Tech Talk: Ghost (Chris Chang)

Proprietary automated market makers (PropAMMs) now control a staggering 90% of aggregated swap volume on Solana, yet most users have no idea how these powerful trading engines actually work. At Breakpoint 2025, Chris Chang, co-founder of Ghost (the team behind Sandwiched.me, Solana's metadata hub), pulled back the curtain on these mysterious liquidity providers, revealing their inner mechanics and showing developers how to reverse engineer their closed-source strategies.

Summary

Top Projects

Lulo

Score: 7,586.13

Bonk

Score: 4,806.12

Sphere

Score: 2,544.42

Moonwalk

Score: 2,026.11

PropAMMs represent a fascinating evolution in decentralized finance infrastructure. While they appear to users as just another automated market maker, they function more like traditional on-chain market makers with sophisticated, actively managed quoting logic. Unlike open-source AMMs where anyone can inspect the code, PropAMMs operate with private strategies and pricing curves, with operators dynamically managing inventory and quotes without accepting external deposits.

The dominance of PropAMMs stems from their competitive advantages: permissionless swapping, extremely tight quotes, low gas costs, and lightning-fast reaction times. Chang noted that typically only two to four PropAMMs dominate the market at any given time, though the competitive landscape can shift rapidly. His team at Ghost became interested in these systems approximately nine months ago and has since developed simulation tools and visualizers available on their Sandwiched.me platform.

What makes PropAMMs particularly intriguing—and potentially concerning—is their closed-source nature. They don't provide decoded or open core interfaces, meaning the only way to see output at any given time is to actually execute a swap or run a simulation. Chang's presentation aimed to democratize this knowledge, arguing that while obfuscated sources raise the technical bar, they shouldn't be impossible to understand.

The presentation also explored the ongoing battle between PropAMMs and "toxic flow"—essentially arbitrage bots that exploit price discrepancies. Chang revealed that atomic arbitrage operations account for 5-10% of atomic arbitrage revenue, and PropAMMs have developed sophisticated countermeasures, including applying different penalty spreads based on which aggregator or router program initiated the swap.

Key Points:

How PropAMMs Structure Their Pricing

Chang walked through the architecture of a typical PropAMM, specifically referencing Screen File as an example. The system operates on multiple layers of protection and pricing mechanics. At the foundation is inventory control, which keeps reserves at a target level, paired with flow control that limits how much can be traded at any given price tier per slot.

The pricing engine uses a sophisticated four-layer approach. Layer one implements stale oracle protection—if the quoted price diverges too far from the oracle reference, the system simply refuses to quote. Layer two adjusts prices based on current reserves versus desired reserves, creating dynamic pricing based on inventory levels. Layer three penalizes quotes as oracle data becomes increasingly stale, ensuring traders pay more during periods of price uncertainty. Finally, layer four handles execution across up to 10 separate liquidity "lungs," each functioning as a distinct tranche with its own spread and capacity.

The Lung-Based Liquidity System

Perhaps the most innovative aspect of PropAMM architecture is the lung-based execution model. Each PropAMM maintains 10 lungs per trading direction, with each lung having its own capacity cap, meters, and spread measured in parts per million (PPM). When a trade comes in, the system fills from the base tranche first, then walks down the ladder to subsequent lungs as needed.

This creates a piecewise pricing function rather than a smooth curve, with explicit state tracked per lung limit. The system includes built-in backfill logic: after being consumed, a lung refills to 50% capacity on the next slot, then fully refills on the following slot if not consumed again. The final output for any trade is calculated as the sum of fills across all 10 lungs, with the system scanning from zero to nine and aggregating partial fills from each tranche.

Reverse Engineering Methodology

Chang outlined a three-pillar approach to understanding PropAMMs: simulation, tracing, and static analysis. The simulation pillar involves executing transactions against known state to answer the question "what would this output given this input mount on this state?" The team uses a modified BPF fork that instruments the virtual machine, logging a structured trace of every meaningful action including function calls, returns, and read/write operations.

From these traces, developers can rebuild code summaries showing which functions were invoked, how often, and how each function interacts with stack, heap, and account data. The static analysis pillar uses Binary Ninja with specialized plugins to decompile Solana programs. By topologically sorting code trees, AI agents can start from the outputs, translate them first, and work backward through the logic. Chang emphasized creating repeatable pipelines where every swap simulation becomes a "case file" for AI agents to analyze.

The Toxic Flow Arms Race

PropAMMs are engaged in constant warfare against toxic flow—arbitrage activity that exploits their pricing. Chang revealed that PropAMMs implement instruction introspection to examine full transactions, including who caused the swap, what swaps are present, and their ordering within the transaction. Many PropAMMs apply different pricing based on which aggregator program invoked them.

Specific examples include one PropAMM that adds 25 basis points penalty if another PropAMM swap appears in the same transaction, while another adds 3.6 basis points for certain router types. The arbitrage community has responded by evolving from single transactions to multiple transaction strategies, with multi-transaction atomic arbitrage now representing the majority of such activity. This cat-and-mouse game continues to evolve, with both sides developing increasingly sophisticated tactics.

Facts + Figures

• PropAMMs represent 90% of aggregated swap volume on Solana
• Typically only 2-4 PropAMMs dominate the market at any given time
• Atomic arbitrage operations generate 5-10% of atomic arbitrage revenue
• PropAMMs use up to 10 liquidity "lungs" per trading direction
• One PropAMM applies a 25 basis point penalty when detecting another PropAMM in the same transaction
• Another PropAMM applies a 3.6 basis point penalty for certain router types
• Lungs refill to 50% capacity in the first slot after consumption
• Lungs fully refill by the second slot if not consumed again
• Ghost team has been researching PropAMMs for approximately nine months
• Pricing spreads are measured in parts per million (PPM)
• Multi-transaction atomic arbitrage has grown to become the majority of arbitrage activity
• Ghost will soon release unique data executing unchanged transaction swaps between PropAMMs

Top quotes

• "PropAMMs now represent 90% of aggregated swap volume. They are a hot topic on Solana."

• "They behave more like an on-chain market makers with active coding logic. They're close source with private strategies and curves."

• "Closed and obfuscated sources can raise the bar, but they shouldn't be impossible to read."

• "Think about each lung as a separate tranche of liquidity with its own spread."

• "This is why quoting is not smooth—it's actually piecewise with explicit state per lung limit."

• "Reacting to toxic flow is one of the top priorities for PropAMMs."

• "Many PropAMMs use the instruction system to inspect full transactions—who caused me and what the swap present and ordering."

• "The arbitrage bot that faces full discrimination—they start from one transaction to multiple transactions."

• "For every single swap simulation, it becomes a case file for your AI agent to run through."

Questions Answered

What is a PropAMM and how is it different from a regular AMM?

A PropAMM (Proprietary Automated Market Maker) looks like a standard AMM from a user's perspective, but operates fundamentally differently under the hood. While traditional AMMs use open-source code with publicly viewable curves and strategies, PropAMMs use closed-source private strategies and pricing curves. They function more like on-chain market makers with actively managed quoting logic, where operators dynamically manage inventory and quotes. PropAMMs don't accept external deposits like traditional AMMs, and typically don't have their own front-end interfaces—users interact with them through aggregators instead.

Why do aggregators prefer PropAMMs?

Aggregators route significant volume through PropAMMs because they consistently offer the best prices. PropAMMs quote very tightly, meaning the spread between buy and sell prices is minimal. They also have very low gas costs, making trades cheaper to execute. Additionally, PropAMMs can react to market conditions extremely quickly, adjusting their quotes in response to price movements. This combination of tight spreads, low costs, and fast execution makes PropAMMs highly attractive for aggregators trying to find the best prices for their users.

How do PropAMMs protect themselves from being exploited?

PropAMMs implement multiple layers of protection. At the most basic level, they use oracle protection—if their quoted price diverges too far from reference oracle prices, they simply refuse to trade. They also adjust prices based on current inventory versus target inventory, charging more when reserves are depleted. PropAMMs penalize trades when oracle data becomes stale, reflecting increased uncertainty. Perhaps most sophisticated is their instruction introspection capability, where they examine the full transaction context to identify potential arbitrage activity and apply penalty spreads accordingly.

What are liquidity "lungs" in PropAMM architecture?

Lungs are a clever mechanism PropAMMs use to manage liquidity in tranches. Each PropAMM maintains 10 lungs per trading direction, with each lung representing a separate pool of liquidity at a specific spread level. When a trade comes in, the system fills from the base lung first, then moves to subsequent lungs as capacity is exhausted. Each lung has its own capacity cap and spread in parts per million. This creates piecewise pricing rather than a smooth curve, and includes automatic backfill logic where lungs regenerate capacity over time after being consumed.

Can closed-source PropAMMs be reverse engineered?

Yes, according to Chris Chang, closed-source PropAMMs can be understood through systematic reverse engineering. The approach involves three pillars: simulation (executing transactions against known state), tracing (using instrumented VMs to log structured traces of all operations), and static analysis (using decompilers like Binary Ninja). By combining these techniques with AI agents that can analyze the resulting data, developers can rebuild an understanding of how PropAMMs calculate their outputs. Chang argues that while obfuscation raises the technical bar, it shouldn't make understanding impossible.

What is toxic flow and why do PropAMMs care about it?

Toxic flow refers to trading activity that exploits PropAMMs, primarily arbitrage bots that extract value by trading against PropAMM quotes when they're slightly off-market. This activity accounts for 5-10% of atomic arbitrage revenue. PropAMMs care deeply about toxic flow because it represents a direct cost to their operations—every successful arbitrage trade means the PropAMM traded at a suboptimal price. In response, PropAMMs have developed sophisticated detection mechanisms that examine transaction context and apply penalty spreads when they suspect arbitrage activity, leading to an ongoing arms race between PropAMMs and arbitrageurs.