts raw analytical signals into a transparent, deterministic risk score for both SPL and bonding-curve tokens. Unlike traditional scanners that rely on generic heuristics, the Risk Engine uses a structured, behavior-aware model tailored specifically to Solana.

Its purpose is simple:

Measure risk in a way that accurately reflects how real attacks and failures happen on Solana.

The engine operates on top of the Analysis Pipeline and transforms its outputs into a single, intelligible score with full reasoning.


1. Deterministic, Transparent Scoring

Raphael Sentinel does not use black-box machine learning. Every value in the score is:

  • explainable

  • reproducible

  • based on on-chain evidence

  • derived from modular “risk categories”

This ensures that users can trust the score and understand why a token is flagged as risky.


2. Dual Scoring Models (SPL + Bonding Curve)

Because Solana has two fundamentally different token ecosystems, Sentinel maintains two independent models:

  • SPL Risk Model — for SPL tokens.

  • Bonding Curve Risk Model — for Pump.fun, LetsBonk, and other similar launches

Each model uses different weightings because the threat surfaces do not overlap.


3. SPL Risk Model

SPL tokens are primarily vulnerable to technical and market-structure risks.

The SPL model evaluates six core dimensions:

1. LP Stability

  • LP size

  • lock duration

  • provider concentration

  • liquidity manipulation patterns

Weak LP = higher rug risk.

2. Authority Safety

  • mint authority

  • freeze authority

  • update authority

  • metadata control

Retained authorities = high manipulation potential.

3. Holder Integrity

  • whale concentration

  • Gini coefficient

  • distribution fairness

  • early holder dominance

Centralization usually predicts dump events.

4. Market Integrity

  • wash trading

  • spoofed volume

  • bot-driven liquidity inflation

Fake activity hides instability.

5. Bot/Wash Trading Detection

  • MEV bots

  • snipe bots

  • wash loops

  • circulation patterns

Inorganic behavior = unreliable market.

6. Wallet Quality

  • historical rug participation

  • dev involvement

  • prior manipulative behavior

Bad actors → bad outcomes.

All SPL dimensions produce category scores that are aggregated into a 0–100 SPL Risk Score.


4. Bonding Curve Risk Model

Bonding-curve tokens eliminate LP rug risk — but introduce behavioral manipulation as the dominant threat.

Sentinel evaluates four core categories:

1. Dev Cluster Control

How much supply + influence dev wallets hold, including hidden swarm wallets.

2. Pre-Funded Swarm Behavior

Detection of artificial demand created using:

  • multiple wallets

  • identical timing

  • identical sizes

  • shared funding origin

This is the strongest sign of manipulation.

3. Exit Patterns

  • dump intensity

  • dump sequencing

  • exit velocity

  • asymmetry between dev exits and retail entries

This identifies coordinated exits.

4. Dev Historical Behavior

Dev reputation matters. If the same cluster launched multiple farm tokens → risk skyrockets.

All categories combine into a 0–100 Bonding Curve Risk Score.


5. Confidence Scoring

Each score includes a confidence level, based on:

  • data completeness

  • transaction density

  • cluster clarity

  • market age

  • statistical signal strength

Low confidence does not mean low risk — it means insufficient evidence.


6. Red Flags & Explanation Layer

For every token, Sentinel generates:

  • explicit red flags

  • detailed reasoning

  • evidence

  • affected categories

  • severity rating

This removes ambiguity and turns the risk score into an actionable insight.


7. Real-Time Risk Updates

Risk is not static. The engine recalculates in real time when:

  • dev clusters begin dumping

  • LP changes (SPL)

  • swarm wallets join/leave

  • exit momentum increases

  • new behavior emerges

This offers a huge edge over static scanners.


8. Final Output

Each token receives:

  • a final numeric score (0–100)

  • a risk level (Low, Medium, High, Extreme)

  • confidence level

  • categorized breakdown

  • timeline of risk evolution

  • red flags with explanations

This output flows into the UI, API, alerts, and subscriptions.

Last updated