ChainFeeds PRO reviews AI agent verifiability, PropAMM data, and EIP-8288 recursion design

ChainFeeds PRO reviews AI agent verifiability, PropAMM data, and EIP-8288 recursion design

N
News Editor
2026-10-27 03:31:20
ChainFeeds Research’s PRO issue #163 brings together a wide set of crypto research updates, with three themes standing out: how AI agents can be verified in hostile environments, what cross-chain data says about PropAMMs, and how Ethereum’s proposed EIP-8288 could support more flexible recursive proof aggregation. The report starts with Bettina Warburg’s argument that crypto’s "dark forest" has already shown why verifiability matters more than explanations when software acts with real money and permissions. It then moves to Bitcoin, where Mr.Hodl argues that auditing what a signer actually did can be more direct than simply adding more hardware devices in a multisig setup, while Bitcoin Optech #428 covers Lichen, descriptor support for hash preimages, and Lightning security fixes. On Ethereum, the report highlights research from Category Labs showing PropAMMs outperforming traditional AMMs on short-term markout across Solana and Base, while also potentially offering better pricing to retail-like flow. It also reviews ignotusnemo’s "Paranoid EIP-8288" proposal for heterogeneous recursion with an authenticated class bank, and peersky’s BRO-AMM design, which aims to reduce LP losses from arbitrage through a bonded reconciliation process. The issue closes with MEV research notes and a paper on the geography of concurrent block proposing.

ChainFeeds Research has published PRO issue #163, a weekly roundup covering Bitcoin protocol work, Ethereum governance and research, MEV developments, and recent academic papers. This edition centers on three topics: AI agent verifiability, empirical evidence on PropAMMs across chains, and new design discussion around recursive proofs in EIP-8288.

ChainFeeds PRO reviews AI agent verifiability, PropAMM data, and EIP-8288 recursion design 2

AI agents and the "dark forest" problem

Animal Ventures co-founder Bettina Warburg begins with the well-known 2020 story "Ethereum is a Dark Forest." In that episode, a Paradigm researcher tried to recover about $12,000 from a contract, only to have the transaction spotted and front-run after it entered the public mempool. At the time, the predators were relatively simple bots that watched, copied, and paid higher gas. Warburg argues that those systems are now evolving into AI agents capable of forming their own strategies.

Her point is that crypto has spent more than a decade building an unusually hostile testing environment: participants do not trust one another by default, the assets are real, behavior is publicly verifiable, and mistakes carry immediate economic consequences. AI agents are now entering markets, APIs, and software systems with users’ money, credentials, and permissions. In her framing, they are walking into a very similar dark forest.

What crypto can teach AI, she argues, is not primarily how to understand why an agent took an action, but how to prove what it actually did. A bot strategy on Ethereum may be fully private, and even its developer may not fully understand it, yet every transaction input and execution result can still be re-executed and verified by any node. As agents begin shopping, paying, writing code, and managing assets on behalf of users, the bottleneck shifts from what AI can do to what humans can verify.

Warburg calls that gap the Verifiability Gap. If an agent claims it completed a payment within a user’s authorization, its own logs or a provider’s assurances are not enough. Independent evidence is needed to show that the agent had valid authorization, stayed within spending limits, and actually performed the operation it claimed to perform.

She summarizes the security model for an agent economy in four parts: verify code rather than identity, constrain permissions rather than trust behavior, leave verifiable proofs for important actions, and protect information that does not need to be public. Advanced AI Society’s Proof-of-Control effort is presented as one attempt to create that kind of standard, so third parties can verify whether an agent acted within its authorized scope without trusting the developer. In that sense, "Don’t trust, verify" does not remove trust from human relationships; it replaces blind trust in software and intermediaries with verifiable mechanisms.

Bitcoin protocol updates

Verification over simply adding more devices

Mr.Hodl examines a common hardware wallet security practice: if a user does not fully trust one signer, buy several devices from different vendors and use multisig. The report argues that multisig is well suited to permission separation, key-loss resilience, and collaborative control, but it does not prove that a given hardware signer actually executed the user’s instructions correctly. If the real concern is that the signer may be lying, a more direct answer is to verify what happened during signing.

EntropyLab is developing PSBT Signer Audit for that purpose. The tool stores the original PSBT sent to the hardware signer and compares it with the PSBT returned by the signer, checking inputs, outputs, amounts, sighash types, derivation data, change addresses, and whether signatures were altered. It can also check for ECDSA nonce reuse and Anti-Exfil data.

The underlying idea is to replace part of the security model of "buy more devices and spread trust" with "inspect what the signer actually did." Multisig still fits multi-party control and key redundancy, but when the threat model is a malicious signer, verification is presented as the more direct approach.

Other items from Bitcoin Optech Newsletter #428

On Lichen, Adam Gibson proposed a Bitcoin L2 with a single issuing hub and a sequencer that maintains the ledger. Each user has an independent 2-of-2 output with the hub. Even if the hub goes offline or refuses redemption, the user can still try to recover BTC through a Groth16 proof and a BitVMX on-chain dispute mechanism. The design does not remove all trust assumptions, however, because users still rely on the sequencer to maintain ledger state correctly.

On descriptor support for hash preimages, João Dias proposed extending BIP379 so wallet descriptors can directly store values such as sha256(preimage(...)) and other hash preimages. That would let Bitcoin Core wallets import and manage those secrets directly instead of relying on external tools to add preimages through PSBT.

On Lightning security, Eclair 0.14.4 and 0.14.5 fixed several issues involving splicing, HTLCs, trampoline routing, and onion parsing. LND, BDK, and BTCPay Server also recently fixed issues tied to synchronization, permissions, and duplicate payments.

Ethereum research and design work

PropAMMs across chains

Category Labs researchers Ozan Solmaz, Lioba Heimbach, and Jason Milionis studied PropAMMs on Solana, Base, and Monad over nearly a year. In a traditional AMM, LPs provide liquidity passively and pool prices usually update only when a trade occurs. If the external market moves first, arbitrageurs can trade against stale quotes, and that adverse selection keeps transferring value away from LPs.

PropAMMs work differently. A professional maker manages its own inventory and can keep updating on-chain quotes even when no swap takes place. By 2026, PropAMMs accounted for more than half of SOL/USDC trading volume on Solana, which the report treats as evidence that active liquidity is moving beyond the experimental stage and becoming a meaningful on-chain trading mechanism.

The clearest difference appears in post-trade markout. Two seconds after execution, PropAMMs on Solana and Base earned about 0.37 bps and 1.19 bps on average, while traditional AMMs lost about 0.22 bps and 0.62 bps.

The report lists four reasons for that edge: continuous repricing without waiting for arbitrage to restore prices, quote adjustments based on the risk of different order flow, avoidance of cross-venue arbitrage from other on-chain markets, and in some cases execution prices that are worse than displayed quotes. In practice, that makes PropAMMs look more like professional on-chain market makers than fully passive liquidity pools.

That does not automatically mean worse pricing for ordinary users. The study isolated trades where the reference price did not move materially and used them as a proxy for retail flow. On Solana, PropAMMs earned only about 0.26 bps from those trades, while standard AMMs earned about 2.59 bps. The report says that result suggests PropAMMs may actually offer more competitive pricing to traders without an informational edge.

Placed alongside recent PropAMM development on Ethereum, the broader takeaway is that on-chain liquidity is no longer just a model where anyone deposits funds into a fixed curve. Another model is taking shape: professional makers actively manage price and risk, users get tighter spreads, and liquidity depends more heavily on specialized market makers and the builders that support these mechanisms.

Paranoid EIP-8288 and heterogeneous recursion

EIP-8288 aims to recursively aggregate large numbers of signatures and proofs into one final proof that Ethereum validators can verify at low cost. The current direction leans toward a fixed protocol-level verification key, but community member ignotusnemo asks whether aggregation jobs of very different sizes should always use the same maximum-sized proving circuit.

If small aggregation tasks are forced into a relation designed for much larger workloads, proving resources are wasted. Based on the author’s Parano1d implementation, the proposal introduces heterogeneous recursion: the protocol would pre-approve multiple verifier classes with different sizes and structures, and each recursive step could choose the class that best fits the actual workload.

The hard part is that different proof classes do not share the same constraint matrix, witness size, or polynomial commitment parameters. At the same time, a deferred verification claim left by one recursive layer cannot simply disappear because the next layer switches classes. To handle that, the design adds an authenticated class bank.

Under this model, each permitted verifier class has its own accumulator lane. Any unfinished checks produced by one proof remain in the lane tied to that class. Even if later recursion switches to another class, those historical verification obligations stay alive. Terminal verification then checks both the current proof and every still-active lane, so no obligation from the recursive chain is dropped.

The advantage is that EIP-8288 would not have to choose between one huge unified circuit and several mutually incompatible proving systems. It could select different proof classes for different aggregation scales while still keeping them protocol-authenticated. The trade-off is straightforward: more classes mean larger public accumulator state and more work at final verification.

The report says several EIP-8288-specific questions remain open, especially how to merge multiple child proofs, how to merge multiple class banks, how to deduplicate and discard dependencies, and how to map the previously discussed Program/ISA Identity to verifier classes. At this stage, the proposal is framed as an exploration of a more flexible and upgradeable backend architecture for recursive proofs under EIP-8288.

BRO-AMM and post-trade reconciliation

Researcher peersky proposed BRO-AMM to address one of the main sources of LP losses in traditional AMMs: price updates happen through arbitrage. When ETH moves on centralized markets, an AMM may still quote the old price. Arbitrageurs then trade against the pool until the pool price converges back to the market. For LPs, that means paying continuously to get the pool repriced, while arbitrageurs burn a large share of the opportunity competing for block position.

peersky argues that this structure makes the idea of passively depositing capital into a pool and steadily earning fees harder to sustain. BRO-AMM adds a reconciliation stage that resembles block finality. Trades still execute immediately, but the final price is settled later against a trusted market price.

Arbitrageurs must first post a bond. If they merely help the pool correct a reasonable price gap, they receive what the system learns is fair compensation. If their profit exceeds that range, the excess is returned from the bond to LPs. Ordinary traders who do not repeatedly exploit stale prices can build reputation and receive lower fees later.

That keeps incentives for price discovery in place while preventing arbitrageurs from capturing the entire stale-price gap. In six-month simulations using real ETH/USDC data on Ethereum and Arbitrum, BRO-AMM reduced LP losses from arbitrage by about 68% to 79% on Ethereum and about 49% to 55% on Arbitrum. In the Ethereum simulation, LPs moved from an overall loss to profitability, while arbitrageur profits were largely unchanged because the funds returned to LPs mainly came from bidding costs that searchers had previously spent competing for block position.

The report also notes the limits: the design still depends on a trusted signed market price, and the results come from simulation rather than live deployment. Its significance lies in trying to turn arbitrage from an unavoidable LP loss into a price-discovery service that can be settled and constrained after the fact.

MEV notes from The MEV Letter #158

  • "Loss-versus-rebalancing with external proportional transaction costs" studies AMM LVR after adding external transaction costs and asks when LP performance can match an active rebalancing strategy, reaching zero LVR.
  • "How Much Screening Buys a Second? Speed, Selection, and the Maker’s Spread" examines the link between maker speed and order-flow screening. A maker updating quotes every 12 seconds needs to filter about 85% of informed flow to match the break-even spread of a maker updating every second with no screening.
  • "Active Liquidity On Chain: Evidence from PropAMMs Across Chains" shows how PropAMMs use active repricing and order-flow risk management to achieve better short-term markout than traditional AMMs, while also potentially offering better prices to retail flow.
  • "Glamsterdam scaling results on Sepolia, ten days in" reports that after ten days on Sepolia, average gas used per block rose about 134%, median block arrival became faster, and overall median payload execution time fell.
  • Posts including "Cambridge PQ R&D Workshop" summarize progress during Cambridge’s post-quantum research week, covering Ethereum PQ migration, AI-assisted formal verification, fast finality, decoupled consensus, and lean clients.
  • "EIP-8288 as a distributed proving network" describes a model in which the Ethereum mempool gradually takes on the role of a distributed proving network: signatures and proofs are aggregated before reaching the builder, and the builder ultimately produces one proof for lightweight validator verification.
  • "Frame Transactions" argues for shared account standards and wallet support early, so frame transactions can actually be adopted by users and wallets instead of repeating the fragmentation seen around EIP-7702.
  • "Blob Inclusion" asks why blob inclusion can remain unreliable even when blob fees are low, and says ePBS may improve that by giving blobs more time to propagate.
  • The video "Credible Commitments: The Holy Grail of Block Building" discusses MPBC, including builder coordination, operator economics, blob inclusion, and its relationship with FOCIL.

Paper: The Geography of (Concurrent) Block Proposing

In "The Geography of (Concurrent) Block Proposing," Ciamac C. Moallemi, Mallesh M. Pai, Dan Robinson, and Christoph Schlegel study the geographic incentives facing block proposers.

A single proposer tends to locate near major information sources to reduce latency. With multiple concurrent proposers, however, information that everyone receives at the same time has little extra value. What matters is exclusive information received earlier than competitors. The model shows that when there are multiple information sources, different proposers naturally spread across different locations.

The paper’s conclusion is that multiple concurrent proposers may do more than weaken the inclusion power of a single proposer. They may also push block production toward greater geographic decentralization through economic incentives.

This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan.
9300

Disclaimer:

The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.

Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.

ChainFeeds PRO reviews AI agent verifiability, PropAMM data, and EIP-8288 recursion design | Bit.Fan