> For the complete documentation index, see [llms.txt](https://gotts.gitbook.io/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://gotts.gitbook.io/docs/gotts-vaults/vault/research-citations.md). # Research Citations > **Part of**: [Vault PRD](/docs/gotts-vaults/vault.md) | **Last Updated**: 2026-02-20 > > Structured bibliography linking 185+ academic papers and production references to specific PRD mechanisms. This serves as an audit trail for design decisions and helps future contributors understand the theoretical basis for each mechanism. *** ## Citation Index by PRD Mechanism ### DynamicFeeEngine (06-contracts.md §10.3a, D-023, D-029, D-030) | Paper | Authors | Venue | Key Finding | Application | | --------------------------------------------------- | -------------------------------------- | ----------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ | | "Optimal Fees for Liquidity Provision in AMMs" | Campbell, Bergault, Milionis, Nutz | arXiv:2508.08152 (Aug 2025) | Optimal fee is remarkably stable normally but spikes during high volatility — threshold-type dynamic fee schedule is robust and optimal | Three-regime fee model: stable/elevated/spike | | "Optimal Dynamic Fees in AMMs" | Baggiani, Herdegen, Sánchez-Betancourt | arXiv:2506.02869 (Jun 2025, Oxford) | Dynamic fees linear in inventory approximate the true optimum; two distinct fee regimes | Inventory-linear formula for elevated regime | | "AMM and Arbitrage Profits in the Presence of Fees" | Milionis, Moallemi, Roughgarden | arXiv:2305.14604 (rev. Jul 2025); FC 2024 | Square-root decay of arbitrage profits with block time | L2 block-time calibration parameter | | "Modeling LVR via Continuous-Installment Options" | Singh et al. (Toronto) | arXiv:2508.02971 (Aug 2025); AFT 2025 | LVR = theta of perpetual continuous-installment put option; constant-LVR position design achievable; LVR independent of price path over forward windows | LVR-theta fee floor `lvrThetaFloorBps` guarantees LP adverse selection compensation (D-029) | | "Implied Volatility of AMM Fees" | Bichuch, Feinstein | Sep 2025 | Bijective mapping: `sigma = sqrt(8/T * ln(2*sqrt(Px0) / (2*sqrt(Px0) - fee)))`; fee-implied vol as zero-cost on-chain oracle; fixed-for-floating fee swap as novel DeFi derivative | Fee-implied volatility oracle source `_getFeeImpliedVolatility()` (D-030); manipulation detection via divergence from primary oracle | ### PA-AMM Activeness Controller (06-contracts.md §10.3b, D-027) | Paper | Authors | Venue | Key Finding | Application | | -------------------------------- | ----------- | --------------------------- | -------------------------------------------------------------------------------------------- | ----------------------------------------- | | "Partially Active AMMs (PA-AMM)" | (Anonymous) | arXiv:2602.09887 (Feb 2026) | Activeness parameter λ controls LVR-vs-tracking-error tradeoff; efficient frontier derivable | Per-vault lambda with tier-bounded ranges | ### Predictive Exit Timing (10-safety.md Layer 10, agents/09-agents-vault.md F2, D-031) | Paper | Authors | Venue | Key Finding | Application | | --------------------------- | ------------------------------------ | ---------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------- | | "Optimal Exit Time for LPs" | Bergault, Bieber, Sánchez-Betancourt | arXiv (Sep 2025) | LP withdrawal formalized as optimal stopping via HJB quasi-variational inequality; viscosity solution uniqueness proven; pre-emptive exit before arbitrageur price correction | Longstaff-Schwartz Monte Carlo for exit boundary computation; `OptimalExitSignal` in Layer 10 circuit breaker; vault-strategist framework #5 | ### Cross-Vault Defensive Rebalancing (06-contracts.md §10.1a, D-032) | Paper | Authors | Venue | Key Finding | Application | | -------------------------------- | -------------------------- | ---------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------ | | "Defensive Rebalancing for AMMs" | Devorsetz, Herlihy (Brown) | arXiv (Jan 2026) | Arbitrage-free iff Pareto efficient for CFMMs with log-concave trading functions; convex optimization yields unique solution; active/passive pool framework | `CrossVaultCoordinator` at factory level; eliminates inter-vault arbitrage leakage; reputation-gated participation (Verified+) | ### JIT Liquidity Strategy (agents/09-agents-vault.md F2, D-033) | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------- | ------------------------ | ------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------- | | "Strategic Analysis of JIT Liquidity" | Trotti et al. (CMU/UFRJ) | arXiv:2509.16157 (Sep 2025); AFT 2025 | First transaction-level JIT model; passive LPs lose 44% per JIT trade; JIT LPs could earn 69% more by optimizing price impact; if pool price = market price at entry, JIT LP never loses from price impact alone | vault-strategist framework #7: defensive (reduce PA-AMM lambda) and offensive (execute JIT on behalf of depositors) modes | ### LVF-Calibrated Auctions (06-contracts.md §10.4, §10.8, D-034) | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------------- | -------------------------------------- | -------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------- | | "Loss-Versus-Fair: Dutch Auctions on Blockchains" | Moallemi, Robinson (Columbia/Paradigm) | arXiv (Jun 2024); AFT 2024 | LVF >= sigma\*sqrt(dt/2) lower bound; closed-form time-to-fill; GDA profit decomposition ARB = VOL \* LVF+; z0 <= 0 Pareto optimal when value known | `lossFloorBps` in CCA adapter; LaunchFeeHook calibration (never start above NAV); Base 2s blocks yield \~1.2 bps unavoidable loss — confirms chain suitability | ### FM-AMM Batch Clearing (D-035) | Paper | Authors | Venue | Key Finding | Application | | ------------------------- | ---------------- | -------------------------- | ----------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------- | | "Batching Trades on AMMs" | Canidio, Fritsch | arXiv:2307.02074; AFT 2023 | Batch clearing eliminates LVR and sandwich attacks entirely; LP returns slightly exceed Uniswap v3 in simulations | Uniform-price batch auction for vault rebalancing intents | ### Formal Verification of AMM Invariants (10-safety.md V4 Hook Security) | Paper | Authors | Venue | Key Finding | Application | | -------------------------------------------- | ------------------------------ | ---------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------- | | "Formal State-Machine Models for Uniswap v3" | Tranquilli, Gupta | arXiv (Dec 2025) | PTA and FST formalisms; proves K' >= K (rounding always favors pool); verifies absence of rounding arbitrage | Pool Favor Property invariant check in V4 hook verification requirement #4; FST as formal interface specification for agent action spaces | | "Formal AMM Fee Mechanisms with Lean 4" | Dessalvi et al. (DTU/Cagliari) | arXiv (Jan 2026) | 3,500 lines machine-checked proofs; single large swap dominates split swaps under fees (generalized additivity theorem); equilibrium != profit-maximizing swap amounts | Fee monotonicity/output-boundedness verification baseline; single-swap dominance rule for vault rebalancing (06-contracts.md §10.10) | ### MCP Integrity Verification (mcp-server/09-safety.md §6.10, agents/11-safety.md Layer 2.5, D-025) | Paper | Authors | Venue | Key Finding | Application | | ----------------------------------------------- | ---------------------------------- | --------------------------- | ---------------------------------------------------------------------------------------- | -------------------------------------------- | | "Real AI Agents with Fake Memories (CrAIBench)" | Patlan et al. (Princeton/Sentient) | arXiv:2503.16248 (Mar 2025) | Memory injection > prompt injection; 150+ blockchain tasks, 500+ attack cases | Memory integrity hashing outside LLM context | | "TradeTrap" | Zhou et al. | arXiv:2512.02261 (Dec 2025) | "Fake MCP attack" — compromised server causes extreme portfolio concentration | Tool provenance cryptographic signing | | "From Prompt Injections to Protocol Exploits" | (Multi-author) | arXiv:2506.23260 (Jun 2025) | Adaptive prompt injection bypasses defenses >50% of cases; environment-injection 93% ASR | Formal MCP/A2A threat model | | "MCP-Guard" | (Multi-author) | arXiv:2508.10991 | Three-stage detection: static + semantic + fine-tuned E5; 96.01% accuracy | MCP-Guard integration (Stage 1-3) | | "AgentBound: Securing AI Agent Execution" | (Multi-author) | arXiv:2510.21236 | Declarative access control for MCP servers without modifications | Policy-based MCP server hardening | | "Securing the MCP: Risks, Controls, Governance" | (Multi-author) | arXiv:2511.20920 | Per-user auth, provenance tracking, sandboxing, inline DLP | MCP server governance framework | ### Adaptive Circuit Breakers (06-contracts.md §10.2, 10-safety.md, D-026) | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------------ | -------------- | --------------------------------- | -------------------------------------------------------------------------------------- | ------------------------------------------ | | "Systemic Risk in DeFi: Network-Based Fragility" | (Multi-author) | arXiv:2601.08540 (Jan 2026) | DeFi Correlation Fragility Indicator (CFI) from time-varying protocol TVL correlations | CFI-driven adaptive thresholds | | "Circuit Breakers and Market Runs" | (Multi-author) | Review of Finance, vol. 28 (2024) | "Magnet effect" — traders front-run anticipated halts; randomized halts solve this | Randomized trigger bands (±0.5%) | | "Aggregated Systemic Risk Index (ASRI)" | (Multi-author) | arXiv:2602.03874 (Jan 2026) | Unified DeFi-TradFi risk index; detected historical crises with 18-day lead time | ASRI as complementary systemic risk signal | ### Vault Disclosure Standard (06-contracts.md §10.1 VaultConfig, D-024) | Paper | Authors | Venue | Key Finding | Application | | ---------------------------------------------------------- | -------------- | --------------------------- | --------------------------------------------------------------------------------------------------- | -------------------------------------------- | | "Institutionalizing Risk Curation in Decentralized Credit" | Zbandut et al. | arXiv:2512.11976 (Dec 2025) | Two-layer DeFi credit system needs on-chain disclosure; curator concentration creates systemic risk | VaultDisclosure struct mandatory at creation | ### Endgame Defection Detection (06-contracts.md §10.19, 10-safety.md, D-028) | Paper | Authors | Venue | Key Finding | Application | | ---------------------------------------------------- | -------------------------------------- | --------------------------- | ----------------------------------------------------------------------------------------------- | ---------------------------------------------------------------- | | "Understanding LLM Agent Behaviours via Game Theory" | Huynh et al. | arXiv:2512.07462 (Dec 2025) | LLMs shift toward defection in endgames; strategy signatures classifiable >90% confidence | Behavioral regime classifier (cooperative/tit-for-tat/defective) | | "Multi-Agent Risks from Advanced AI" | Hammond, Chan et al. (50+ researchers) | arXiv:2502.14143 (Feb 2025) | Three failure modes: miscoordination, conflict, collusion; single-agent safety doesn't transfer | Multi-agent vault collusion/defection defense | | "Game Theory Meets LLMs: A Systematic Survey" | Sun et al. | arXiv:2502.09053 (Feb 2025) | LLMs exhibit pro-social bias but shallow strategy; inconsistent under uncertainty | Hard guardrails beyond reputation for strategic defense | ### Tier-Differentiated Loss Socialization (06-contracts.md §10.2) | Paper | Authors | Venue | Key Finding | Application | | ---------------------------------------------------- | ----------------- | --------------------------- | ----------------------------------------------------------------------- | ------------------------------------------------------------------------------------- | | "Autodeleveraging: Impossibilities and Optimization" | Chitra (Gauntlet) | arXiv:2512.01112 (Nov 2025) | ADL Trilemma: solvency vs revenue vs fairness impossible simultaneously | Tier-modulated loss policy: solvency-first for lower tiers, fairness-first for higher | ### Covered Call LP Framework (agents/09-agents-vault.md F2) | Paper | Authors | Venue | Key Finding | Application | | --------------------------------------------------- | ------------------------ | ----------------------------- | ------------------------------------------------------------------------------- | --------------------------------------------------- | | "Economic Model of DEX with Concentrated Liquidity" | Hasbrouck, Rivera, Saleh | Management Science (Feb 2025) | Concentrated LP ≈ covered call sold at intrinsic; LPs forgo option time premium | vault-strategist go/no-go criterion | | "Modeling LVR via Continuous-Installment Options" | Singh et al. | arXiv:2508.02971; AFT 2025 | LVR = continuous funding fees of embedded perpetual options | Forward-looking LVR from implied vol term structure | ### Oracle Architecture (06-contracts.md §10.2) | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------------------ | ----------- | ------------------------------------------ | ------------------------------------------------------------------------------ | --------------------------------------- | | "Ormer: Manipulation-resistant Pricing Oracle" | Bai et al. | arXiv:2410.07893 (Oct 2024, rev. Jun 2025) | Median-based pricing replaces vulnerable TWAP with inherent outlier robustness | Median-based NAV pricing on L2 | | "SecPLF: Secure Protocols against Oracle Manipulation" | Xu et al. | arXiv:2401.08520 (Jan 2024) | Per-asset price state tracking — only engage oracle if price outside bound | Threshold-based oracle refresh | | "OVer: Safeguarding DeFi against Oracle Deviations" | Deng et al. | ICSE 2024 | Symbolic analysis detects exploitable oracle scenarios proactively | Formal verification of agent strategies | ### Combinatorial CCA (06-contracts.md §10.4) | Paper | Authors | Venue | Key Finding | Application | | --------------------------------------------------------- | ---------------- | ---------------------------------- | ------------------------------------------------------------------------------------------------ | ------------------------------------------ | | "Fair Combinatorial Auction for Blockchain Trade Intents" | Canidio, Henneke | arXiv:2408.12225 (2024, rev. 2025) | Joint execution creates efficiencies via coincidence of wants; counterfactual fairness benchmark | Factory-coordinated combinatorial CCA bids | ### L2 Parameter Calibration (06-contracts.md §10.18, 04-architecture.md) | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------ | -------------- | --------------------------- | --------------------------------------------------------------------------- | ------------------------------------ | | "Layer-2 Arbitrage: An Empirical Analysis" | (Multi-author) | arXiv:2406.02172 (Jun 2024) | LVR overestimates L2 arbitrage by 5x; actual arbitrage 0.03–0.05% of volume | L2-specific parameter table | | "Optimistic MEV in Ethereum Layer 2s" | (Multi-author) | arXiv:2506.14768 (Jun 2025) | Cyclic arbitrage >50% of gas on Base/Optimism Q1 2025 | L2 MEV dynamics for vault strategies | *** ## Auction Mechanisms (Foundational) Papers informing CCA design, shill-proof auctions, and Dutch auction calibration: | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------------------ | ------------------------------- | -------------------------------- | ------------------------------------------------------------------------------------- | ------------------------------------------------------ | | "Shill-Proof Auctions" | Komo, Kominers, Roughgarden | EC 2025; arXiv:2404.00475 | Dutch auction is the unique revenue-optimal shill-proof mechanism | Formal justification for Dutch-style CCA | | "Loss-Versus-Fair: Dutch Auctions on Blockchains" | Moallemi, Robinson | AFT 2024 | Closed-form Dutch auction efficiency; GDA analysis for continuous token sales | CCA parameter calibration (decay rate, starting price) | | "Transaction Fee Mechanism Design Post-MEV" | Bahrani, Garimidi, Roughgarden | AFT 2024 | SAKA mechanism — deterministic, DSIC, sybil-proof, \~50% welfare guarantee | Vault-block builder interaction model | | "Credible, Optimal Auctions via Public Broadcast" | Ferreira, Kulkarni | AFT 2024 | Public blockchains enable larger class of credible auctions | Theoretical basis for on-chain CCA | | "Collusion-Resilience in TFM Design" | Chung, Roughgarden, Shi | EC 2024; arXiv:2402.09321 | Resilience to side payments imposes fundamental constraints | Agent-builder collusion resistance | | "Quantifying Price Improvement in Order Flow Auctions" | Bachu, Wan, Moallemi | arXiv:2405.00537; CAAW 2025 | 4–5 bps average improvement from auction-enhanced interfaces | CCA value-add evaluation methodology | | "Augmenting Batch Exchanges with CFMMs" | Ramseyer, Goyal, Goel, Mazières | EC 2024; arXiv:2210.04929 | Joint Price Discovery — impossibility of Pareto + price coherence + local computation | CCA design property tradeoffs | | "Batching Trades on AMMs" | Canidio, Fritsch | AFT 2023; arXiv:2307.02074 | Batch trading eliminates sandwich attacks and reduces LVR | Token launch batch execution protection | | "Auctions with Tokens" | Canidio | arXiv:2301.13794 (rev. Jun 2025) | Deflationary token monetary policy yields higher utility for risk-averse auctioneers | Vault token deflationary economics | *** ## AMM Theory (Foundational) Papers informing LP strategy, LVR modeling, and hook design: | Paper | Authors | Venue | Key Finding | Application | | --------------------------------------------- | -------------------------------------- | ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------ | | "Myersonian Framework for Optimal LP in AMMs" | Milionis, Moallemi, Roughgarden | ITCS 2024; arXiv:2303.00208 | Monopolist LP ≈ optimal auction design with "no-trade gap" | Vault LP demand curve calibration | | "Equilibrium Reward for LP in AMMs" | Aqsha, Bergault, Sánchez-Betancourt | arXiv:2503.22502 (Mar 2025, Oxford) | Stackelberg game closed-form equilibrium rewards | Optimal liquidity provision timing | | "Cost of Permissionless LP in AMMs" | Ma, Crapis (Ethereum Foundation) | arXiv:2402.18256 (2024) | Price of anarchy O(N) — permissionless provision efficient for long-tail tokens | Focus on agent-launched tokens with few LPs | | "Price-Aware AMMs" | Bergault et al. (Paris Dauphine/Swaap) | arXiv:2405.03496 (May 2024) | External oracle info optimizes quotes; Heston-Bates, jumps, Hawkes models | V4 hook oracle-aware pricing | | "Unified Hedging of IL" | Lipton, Lucic, Sepp | Digital Finance vol. 7 (2025); arXiv:2407.05146 | IL perfectly hedgeable with European vanilla options | AI-managed IL hedging via on-chain options | | "Strategic Analysis of JIT Liquidity" | Trotti et al. (CMU/UFRJ) | arXiv:2509.16157 (Sep 2025); AFT 2025 | First transaction-level JIT model; passive LPs lose 44% per JIT trade; JIT LPs could earn 69% more; JIT LP never loses from price impact if pool=market at entry | JIT-style AI agent LP capital allocation; defensive/offensive vault strategy (D-033) | | "pm-AMM for Prediction Markets" | Robinson, Moallemi (Paradigm) | November 2024 | Given asset model, derive optimal AMM invariant — uniform AMMs where LVR ∝ portfolio value | V4 hooks with asset-type-specific invariants | | "Optimal Routing in Presence of Hooks" | Chitra et al. | Working paper (2025) | Mean-variance optimization for splitting trades between risky hooks and guaranteed CFMMs | Vault rebalancing: CCA vs AMM allocation | | "Optimal Execution on Uniswap v2/v3" | Baude, Challet, Toke | arXiv (Jan 2026) | Closed-form optimal execution for CPAMM; multi-layer DP framework for concentrated liquidity; transient impact model | vault-strategist optimal execution sizing for rebalances | | "Efficiency of Constant Log Utility MMs" | Papireddygari et al. | arXiv (Oct 2025); AAMAS 2026 | CLUM pricing is #P-hard; (1+epsilon)-approximation algorithm; polynomial-time oracle for interval securities | Theoretical complexity bounds for agent pricing computations | | "Dynamic Exponent Market Maker (DEMM)" | Kositwattanarerk | arXiv (Jul 2025) | AMM where constant-product exponent adjusts dynamically with deposits; enables personalized portfolios in single pool | Future: personalized vault LP strategies (flash loan vulnerability unsolved) | | "Orbital: Multi-Asset AMM" | White, Robinson, Moallemi | arXiv:2510.05428 | Center-curve superellipse for 2 to 10,000+ stablecoins; concentrated liquidity in arbitrary dimensions | Future: multi-asset stablecoin vault strategies | | "Semi-fungible Asset Clearing" | Diamandis, Chitra, Angeris | arXiv:2505.19298 | Convex optimization for pricing assets with partial ordering (different-expiry bonds, staked ETH variants) | Vaults holding heterogeneous yield-bearing positions | *** ## Reputation & Identity Systems | Paper | Authors | Venue | Key Finding | Application | | --------------------------------------------------------------------- | ------------------------------- | -------------------------------- | --------------------------------------------------------------------------------------------- | ----------------------------------------------- | | "ERC-8004: Trustless Agents" | Crapis, De Rossi, Ellis, Reppel | EIP (Aug 2025), mainnet Jan 2026 | Identity (ERC-721), Reputation (feedback), Validation (check hooks) registries | Universal identity gate for vault access | | "Deep Reputation Scoring: zScore-Based Wallet Ranking" | Kandaswamy et al. | arXiv:2507.20494 (Jul 2025) | Dual LP + Swap scores; deep residual neural network refinement | 5-tier system scoring blueprint | | "zScore: Universal Decentralised Reputation" | Udupi et al. | arXiv:2503.05718 (Mar 2025) | Validated against Feb 2024 Aave V3 liquidation — predicted risky users | Empirical validation of tier gating approach | | "Attribute-Based Threshold Issuance Anonymous Counting Tokens (tACT)" | (Multi-author) | IACR ePrint 2024/1024 | Sybil-resistant identity with threshold issuance, unlinkable multi-show, constant-size proofs | Sybil defense for multi-registration prevention | | "Decentralized Society: Finding Web3's Soul" | Weyl, Ohlhaver, Buterin | SSRN (2022, cited through 2026) | Soulbound Tokens for undercollateralized lending (reputation as collateral) | SBT-based vault milestone credentials | *** ## Agent-Specific DeFi & Security | Paper | Authors | Venue | Key Finding | Application | | --------------------------------------------------------------- | --------------------------------- | ------------------------------ | ---------------------------------------------------------------------------------------------- | -------------------------------------------- | | "Autonomous Agents on Blockchains: Standards, Execution, Trust" | Multi-institutional SLR | arXiv:2601.04583 (Jan 2026) | Six-stage pipeline (observe→reason→plan→sign→submit→confirm); MCP/UTCP/A2A review | Vault agent interaction pipeline model | | "Mechanism Design for LLMs" | Duetting et al. (Google Research) | ACM WWW 2024; arXiv:2310.10826 | Payment monotonicity ↔ consistent aggregation; higher reputation → monotonically better access | Incentive-compatible reputation-gated access | | "Multi-Agent AI Architecture for DeFi Investment" | Liu et al. | 2025 (SuperIntent) | Intent-driven multi-agent architecture; CryptoRLPM achieves 83% higher returns vs BTC | Multi-agent vault management template | *** ### Reputation & Behavioral Scoring (D-028, agents/08-agents-economy.md) | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------------------ | ----------------------------------- | ----------------------------------- | ------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------- | | "Deep Reputation Scoring: zScore-Based Wallet Ranking" | Kandaswamy et al. | arXiv:2507.20494 (Jul 2025) | Behavioral zScores decomposing LP behavior into volume, frequency, holding time, withdrawal patterns via deep residual networks | Closest academic work to reputation-weighted economics; informs ERC-8004 agent identity scoring layer | | "Equilibrium Reward for LP in AMMs" | Aqsha, Bergault, Sánchez-Betancourt | arXiv:2503.22502 (Mar 2025, Oxford) | Stackelberg stochastic game: LPs should add liquidity only when higher liquidity attracts more noise trading | vault-strategist framework #6: `marginal_noise_trading(dL) > marginal_LVR_cost(dL)` check before liquidity increases | ## Vault Economics & Curator Layer | Paper | Authors | Venue | Key Finding | Application | | ---------------------------------------------------- | ----------------------------------- | -------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------- | | "Delegated Portfolio Management with Random Default" | Gennaro, Mastrolia | arXiv:2410.13103 (Oct 2024) | BSDEs for optimal compensation under uncertain horizons (crashes, hacks, withdrawals) | Fee structures for uncertain agent operational horizons | | "am-AMM: Auction-Managed AMM" | Adams, Moallemi, Reynolds, Robinson | arXiv:2403.03367 (2024); FC 2025 | Harberger lease auctions vault management rights; winning bidder pays rent, sets fees, captures per-block arbitrage; attracts higher equilibrium liquidity than fixed-fee models | StrategyAuctionModule (D-012); vault-auctioneer agent; deployed in production as Bunni v2 | *** ## Risk Management | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------ | ------------------------------- | --------------------------- | ------------------------------------------------------------------- | -------------------------------------- | | "Systemic Risk in DeFi" | (Multi-author) | arXiv:2601.08540 (Jan 2026) | CFI metric from time-varying correlation networks | Adaptive circuit breaker thresholds | | "ERC-7265: Circuit Breaker Standard" | Diyahir, tcb\_00, real\_philogy | EIP (2023–2025) | Temporary halts on token outflows; delay settlement or revert modes | Identity-gated outflow circuit breaker | *** ## Composability & L2 Design | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------------------------ | ----------------------------------- | --------------------------- | -------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------- | | "Uniswap v4 Core Whitepaper" | Adams et al. | August 2024 | Hooks, singleton model, flash accounting, dynamic fees, native ETH | Foundational V4 hook architecture | | "Overview of Uniswap v4 for Researchers" | Bachu, Hasbrouck, Saleh, Wan | SSRN (Feb 2025) | Academic mapping of hooks to concentrated liquidity capabilities | Hook-PoolManager interaction reference | | "Framework for Combined L2 Transaction Posting and Pricing" | Wang, Crapis, Moallemi | FC 2025; arXiv:2505.19556 | Unified L2 batching and pricing optimization | Vault operation timing optimization on Base | | "Quantifying Value of Revert Protection" | Zhu, Wan, Moallemi, Robinson, Bachu | FC 2025; arXiv:2410.19106 | Economic value of guaranteed execution | Revert protection for multi-step vault operations | | "Multidimensional Blockchain Fees Are Optimal" | Angeris, Diamandis, Moallemi | AFT 2025; arXiv:2402.08661 | Separate resource pricing is nearly optimal | Multi-dimensional gas optimization on L2 | | "Interoperable Architecture for Digital Identity Delegation" | Saavedra | arXiv:2601.14982 (Jan 2026) | Delegation Grants as first-class authorization artifacts | Tier-scoped depositor-to-agent authority | | "Framework for Combined L2 Transaction Posting and Pricing" | Wang, Crapis, Moallemi | arXiv:2505.19556; FC 2025 | Joint optimization of L2 batching and pricing; directly relevant to vault operation timing on Base | Vault rebalance timing optimization; L2 fee posting cost minimization | *** ## TEE Security & Agent Wallet (D-036, D-040) | Paper | Authors | Venue | Key Finding | Application | | ---------------------------------------------------------------------- | ---------------------- | ------------------------------------------ | ------------------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------- | | "BadRAM: Practical Memory Aliasing Attacks on TEEs" | De Meulemeester et al. | IEEE S\&P 2025, pp. 4117-4135 | DDR4/DDR5 SPD chip tampering <$10 breaks AMD SEV-SNP; some DIMMs enable software-only SSH attack | Prefer Intel TDX over AMD SEV-SNP; require SPD-locking verification | | "Battering RAM: Low-Cost Interposer Attacks on Confidential Computing" | Van Bulck et al. | IEEE S\&P 2026 (accepted Sep 2025) | DDR4/DDR5 memory interposer <$50 breaks Intel TDX, AMD SEV-SNP, AND NVIDIA CC; forges attestation quotes | TEEs are defense-in-depth only; time-delayed proxy is primary security primitive | | "SEAgent: Taming Privilege Escalation in LLM-Based Agent Systems" | Ji et al. | arXiv:2601.11893 (Jan 2026) | Confused deputy problem; inter-agent trust exploitation 84.6% ASR vs 46.2% direct prompt injection; ABAC framework | Agent authorization boundary as critical attack surface; ABAC policy model | | "Progent: Programmable Privilege Control for LLM Agents" | Shi et al. | arXiv:2504.11703 (Apr 2025) | First privilege control DSL with Z3 SMT formal verification for fine-grained tool access | Formal wallet permission specification with provable guarantees | | "CaMeL: Defeating Prompt Injections by Design" | Debenedetti et al. | arXiv:2503.18813 (Mar 2025; 66+ citations) | Capability-based security separating control/data flow; 77% task completion with provable security | Capability tokens for vault agent transaction authorization | | "Design Patterns for Securing LLM Agents against Prompt Injections" | Beurer-Kellner et al. | arXiv:2506.08837 (Jun 2025) | Six principled patterns; general-purpose agents with broad capabilities inherently insecurable | Action-Selector + Plan-Then-Execute patterns; validates narrow-scope vault agent design | ### ERC-4626 Security Validation | Paper | Authors | Venue | Key Finding | Application | | -------------------------------------------------------------- | ---------------- | --------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------- | | "Smart Contract and DeFi Security Tools" | Chaliasos et al. | ICSE 2024; arXiv:2304.02981v2 | Only 8% of $2.3B losses preventable by automated tools; oracle/logic vulns undetectable | Validates internal asset accounting as primary security invariant | | "DeFiAligner: Leveraging Symbolic Analysis and LLMs" | Gan et al. | AFT 2024, LIPIcs Vol. 316 | Detects specification-implementation inconsistencies via symbolic + LLM analysis | Verify ERC-4626 spec consistency for totalAssets/convertToShares/convertToAssets | | "SoK: Decentralized Finance Attacks" | Zhou et al. | IEEE S\&P 2023, pp. 2444-2461 | 103 of cataloged attacks were non-atomic — defenders have rescue time window | Partially validates 6-hour profit unlock as providing viable defense windows | | "OVer: Safeguarding DeFi against Oracle Deviations" (extended) | Deng et al. | arXiv:2401.06044; Bank of Canada Staff Discussion Paper 2024-10 | SMT solver-based parameter safety ranges; current protocol params inadequate for significant deviations | Apply OVer methodology to rate-limiting thresholds on convertToAssets bounds | *** ## AMM Fee Calibration (D-037, D-038) | Paper | Authors | Venue | Key Finding | Application | | --------------------------------------------------------------- | -------------------- | --------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------- | | "LVR under Deterministic and Generalized Block-Times" | Nezlobin, Tassy | arXiv:2505.05113 (May 2025) | Constant block-time uniquely minimizes asymptotic LVR; Corollary 3.1 for PoS; fees 5-15% lower than Poisson formula | Use constant-block-time formula for Base; L2 fees lowered from 8 to 5 bps (D-037) | | "IL and LVR II: Arbitrage Time" | Loesch | arXiv:2502.04097 (Feb 2025) | Three temporal regimes; arbitrage time tau\_arb = f^2/sigma^2 | Principled sigma\_low/sigma\_high threshold calibration | | "Adaptive Curves for Optimally Efficient Market Making" | Nadkarni et al. | arXiv:2406.13794 (Jun 2024, updated Mar 2025) | Kalman filter on trader behavior (Glosten-Milgrom); static CFMMs only optimal when inter-block time vanishes | Kalman filter replacement for EWMA sigma estimation (D-038) | | "Measuring Arbitrage Losses and Profitability of AMM Liquidity" | Fritsch, Canidio | arXiv:2404.05803 (Apr 2024; WWW'24 Companion) | Fees don't sufficiently compensate in most large pools; 20-70% loss decrease from 12s to 100ms blocks; NOT uniform sqrt | Pair-specific LVR-block-time calibration tables | | "Game Theoretic Liquidity Provisioning in CLMMs" | Tang et al. | arXiv:2411.10399 (Nov 2024; SIGMETRICS 2025) | NE follows waterfilling strategy; LPs improve $116/day; code: WeizhaoT/CLMM-game | Default vault-strategist range selection algorithm (D-041) | | "Efficient LP with Deep Reinforcement Learning" | Xu, Brini | arXiv:2501.07508 (Jan 2025) | PPO agent dynamically adjusts concentrated liquidity; rolling window for regime shifts | PPO overlay for non-stationary dynamics on waterfilling baseline | | "Optimal Risk-Aware Interest Rates for Decentralized Lending" | Baude, Challet, Toke | arXiv:2502.19862 (Feb 2025) | Optimal rates via stochastic control; liquidity risk penalty function for idle capital | Rehypothecation cap at 60% of idle capital (D-042) | *** ## Circuit Breaker Design (D-043) | Paper | Authors | Venue | Key Finding | Application | | ----------------------------------------------------- | ------------------------------ | --------------------------------------------------------- | ------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------- | | "The Dark Side of Circuit Breakers" | Chen, Petukhov, Wang, Xing | Journal of Finance, 79(2), 2024, pp. 1405-1455 | Near threshold: volatility rises drastically, negative skewness, trading spikes — magnet effect | Reinforces randomized trigger bands; prohibits binary halts | | "Circuit Breakers and Market Runs" (extended) | Bongaerts, De Luca, Van Achter | Review of Finance, 28(6), 2024, pp. 1953-1989 | Forward-looking welfare-optimized breaker; too-tight binary breakers create rather than mitigate runs | DeFi circuit breakers must never impose hard binary halts; graduated dampening essential | | "When Do Circuit Breakers Stabilize Markets?" | Hu, Ming | Annals of Economics and Finance, 26(2), 2025, pp. 573-613 | Mild shocks: dampen effectively; severe crises: amplify fluctuations | Regime-aware dampening curve shapes (alpha varies by CFI) | | "Stablecoin Runs and the Centralization of Arbitrage" | Ma, Zeng, Zhang | NBER Working Paper 33882 (May 2025) | Concentrated arbitrage amplifies run risks; rate-limiting redemptions reduces strategic complementarity | Strongest academic justification for withdrawal dampening in DeFi vaults | | "Hyper-heuristic Driven Smart Contracts for DeFi" | (Multi-author) | Frontiers in Blockchain (2025) | Self-adjusting contracts; Compound Proposal 62 failure ($90M+ from governance latency >15K blocks) | Autonomous circuit breaker threshold adjustment; security bounds critical | *** ## Reputation & Game Theory (D-044, D-045, D-028 refinements) | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------------------------------------ | ----------------------------- | ----------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------- | | "TraceRank: Sybil-Resistant Service Discovery for Agent Economies" | Shi | arXiv:2510.27554 (Oct 2025) | Payments-as-endorsements; reputation propagates through payment flows; integrates ERC-8004 | Vault interactions as reputation seed scores (D-044) | | "On Sybil-Proof Mechanisms" | Schlegel, Mattauch, Moldovanu | arXiv:2407.14485v3 (Jul 2024, updated May 2025) | Only Sybil-proof mechanism = second-price auction with symmetric tie-breaking; proportional allocation can't be both Sybil-proof and IC | Reputation-weighted allocation accepts Bayesian relaxation or auction-like mechanisms | | "Nicer Than Humans: LLMs in the Prisoner's Dilemma" | Fontana et al. | arXiv:2406.13605v2 (Sep 2024) | Cooperation emerges sharply when defection drops below 30%; Llama3 consistently exploitative | 30% defection rate threshold for reputation penalties | | "Strategic Intelligence in LLMs: Evidence from Evolutionary Game Theory" | (Multi-author) | arXiv:2507.02618 (Jul 2025) | Gemini correctly identifies high termination probability and rationally defects; OpenAI models cooperative | Mask termination horizon; use probabilistic continuation for vault management periods | | "Repeated Games with Partner Choice" | Graser et al. | PLOS Computational Biology (Feb 2025) | No equilibrium where all start cooperating in round 1; mandatory testing phases | Validates starting new agents at low reputation with mandatory probationary period | | "Dynamics of Cooperation in Concurrent Games" | Rossetti et al. | Nature Communications 16, Article 1524 (2025) | Cross-game spillover reduces cooperation; cognitive constraints link behavior across games | Protocol-specific reputation with controlled cross-protocol signaling (D-045) | *** ## Factory Economics & Insurance (D-039, D-046, D-047) | Paper | Authors | Venue | Key Finding | Application | | ----------------------------------------------------------------- | ------------------------------ | ----------------------------------------------------------------- | --------------------------------------------------------------------------- | -------------------------------------------------------------------------- | | "A Unified Theory of Decentralized Insurance" | Feng, Liu, Taylor | Insurance: Mathematics and Economics, Vol. 119, 2024, pp. 268-297 | Hierarchical structures outperform flat pools under many conditions | Hierarchical tranched insurance: junior/senior structure (D-039) | | "DeFi Composability as MEV Non-Interference" | Bartoletti et al. | arXiv:2309.10781v2 (Jan 2024; FC 2024) | Formal composability = MEV non-interference for compound contracts | Formal verification for vault-of-vaults compositions at factory level | | "Disentangling DeFi Compositions" | Kitzler et al. | ACM Transactions on the Web (2024) | Large-scale composition measurement; contagion case study | Meta-vault composition depth limit of 2 (D-046) | | "Determinants of Funding Liquidity Risk in Decentralized Lending" | Nguyen et al. | Global Finance Journal, Vol. 64 (Mar 2025) | Lower deposit concentration may exacerbate funding liquidity risk | Time-weighted withdrawal mechanisms; soft lockups for small deposits | | "Decentralised Dealers?" | Aquilina et al. | BIS Working Paper 1227 (2024) | 65-85% of V3 liquidity from professional group; retail LPs earn less | Validates vault-as-curator; performance-based fees (hurdle + carry, D-047) | | "A Theory of Lending Protocols in DeFi" | Bartoletti, Ferrando, Ferraris | arXiv:2506.15295 (Jun 2025) | Formal game-theoretic model; incentive alignment harder with pool diversity | Verify factory markets satisfy solvency invariants before activation | | "Cross-Chain Arbitrage: The Next Frontier of MEV" | Oz et al. | arXiv:2501.17335 (Jan 2025, revised Jun 2025) | 242,535 cross-chain arbs totaling $868.64M; top 5 addresses >50% of trades | Cross-vault rebalancing as auction rather than open MEV | | "Delegated Portfolio Management with Random Default" (extended) | Gennaro, Mastrolia | arXiv:2410.13103 (Oct 2024) | BSDEs for optimal compensation under uncertain horizons | Fee structure for uncertain agent operational horizons | *** ## Identity-Gated Pools (Additional Evidence) | Paper | Authors | Venue | Key Finding | Application | | --------------------------------------------------- | -------------------------------- | ------------------------------------------- | --------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------ | | "Cost of Permissionless LP in AMMs" (extended) | Ma, Crapis (Ethereum Foundation) | arXiv:2402.18256 (2024); FC 2024 | Price of anarchy O(N); permissionless provision efficient for illiquid tokens but wasteful for liquid assets | Most important theoretical paper for identity gating; O(N) welfare loss justifies controlled LP entry | | "Privacy-Preserving On-chain Permissioning" | (Multi-author) | arXiv:2510.05807 (Oct 2025) | SSI + ZKPs + ABAC for privacy-preserving permissioning; multiple proof types | Future: commit-and-prove ZKP credential verification at pool entry | | "BEAT-MEV: Epochless Batched Threshold Encryption" | Bormet et al. | IACR ePrint 2024/1533; USENIX Security 2025 | <2ms encryption, <440ms decryption for 512 txs; eliminates epoch costs | Encrypted order flow complement to identity gating for defense-in-depth | | "Centralizing Effects of Private Order Flow on PBS" | Gupta, Pai, Resnick | arXiv:2305.19150; AFT | Private order flow complementary to top-of-block; top builder 18%→\~50% share in one year | Identity-gated pools should maintain builder diversity | | "Thorns in the Rose: Security Risks in V4 Hooks" | BlockSec | 2024 | 36% of community hooks vulnerable; flawed access control + input validation | All identity-gating hooks require rigorous onlyPoolManager + audit | | Cork Protocol exploit | BlockSec analysis | May 28, 2025; $11M loss | Hook callback functions called directly without routing through PoolManager due to missing `onlyPoolManager` modifier | Canonical production reference for hook access control requirement; identity cache design in VaultHook | *** ## CCA/Auction Extensions | Paper | Authors | Venue | Key Finding | Application | | -------------------------------------------------- | ------------------ | ----------------------------------------- | -------------------------------------------------------------------------------------- | ------------------------------------------------------------------ | | "Uniswap Liquidity Launchpad / CCA" | Adams et al. | Uniswap whitepaper (Nov 2025) | First production CCA; $59M Aztec sale; Q(t) release schedule; LBP Strategy module | CCA integration blueprint; bid-spreading mechanism removes sniping | | "Mechanism Design for Automated Market Makers" | Chan, Wu, Shi | AFT 2025, LIPIcs Vol. 354 | Batch-processing AMMs with constant potential function achieve provable MEV-resilience | Strengthens FM-AMM batch clearing path (D-035) | | "Loss-Versus-Fair: Dutch Auctions" (L2 correction) | Moallemi, Robinson | AFT 2024, LIPIcs Vol. 316, pp. 18:1-18:17 | LVF ≈ sigma^2 \* dt / (2\*delta); L2 (2s blocks) reduces LVF \~6x vs Ethereum | Steeper CCA decay rates on Base with equivalent efficiency | *** ## Recursive Lending & Leverage Loops (D-048) | Paper / Source | Authors | Venue | Key Finding | Application | | ------------------------------------------------------------- | -------------------- | --------------------------- | ---------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------- | | Morpho Blue lending volume data | Morpho Labs | Production data (2025) | 64% of Morpho volume comes from looping (recursive lending) strategies | Validates recursive lending as primary DeFi leverage mechanism | | Aave V3 E-Mode specification | Aave | Production docs | 90% LTV for ETH-correlated assets on Base; health factor mechanics | RecursiveLendingAdapter parameter calibration (90% LTV, HF thresholds) | | "Optimal Risk-Aware Interest Rates for Decentralized Lending" | Baude, Challet, Toke | arXiv:2502.19862 (Feb 2025) | Stochastic control with liquidity risk penalty function for optimal lending rates | Informs `maxLeveragedAumBps` (40%) and health factor floor calibration | | Sommelier Real Yield ETH vault | Sommelier Finance | Production data (2023-2025) | Off-chain ML + on-chain execution pattern; outperformed stETH holders by 2.25x; avoided losses during USDC depeg and Curve exploit | Reference architecture for pre-emptive deleveraging with off-chain signals | | Contango flash-loan looping | Contango | Production on Base (2025) | cPerps with funding rates 3x lower than centralized exchanges via automated flash-loan looping | Integration target for atomic loop execution | *** ## Perpetual Yield Tranches (D-049) | Paper / Source | Authors | Venue | Key Finding | Application | | --------------------------------------------- | ------------------------------------- | ---------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | | Idle Finance Perpetual Yield Tranches | Idle Finance (rebranded Pareto, 2025) | Production (Dec 2021+) | Senior (AA) receives protected yield; Junior (BB) provides first-loss and earns up to 90% of yield via adaptive split. Perpetual, no-lock, ERC-4626 compatible. \~$14.3M TVL with 19-40% yields on wstETH | TrancheModule.sol design: adaptive yield split, Senior/Junior token structure | | Goldfinch junior/senior tranche model | Goldfinch | Production (\~$170M originated) | Backers (junior) as first-loss capital; Senior Pool auto-diversifies; waterfall directs repayments Senior-first; 20% of Senior yield reallocated to Junior | Waterfall mechanics and junior-first loss absorption pattern | | Centrifuge DROP/TIN tranching | Centrifuge | Production ($1.2B+ TVL on Base V3) | DROP (senior) / TIN (junior) token tranching for real-world assets; JAAA product exceeding $1B TVL | Deployed-at-scale validation of on-chain tranche structures | | "A Unified Theory of Decentralized Insurance" | Feng, Liu, Taylor | Insurance: Mathematics and Economics, Vol. 119, 2024 | Hierarchical structures outperform flat pools under many conditions | Theoretical basis for tranche superiority over flat risk pools (complements D-039) | *** ## ve-Token Economics (D-050) | Paper / Source | Authors | Venue | Key Finding | Application | | -------------------------- | ------------- | ------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------- | | Velodrome Finance codebase | Velodrome | Production on Optimism (2022+) | Battle-tested ve(3,3) implementation; gauge voting, bribe market, emission allocation infrastructure | Fork path for veVAULT governance contracts | | Curve Finance veCRV | Curve Finance | Production (2020+) | Pioneered vote-escrowed tokenomics; 20+ protocols adopted variants; lock duration determines voting power with linear decay | Original ve-token mechanism design | | "Novel tokenomics" section | Research | `research/defi-primitives2.md` | Agent bonding (operators stake factory tokens, slashed for poor performance); revenue sharing (50% veVAULT, 30% insurance, 20% treasury) creates self-reinforcing loop | Revenue split and agent bonding design parameters | *** ## Pendle PT/YT Integration (D-051) | Paper / Source | Authors | Venue | Key Finding | Application | | ----------------------------------- | -------------- | ---------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------- | | Pendle V2 yield tokenization | Pendle Finance | Production ($11B TVL peak, 58% market share) | PT/YT mechanism splits yield-bearing assets into fixed-rate Principal Tokens and variable Yield Tokens via Standardized Yield (SY) wrappers | PendleAdapter.sol core PT/YT interaction pattern | | Term Structure PT looping strategy | Term Structure | Production tutorial (2025) | PT-sUSDe at 15% APY as collateral → borrow at 10% → reinvest → 3.1x leverage boosts yield from 15% to 26% | Fixed-rate capture strategy with validated leverage targets | | Pendle Prime permissionless listing | Pendle Finance | Production (Mar 2025) | Anyone can list assets without team approval; programmatic listing enables automated vault share tokenization | Auto-listing vault share tokens for PT/YT market creation | | Pendle Boros V3 | Pendle Finance | Production on Arbitrum (2025); not yet on Base | Funding rate trading via tokenized Yield Units; 5.98-11.4% fixed APR on cross-exchange funding arbitrage | Deferred until Boros deploys on Base; no cross-chain dependency in v1 | *** ## Credit Delegation (D-052) | Paper / Source | Authors | Venue | Key Finding | Application | | ---------------------------------------- | ------------------------------- | ------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------- | | Aave V3 credit delegation | Aave | Production docs | `approveDelegation()` + `borrow(onBehalfOf)` enables uncollateralized borrowing; delegatee cannot abuse to liquidate delegator (health factor check) | CreditDelegationAdapter.sol core delegation mechanics | | ERC-8004 reputation as trust signal | Crapis, De Rossi, Ellis, Reppel | EIP mainnet Jan 2026 | On-chain reputation registry with feedback/summary interface; tiered trust levels | Reputation score as delegation eligibility gate (minReputationScore = 100) | | "Reputation-weighted collateral" concept | Research | `research/defi-primitives2.md` | Trust progression: new 150% → 6mo clean 120% → 12mo audited 100% → institutional 80% with insurance | Collateral tier reduction schedule tied to reputation score | *** ## Policy Cage Architecture (D-053) | Paper / Source | Authors | Venue | Key Finding | Application | | -------------------------------------------------- | ------------- | --------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------- | | Theoriq AlphaVault | Theoriq | Production (Dec 2025) | Smart contract "policy cages" define exactly what AI agents can do; 65M+ agent requests across 2.1M wallets; $THQ reputation token with slashing | PolicyCage.sol hard boundary pattern | | Yearn V3 vault factory | Yearn Finance | Production ($447-562M TVL) | Role-based access (ADD\_STRATEGY\_MANAGER, DEBT\_MANAGER, REPORTING\_MANAGER, EMERGENCY\_MANAGER); Debt Allocator with APR oracles for automated strategy rotation; VaultFactory via create2 | Role mapping and on-chain constraint enforcement | | "Efficient LP with Deep Reinforcement Learning" | Xu, Brini | arXiv:2501.07508 (Jan 2025) | PPO agent dynamically adjusts concentrated liquidity; outperforms heuristics in 7/11 out-of-sample windows; rolling window for regime shifts | RL overlay operating within policy cage constraints | | "TD3-BC for Aave V2/V3 Interest Rate Optimization" | Qu et al. | arXiv:2506.00505 (Jun 2025) | TD3-BC handles historical stress events including May 2021 crash | Validates RL for lending rate optimization within policy boundaries | | Numerai tournament model | Numerai | Production (2017+) | Stake-weighted ensemble allocation across competing data scientist models; underperformers slashed | Deferred Numerai-inspired competing strategy extension | *** ## Cross-Vault Liquidity Router (D-054) | Paper / Source | Authors | Venue | Key Finding | Application | | ----------------------------------------------------------------- | ----------------- | ------------------------------------------ | ----------------------------------------------------------------------------------------------------- | --------------------------------------------------------------- | | Morpho Blue shared liquidity | Morpho Labs | Production ($13B deposits in 2025) | Shared liquidity across markets; utilization-curve interest rate model | Design precedent for inter-vault lending rate model | | Fluid Protocol shared liquidity layer | Fluid (Instadapp) | Production (2025) | Cross-protocol shared liquidity reducing idle capital waste | Design precedent for factory-level capital sharing | | "Determinants of Funding Liquidity Risk in Decentralized Lending" | Nguyen et al. | Global Finance Journal, Vol. 64 (Mar 2025) | Fragmented deposit bases increase run risk; rate-limiting mechanisms reduce strategic complementarity | Motivates `maxBorrowBps` (20%) and `minIdleReserve` constraints | *** ## BondMM Fixed-Rate Lending Hook (D-055) | Paper / Source | Authors | Venue | Key Finding | Application | | --------------------------------------------------------------------------------------------- | -------------- | --------------------------------- | --------------------------------------------------------------------------------------------------------------------- | ------------------------------------------ | | "Design of a Decentralized Fixed-Income Lending AMM Protocol Supporting Arbitrary Maturities" | (UMass Boston) | arXiv:2512.16080 (Dec 2025) | Single liquidity pool for multiple maturities; superior capital efficiency; better price impact for borrowers/lenders | BondMMHook.sol V4 hook architecture | | BondMM paper | (UMass Boston) | cs.umb.edu/\~duc/research/bondmm/ | Foundational BondMM design; flexible maturity dates through one pool; improved financial stability | Original fixed-income AMM design reference | *** ## Production Protocol References | Protocol/System | Relevance | Application | | -------------------------------------- | ---------------------------------------------------------------------------------------------- | -------------------------------------- | | **Doppler** (Whetstone Research, 2025) | V4 hooks for Dutch auction dynamic bonding curves; three slug types; post-auction LP migration | CCA + V4 pool lifecycle model | | **Angstrom** (Sorella Labs, 2025) | App-Specific Sequencer via V4 hooks; dual auction per block; MEV→LPs; EigenLayer enforcement | MEV-protected trading for vault pools | | **Bunni v2** | am-AMM production implementation; 59% of V4 hook volume; BidDog open-source reference | StrategyAuctionModule direct precedent | | **ClankerHook** | 164K+ auto-created pools; $3.1B+ volume; 2-block anti-snipe | NAVAwareHook + LaunchFeeHook precedent | | **Morpho** | $5.8B TVL; permissionless vault factory; four-role governance; adapter architecture | Factory + governance model precedent | | **CCA / Aztec** | $59M raised from 17,000 bidders across 191 countries; zero detected sniping | Production CCA validation at scale | *** ## NAV Guardrail System (D-062) | Paper / Source | Authors | Venue | Key Finding | Application | | --------------------------------- | -------------------------------------- | -------------------- | ------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------ | | LVR research (composite) | Milionis, Moallemi, Roughgarden et al. | Multiple (2022-2025) | NAV-priced markets using on-chain spot state are vulnerable to flash-loan manipulation; TWAPs and discrete snapshots reduce attack surface | Discrete NAV snapshots instead of live `convertToAssets()` | | Yearn V3 profit locking/unlocking | Yearn Finance | Production | Profit max unlock time smooths share price jumps; instantaneous unlock creates sandwich incentives | Rate-of-change clamp extends profit unlock to NAV pricing boundary | | ERC-7265 circuit breaker pattern | Community standard | ERC draft | Rate-limited token outflows prevent rapid drain; would have prevented $195M Euler exploit | Market safety mode integrates with ERC-7265 interface | | "Flash Boys 2.0" | Daian et al. | IEEE S\&P 2020 | Flash loans enable enormous temporary positions for oracle manipulation within single transactions | No-spot-assumptions policy (D-067); staleness gates | ## Risk-Adjusted Strategy Scoring Framework (D-063) | Paper / Source | Authors | Venue | Key Finding | Application | | ------------------------------------ | ------------------------------------------------------------------------------- | ------------------------------------- | -------------------------------------------------------------------------------------------- | -------------------------------------------------- | | LVR-theta perpetual option framework | Singh et al. | arXiv:2508.02971 (Aug 2025, AFT 2025) | LVR = theta of a perpetual continuous-installment put option; provides fee floor calibration | LVR penalty component in risk scoring function | | Fee-implied volatility oracle | Bichuch, Feinstein | arXiv (Sep 2025) | Bijective mapping between swap price and implied volatility; zero-cost on-chain vol oracle | Volatility input for risk penalty computation | | Dynamic fee consensus (three-paper) | Campbell-Bergault-Milionis-Nutz; Baggiani et al.; Milionis-Moallemi-Roughgarden | arXiv (2023-2025) | Threshold-type dynamic fees approximately optimal; closed-form solutions | Fee regime inputs for yield estimation | | Morpho vault risk framework | Morpho Labs | Production docs | Curator/allocator/guardian roles; exposure caps; adapter isolation | Governance model for scoring framework enforcement | ## Intent-Based Rebalance Auction (D-064) | Paper / Source | Authors | Venue | Key Finding | Application | | ------------------------- | ---------------- | --------------------------- | ----------------------------------------------------------------------------------------- | ------------------------------------------------------------------- | | UniswapX intent auction | Uniswap Labs | Production (2024) | Signed intent + filler competition + on-chain settlement constraints; MEV internalization | Solver competition model for vault rebalances | | "Batching Trades on AMMs" | Canidio, Fritsch | arXiv:2307.02074 (AFT 2023) | Batch clearing eliminates sandwich attacks; LP returns slightly exceed Uniswap v3 | Fallback to FM-AMM batch clearing when auction has no valid winners | | "Flash Boys 2.0" | Daian et al. | IEEE S\&P 2020 | Vault rebalances create predictable MEV targets; bot competition extracts value | Motivation: rebalances need MEV-aware execution | ## Three-Lane Key Architecture (D-065) | Paper / Source | Authors | Venue | Key Finding | Application | | ----------------------- | ---------------------- | -------------- | ----------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------- | | Battering RAM | Van Bulck et al. | IEEE S\&P 2026 | <$50 memory interposer breaks Intel TDX, AMD SEV-SNP, and NVIDIA CC | Hot lane (TEE) is defense-in-depth, not defense-in-total | | BadRAM | De Meulemeester et al. | IEEE S\&P 2025 | <$10 SPD chip tampering breaks AMD SEV-SNP; some DIMMs allow software-only attacks | Reinforces need for cold lane (elevated owner authority) and fail-safe lane (cancel-only) | | Safe Transaction Guards | Safe (Gnosis) | Production | Programmable restrictions on M-of-N control; guard validates every tx before execution | Cold lane implementation pattern | | ERC-7579 SmartSessions | Rhinestone, Biconomy | Production | Session key modules with composable hooks (spending limits, allowed targets, time bounds) | Hot lane session key scoping | ## Canonical Event Schema + Health Attestations (D-066) | Paper / Source | Authors | Venue | Key Finding | Application | | ----------------------------------------------------------- | ------------------ | ---------------- | -------------------------------------------------------------------------------------- | ----------------------------------------------------------------- | | ERC-7265 circuit breaker interface | Community standard | ERC draft | Standardized rate-limited token outflow pattern; composable across protocols | Health attestation interface design | | Trail of Bits "Building Secure Contracts" | Trail of Bits | Handbook (2024) | Standard event schemas and invariant monitoring reduce mean time to detect and respond | Canonical event taxonomy design | | "Systemic Risk in DeFi: A Network-Based Fragility Analysis" | arXiv:2601.08540 | arXiv (Jan 2026) | DeFi Correlation Fragility Indicator (CFI); cross-protocol correlation concentration | Health attestations enable factory-level systemic risk monitoring | ## ERC-7540 Emergency Exit UX (D-068) | Paper / Source | Authors | Venue | Key Finding | Application | | ----------------------------- | ------------------ | ------------------------ | -------------------------------------------------------------------------------------------- | --------------------------------------------------------------- | | Lido stETH withdrawal queue | Lido DAO | Production ($28B+ TVL) | FIFO withdrawal queue with NFT receipt + claim pattern; queue status visible on-chain | Withdrawal receipt NFT UX reference | | ERC-7540 async deposit/redeem | Community standard | ERC (Finalized Jun 2024) | request -> claim lifecycle for illiquid vault operations; `requestId` for batched settlement | Three-option exit UX (withdraw now / request exit / force exit) | | Centrifuge RWA vaults | Centrifuge | Production (\~$250M TVL) | ERC-7540 for institutional-grade async deposit/withdrawal with real-world asset settlement | Production ERC-7540 implementation reference | *** ## VIP1 Research Integration (Feb 2026) Citations added from VIP1 gap analysis (`research/VIP1.md`), organized by topic area. ### V4 Hook Security (Production Incidents) | Source | Details | Key Finding | Application | | ------------------------- | ----------------------- | ------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------- | | **Cork Protocol exploit** | May 28, 2025, $11M loss | Hook callback functions lacked `onlyPoolManager` access control; attacker called hook functions directly bypassing PoolManager | Canonical cautionary tale for V4 hook security requirement #2; referenced in 06-contracts.md VaultHook security note and 10-safety.md | ### ERC-4626 Testing & Verification | Source | Details | Key Finding | Application | | --------------------------------------- | ----------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------ | | **a16z/erc4626-tests** | github.com/a16z/ERC4626-tests | Property test suite validating round-trip properties (no free profit from deposit→withdraw), preview function accuracy, caller independence, and functional correctness. Solodit tracks 265 ERC-4626 audit findings (169 HIGH/MEDIUM) | Phase 5 mandatory gate: suite MUST pass for all vault templates before mainnet deployment (14-milestones.md) | | **OpenZeppelin virtual shares pattern** | OpenZeppelin Contracts v5.x | `_decimalsOffset()` override (recommended 3-6 for high-value vaults) prevents inflation attacks; combined with internal asset accounting | Already adopted via D-017; confirmed as gold standard by VIP1 | ### Agent Coordination Protocols | Source | Details | Key Finding | Application | | ------------------------------------------ | -------------------------------------- | ----------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------- | | **Virtuals ACP** (Agent Commerce Protocol) | 18K+ agents, $470M+ agentic GDP | Four-phase lifecycle (Request → Negotiation → Transaction → Evaluation) with smart contract escrow and evaluator agents | Agent coordination stack layer for strategy agreements (16-synergies.md Section 21.10); deferred Track D integration (D-071) | | **x402 protocol** (Cloudflare) | \~500K payments/week, as low as $0.001 | HTTP 402 "Payment Required" for machine-to-machine micropayments; ideal for data feeds, analytics, strategy signals | Micropayments layer in agent coordination stack (16-synergies.md Section 21.10) | ### Agent Security Threat Model | Source | Details | Key Finding | Application | | --------------------------- | ------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------- | | **AIXBT hack** | March 18, 2025, 55.5 ETH (\~$105K) lost | Agent infrastructure attacks target admin interfaces, not AI models. Attacker gained dashboard access and queued malicious replies triggering wallet fund sends. Wallet had no per-transaction spending limit. | Validates session key spending caps and TEE policy enforcement as mandatory (03-custody.md); motivates time-delayed proxy as primary defense | | **SCONE-bench** (Anthropic) | Claude Opus 4.5, 65% exploit success rate, $1.22 avg scanning cost | AI models autonomously exploit post-June 2025 contracts at 55-65% success rates. Exploit revenue doubles every 1.3 months; costs fall 22% every 2 months | Existential threat model for permissionless vault factory; validates defense-in-depth approach and formal verification requirements | ### Multi-Agent Cascading Failure | Source | Details | Key Finding | Application | | ----------------------- | ------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------- | | **Galileo AI research** | December 2025 | A single compromised agent can poison 87% of downstream decision-making within 4 hours in multi-agent systems. Real-world procurement agent compromise led to $3.2M in fraudulent orders | Cross-vault agent isolation requirements, anomaly-driven escalation (10-safety.md Multi-Agent Isolation section) | | **Lakera AI research** | November 2025 | Memory injection attacks corrupt agent long-term memory, creating persistent false beliefs about security policies. "Sleeper agent" attacks prime agents over weeks before trigger activation | Periodic policy re-verification from on-chain state every 6 hours; memory integrity hash verification (10-safety.md) | ### Base L2 Features | Source | Details | Key Finding | Application | | --------------- | ---------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------ | | **Flashblocks** | Live since July 2025 on Base | 200ms sub-blocks within 2-second blocks change event timing behavior; some developers report app breakage with Flashblocks-enabled RPC endpoints | Flashblocks compatibility testing required for all hook-based event handling (12-dependencies.md Base L2 Risk Acknowledgments) | ### MCP Protocol Extensions | Source | Details | Key Finding | Application | | ------------------------ | ------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- | | **MCP Elicitation spec** | MCP draft spec, introduced June 2025 | Servers can request structured user input mid-execution via JSON Schema (Form Mode) or redirect to external URLs (URL Mode); enables human-in-the-loop confirmation for state-changing operations | Write confirmation for vault operations above configurable threshold (08-mcp-tools.md Design Rationale); complements time-delayed proxy | ### Async Redemption Extensions | Source | Details | Key Finding | Application | | ------------ | ------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- | | **ERC-7887** | ERC Draft, February 2025 | Adds cancellation support (`cancelRedeem()`/`cancelDeposit()`) for pending ERC-7540 requests; allows users to exit long-dated positions before settlement | Deferred enhancement for CCA bids where agent strategy changes mid-auction (D-070, Track A) | *** ## Self-Improvement Loop (07-tools-advanced.md Self-Improvement Loop Tools) Papers informing the dual-timescale self-improvement architecture: verbal reflection (inner loop) and Bayesian parameter optimization (outer loop). | Paper / Source | Authors | Venue | Key Finding | Application | | ------------------------------------------------------------------------------------- | --------------- | --------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------- | | "CryptoTrade: A Reflective LLM-based Agent to Guide Zero-shot Cryptocurrency Trading" | Li, Chen et al. | EMNLP 2024 | Verbal reflection after each trade: LLM reviews recent performance, identifies what worked/failed, generates strategic guidance that feeds into subsequent decisions. In-context learning without model weight updates. Outperforms baselines on BTC, ETH, SOL, DOGE | Inner loop (per-execution) pattern for `record_execution` → `compare_predicted_vs_actual` → `get_failure_patterns` cycle | | "Preference-based Bayesian Optimization (PRBO)" | Liu, Cartlidge | arXiv:2208.02901 | "Bandit-over-bandit" approach: inner Gaussian Process tunes continuous parameters (tick ranges, rebalance thresholds, fee sensitivity) while outer multi-armed bandit handles non-stationarity by detecting regime changes and resetting optimization bounds | Outer loop (periodic) pattern for `tune_parameters` tool; handles market regime shifts via rolling-window optimization | | "SAFLA: Self-Adaptive Feedback Loop Architecture" | — | MCP-compatible framework | MCP-compatible feedback loop with safety constraints: hard limits on maximum parameter change per iteration, minimum diversification, maximum position sizes. Prevents runaway self-modification. | Safety constraints in `tune_parameters`: max 10% parameter change per iteration, bounded drift, risk constraint verification | | "Efficient LP with Deep Reinforcement Learning" | Xu, Brini | arXiv:2501.07508 (Jan 2025) | PPO-trained agent outperforms heuristic passive LP strategies in 7 of 11 out-of-sample windows using rolling-window training. Rolling windows essential — strategies from past regimes fail in new conditions. | Validates rolling-window approach for `tune_parameters`; regime-detection component critical for LP optimization | | "Soft Actor-Critic Entropy Regularization for Crypto Volatility" | — | Nature (2026) | SAC RL agents with entropy regularization adapt most effectively to changing crypto volatility regimes | Regime-adaptive strategy optimization for `classify_regime` → `tune_parameters` pipeline | *** ## Memory Architecture Foundations (shared/memory-architecture.md §1) Papers informing the cognitive architecture design, tiered memory, and virtual context management of the DeFi Brain. | Paper | Authors | Venue | Key Finding | Application | | ----------------------------------------------------------------------- | ------------------ | ------------------------------- | -------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------ | | "Cognitive Architectures for Language Agents (CoALA)" | Sumers et al. | arXiv:2309.02427; TMLR 2024 | Formalizes LLM agents as having working, episodic, semantic, and procedural memory — mapping to classical SOAR/ACT-R | Four-tier memory design: LRU cache, LanceDB episodes, SQLite insights, skill library | | "MemGPT: Towards LLMs as Operating Systems" | Packer et al. | arXiv:2310.08560; ICLR 2024 | OS-inspired virtual context management; agent self-directs memory reads/writes via interrupt-driven control flow | Context injection middleware pattern; tiered archival/recall storage | | "Bridging SOAR Design Patterns to Modern Agentic LLM Systems" | Wray, Kirk & Laird | arXiv:2505.07087 (May 2025) | Observe-decide-act loops, knowledge compilation, and metacognitive reflection have direct LLM-agent analogs | Validates Reflexion/ExpeL as modern instantiations of classical cognitive patterns | | "A Survey on the Memory Mechanism of Large Language Model Based Agents" | Zhang et al. | arXiv:2404.13501; ACM TOIS 2024 | Catalogs memory mechanisms across sources, forms, and operations; provides design taxonomy | Taxonomy for DeFi Brain memory subsystem design | | "A-MEM: Agentic Memory for LLM Agents" | Xu et al. | arXiv:2502.12110; NeurIPS 2025 | Zettelkasten-inspired agentic memory with dynamic interconnected knowledge networks | Cross-episode insight linking in consolidation loop | | "FadeMem: Forgetting-Aware Memory for Long-Horizon Agents" | — | 2025 | Ebbinghaus-curve decay achieves 45% storage reduction while maintaining superior retrieval performance | Validates DeFi Brain decay architecture | *** ## Retrieval-Augmented Generation Enhancements (shared/memory-architecture.md §2) Papers informing retrieval quality improvements, embedding upgrades, and hybrid search strategies. | Paper | Authors | Venue | Key Finding | Application | | --------------------------------------------------------------------------------------------------------------- | ---------------- | ------------------------------ | ------------------------------------------------------------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------- | | "HippoRAG: Neurobiologically Inspired Long-Term Memory for LLMs" | Gutierrez et al. | arXiv:2405.14831; NeurIPS 2024 | Knowledge graph + Personalized PageRank; 10-30x cheaper, 6-13x faster, 20% multi-hop QA improvement | Tier 3: associative cross-episode retrieval via `graphology` | | "HippoRAG 2: Towards Deeper Passage Integration" | Gutierrez et al. | arXiv:2502.14802; ICML 2025 | Deeper passage integration into knowledge graph index | Extension of HippoRAG with richer context | | "GraphRAG: From Local to Global" | Edge et al. | arXiv:2404.16130 | Entity-relationship graphs + Leiden community detection + hierarchical LLM summaries | Research context: informs graph-based retrieval design; community detection patterns applicable via `graphology` npm | | "RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval" | Sarthi et al. | arXiv:2401.18059; ICLR 2024 | Recursive clustering + summarization; multi-resolution retrieval; +20% absolute accuracy on QuALITY | Multi-resolution memory: trade-level to strategy-level | | "Self-RAG: Learning to Retrieve, Generate, and Critique" | Asai et al. | arXiv:2310.11511; ICLR 2024 | Reflection tokens enable LLM to decide when to retrieve and critique relevance; only 2% correct from unsupported sources | Tier 2: retrieval quality self-evaluation | | "Corrective Retrieval Augmented Generation (CRAG)" | Yan et al. | arXiv:2401.15884 | Lightweight retrieval evaluator triggers web search fallback when quality is "Incorrect" | Tier 2: fall back to live on-chain data when cached knowledge stale | | "Active Retrieval Augmented Generation (FLARE)" | Jiang et al. | arXiv:2305.06983; EMNLP 2023 | Forward-looking active retrieval using low-confidence tokens as queries | Multi-step DeFi reasoning with step-wise data refresh | | "MemoRAG: Moving Towards Next-Gen RAG via Memory-Inspired Knowledge Discovery" | Qian et al. | arXiv:2409.05591; WWW 2025 | Dual-system: lightweight LLM for global memory + stronger LLM for answers; outperforms GraphRAG | Alternative global memory architecture | | "SPLADE: Sparse Lexical and Expansion Model" | Formal et al. | arXiv:2107.05720; SIGIR 2021 | Learned sparse-dense hybrid retrieval with expansion | LanceDB native hybrid search captures most benefits without SPLADE inference | | "ColBERT v2: Effective and Efficient Retrieval via Lightweight Late Interaction" | Santhanam et al. | arXiv:2112.01488; NAACL 2022 | Late-interaction MaxSim scoring; cross-encoder quality at retrieval speed | No ONNX/Node.js path exists (Transformers.js Issue #851); deferred | | "Matryoshka Representation Learning (MRL)" | Kusupati et al. | arXiv:2205.13147; NeurIPS 2022 | First m dimensions independently useful; 64-dim fast scan + full rerank = 14x speedup | Tier 1: nomic-embed-text-v1.5 upgrade for adaptive-dimension retrieval | | "Contextual Retrieval" | Anthropic | Blog post, September 2024 | LLM-generated chunk context prepended before embedding; 49% retrieval failure reduction with hybrid BM25 | Tier 2: pre-embedding contextual enrichment | | "TSDAE: Using Transformer-based Sequential Denoising Auto-Encoder for Unsupervised Sentence Embedding Learning" | Wang et al. | arXiv:2104.06979 | Up to 8 points improvement on domain STS via unsupervised pre-training | Research context: if community-published DeFi ONNX models become available, load as drop-in replacement in Transformers.js | *** ## Self-Improvement and Lifelong Learning (shared/memory-architecture.md §3) Papers informing the Reflexion/ExpeL self-improvement engine, skill libraries, and planning. | Paper | Authors | Venue | Key Finding | Application | | ----------------------------------------------------------------------------- | --------------- | ---------------------------------------- | ------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------- | | "Reflexion: Language Agents with Verbal Reinforcement Learning" | Shinn et al. | arXiv:2303.11366; NeurIPS 2023 | Intra-task verbal self-reflection; 91% pass\@1 HumanEval; 97% on decision-making benchmarks | Inner loop: per-operation reflections in LanceDB episodic memory | | "ExpeL: LLM Agents Are Experiential Learners" | Zhao et al. | arXiv:2308.10144; AAAI 2024 | Cross-task insight extraction via ADD/UPVOTE/DOWNVOTE/EDIT on natural-language insight pool | Outer loop: consolidation of episodes into semantic insights | | "ReAct: Synergizing Reasoning and Acting in Language Models" | Yao et al. | arXiv:2210.03629; ICLR 2023 | Interleaves reasoning traces with actions; +34% on ALFWorld; no learning | Baseline agent loop pattern | | "Voyager: An Open-Ended Embodied Agent with Large Language Models" | Wang et al. | arXiv:2305.16291; TMLR 2024 | Ever-growing skill library indexed by embeddings; 15.3x faster capability unlock; zero-shot transfer | Tier 2: composable DeFi strategy library in LanceDB | | "Language Agent Tree Search (LATS)" | Zhou et al. | arXiv:2310.04406; ICML 2024 | MCTS for language agents; LLM as actor + value function; 92.7% pass\@1 HumanEval | MCTS-style planning for multi-step vault operations | | "Trial and Error: Exploration-Based Trajectory Optimization (ETO)" | Song et al. | arXiv:2403.02502; ACL 2024 | Contrastive trajectory optimization: SFT + exploration + DPO on failure-success pairs; >5% improvement | Learning from failed DeFi transactions via contrastive pairs | | "STaR: Bootstrapping Reasoning With Reasoning" | Zelikman et al. | arXiv:2203.14465; NeurIPS 2022 | Iterative rationale generation and filtering for self-improving reasoning | Bootstrapped reasoning for strategy evaluation | | "Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking" | Zelikman et al. | arXiv:2403.09629 | Extends STaR to generate internal rationales at every token position | Internal reasoning improvement | | "Thought Cloning: Learning to Think while Acting by Imitating Human Thinking" | Hu & Clune | arXiv:2306.00323; NeurIPS 2023 Spotlight | Clones both thoughts and actions; Precrime Intervention detects unsafe plans before execution | Precrime for dangerous DeFi transactions | *** ## DeFi-Specific ML and Quantitative Methods (shared/memory-architecture.md §4, new papers only) Papers informing RL market making, forecasting, regime switching, and portfolio optimization. Papers already cited elsewhere (LVR by Milionis, IL hedging by Lipton, BIS WP 1227) are not duplicated here. | Paper | Authors | Venue | Key Finding | Application | | -------------------------------------------------------------------------------------- | ------------------------------ | ---------------------------------------------------- | -------------------------------------------------------------------------------------------------- | ----------------------------------------------------------- | | "Market Making via Reinforcement Learning" | Spooner et al. | arXiv:1804.04216; AAMAS 2018 | TD RL with asymmetrically dampened PnL reward for market making | RL-based LP strategy within policy cage constraints | | "Reinforcement Learning for Market Making in Multi-Agent Dealer Market" | Ganesh et al. | arXiv:1911.05892 | Multi-agent RL where agent learns competitor pricing dynamics | Competitive vault strategy optimization | | "Temporal Fusion Transformers for Interpretable Multi-Horizon Time Series Forecasting" | Lim et al. | International Journal of Forecasting, 2021 | TFT architecture with variable selection, temporal attention, and quantile outputs | Multi-horizon crypto price and volatility forecasting | | "Temporal Fusion Transformers with On-Chain Features for Cryptocurrency Forecasting" | — | MDPI, 2025 | Integrates on-chain features (SOPR, TVL, active addresses, exchange net flow, HODL waves) into TFT | On-chain feature engineering for vault strategy signals | | "Adaptive TFT with Dynamic Segmentation" | Peik et al. | arXiv:2509.10542 | Dynamic segmentation significantly outperforms standard TFT/LSTM on 10-minute ETH-USDT | High-frequency vault rebalancing signal generation | | "Markov-Switching GARCH Models for Cryptocurrencies" | Ardia et al. | Research in International Business and Finance, 2019 | MS-GARCH with asymmetries outperforms 1,000+ standard GARCH variants for crypto VaR/ES | Regime-switching volatility estimation for fee optimization | | "HMM + RL Portfolio Management with Regime Detection" | — | Digital Finance, 2024 | Combines HMM regime detection with RL portfolio management using three volatility-defined regimes | `classify_regime` + adaptive strategy selection | | "CVaR-Constrained Deep Reinforcement Learning for Portfolio Optimization" | — | Economic Modelling, 2022 | CVaR-DRL outperforms traditional techniques on crypto portfolios with non-normal returns | Risk-constrained vault strategy optimization | | "Conditional Drawdown at Risk (CDaR)" | Chekhlov, Uryasev & Zabarankin | Quantitative Finance, 2005 | Controls path-dependent drawdowns via LP formulation; essential where rebalancing is costly | Drawdown management for gas-constrained DeFi vaults | | "Deep RL Portfolio Management: SAC and DDPG with LSTM" | — | arXiv:2511.20678, 2025 | SAC/DDPG with LSTM outperform Markowitz MPT on risk-adjusted metrics across BTC/ETH/LTC/DOGE | Multi-asset vault portfolio management | *** ## Causal and Temporal Reasoning for On-Chain Signals (shared/memory-architecture.md §5) Papers informing causal analysis of DeFi signals, temporal knowledge graphs, and continual learning. | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------------------------- | ------------------- | ------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------ | | "Granger Causality in DeFi: TVL, Volume, and Valuations" | — | MDPI International Journal of Financial Studies, 10(4), 2022 | TVL does not Granger-cause future valuations; bidirectional relationship between valuations and GMV | TVL as lagging indicator; focus memory on volume signals | | "Chainlets: Bitcoin Transaction Sub-Graphs and Causality" | Dixon et al. | Digital Finance, 2019 | Extreme chainlet activity Granger-causes intraday price volatility; GARCH + chainlets improves VaR | On-chain transaction pattern features for volatility prediction | | "Transfer Entropy for Cryptocurrency and Social Media" | Chalkiadakis et al. | Royal Society Open Science, 2021 | Transfer entropy detects nonlinear causality between sentiment and crypto returns beyond Granger | Nonlinear causal signals from social/governance events | | "CD-NOTS: Causal Discovery for Nonstationary Time Series" | Sadeghi et al. | arXiv:2312.17375 | Rejects stationarity assumptions invalidating standard methods in crypto; nonstationary causal discovery | Causal structure learning for DeFi protocol interactions | | "GPT-4 for Causal DAG Generation with Do-Calculus Pruning" | Sokolov et al. | SSRN:4679414 | GPT-4 generates initial causal DAGs; do-calculus rejects 39 of 89 edges against S\&P 500 data | Human-AI hybrid causal modeling for DeFi | | "Temporal Knowledge Graph Reasoning: A Survey" | Cai et al. | arXiv:2403.04782 | Covers TKG completion across embedding-based, GNN-based, and rule-based approaches | Tier 3: temporal knowledge graph architecture selection | | "MemoTime: Temporal Knowledge Graphs with Continual Memory" | Tan et al. | arXiv:2510.13614 | Tree of Time hierarchical decomposition; up to 24% improvement; Qwen3-4B matches GPT-4-Turbo | Tier 3: temporal reasoning for DeFi event sequences | | "GenTKG: Generative Forecasting on Temporal Knowledge Graphs" | Liao et al. | NAACL 2024 | RAG for TKG forecasting with temporal relevance scoring | Tier 3: predictive event modeling | | "Event-Driven Structured Memory for Financial Reasoning" | Zhou et al. | arXiv:2105.12825; ACL Findings 2021 | Structured (Actor, Action, Object) event tuples with causal links outperform sentiment-based approaches | Structured event storage in DeFi Brain | | "Continual Learning for Large Language Models: A Survey" | Shi et al. | arXiv:2404.16789; ACM Computing Surveys 2025 | Vertical (general->domain) and horizontal (across time) continuity; taxonomy of CL approaches | Framework for DeFi Brain knowledge evolution | | "Overcoming Catastrophic Forgetting (EWC)" | Kirkpatrick et al. | PNAS, 2017 | Elastic Weight Consolidation prevents forgetting via Fisher information penalty; 45.7% forgetting reduction on KG embeddings | Research context: informs Ebbinghaus decay design as retrieval-level forgetting mitigation (no model fine-tuning needed) | | "Self-Synthesized Rehearsal (SSR)" | Huang et al. | ACL 2024 | LLM generates synthetic rehearsal instances when original data unavailable | Maintaining memory coverage without original episode access | *** ## Production Reliability and Uncertainty Quantification (shared/memory-architecture.md §6) Papers informing circuit breakers, drift detection, conformal prediction, and multi-agent safety. | Paper | Authors | Venue | Key Finding | Application | | ----------------------------------------------------------- | -------------------- | ---------------------------------------------- | ------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------- | | "A Gentle Introduction to Conformal Prediction" | Angelopoulos & Bates | arXiv:2107.07511; FnT ML 2023 | Split CP, conformalized quantile regression, and time-series applications; distribution-free coverage guarantees | Tier 3: conformal prediction for position sizing | | "Adaptive Conformal Inference for Crypto VaR" | Fantazzini | Journal of Risk and Financial Management, 2024 | 4 ACI algorithms across 4,000 crypto-assets; FACI and SF-OGD provide precise VaR where GARCH fails | Tier 3: crypto-calibrated conformal prediction | | "Conformal Predictive Portfolio Selection" | Kato et al. | arXiv:2410.16333 | CP prediction intervals for portfolio selection at specified error rates | Tier 3: uncertainty-aware vault allocation | | "ConU: Conformal Uncertainty in LLMs" | Wang et al. | arXiv:2407.00499; EMNLP 2024 | Conformal prediction applied to black-box LLMs via self-consistency uncertainty measures | LLM confidence calibration for DeFi decisions | | "DeLLMa: Decision Making Under Uncertainty with LLMs" | Liu et al. | arXiv:2402.02392; ICLR 2025 | Identifies unknown states, elicits utility functions, maximizes expected utility; 40% accuracy improvement | Decision-making framework for vault operations | | "AgentSpec: Runtime Safety DSL for LLM Agents" | Poskitt et al. | ICSE 2026 | DSL for generating guardrail code from predicate functions; code-enforced invariants mandatory for financial agents | Validates PolicyCage over prompt-based safety constraints | | "Improving Factuality and Reasoning via Multi-Agent Debate" | Du et al. | arXiv:2305.14325; ICML 2024 | 3+ LLM instances debating over 2+ rounds significantly reduces hallucinations | Tier 2: multi-agent debate before high-value transactions | *** ## ERC-8004 Enhancements, A2A Protocol, and Yield Service Discovery (D-080 through D-088) > **Source**: `research/8004-enhancements.md` ### Standards and Protocols | Source | Details | Key Finding | Application | | ------------------------------ | ------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------- | | **ERC-8004: Trustless Agents** | De Rossi, Crapis, Ellis, Reppel -- EIP mainnet Jan 2026 | Three singleton registries (Identity, Reputation, Validation). Identity Registry stores `MetadataEntry` structs. Reputation Registry supports signed fixed-point feedback with tags and `feedbackURI`. | D-080, D-082, D-084, D-087: structured metadata tags, continuous feedback tracks, infrastructure identity, yield leaderboard | | **EIP-4361: SIWE** | Ethereum Foundation -- EIP Final | Sign-In with Ethereum. Structured message signing for wallet-based authentication. Domain binding, nonce replay prevention. | D-085: SIWE + JWT for remote MCP server authentication | | **A2A Protocol** | Google -- Linux Foundation, spec v0.1.0 | Agent-to-Agent protocol. JSON-RPC 2.0 over HTTP. Agent Cards, Tasks with state machines, multi-turn dialogue. Complementary to MCP. | D-083, D-086: A2A for complex negotiations, Agent Card in ERC-8004 registration file | | **OASF Taxonomy** | agent0lab -- GitHub | Open Agent Skills Framework. Hierarchical taxonomy for agent skills and domains. Used in ERC-8004 registration files for discovery. | D-080, D-086: OASF classification for vault manager discovery | | **MCP Authorization Spec** | Anthropic -- MCP draft spec | OAuth 2.1 as standard authentication. STDIO transports should not use OAuth. | D-085: SIWE layered on OAuth 2.1 for crypto-native agents | ### SDKs and Tools | Source | Details | Key Finding | Application | | ------------------------------ | -------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------- | | **Agent0 SDK (`@agent0/sdk`)** | De Rossi et al., MIT license | TypeScript SDK for ERC-8004. Identity registration, metadata management, MCP/A2A endpoint advertising, OASF classification, feedback submission, subgraph-powered agent discovery. Alpha version. | D-080: canonical ERC-8004 implementation | | **Privy Agentic Wallets** | Privy -- `@privy-io/node` | `createViemAccount` produces LocalAccount backed by secure enclave. P-256 authorization key pattern. EIP-712 support for ERC-8004 wallet verification. | D-081, D-088: Privy wallet integration for Agent0 SDK | | **8004scan Explorer** | agent0lab -- | On-chain explorer for ERC-8004 agents. Verification and discovery interface. | Discovery UX for registered agents | ### Integrations and Ecosystem | Source | Details | Key Finding | Application | | ------------------------------------ | ------------------------------------- | -------------------------------------------------------------------------------------------------- | --------------------------------------------------------- | | **DefiLlama Yields API** | DefiLlama -- `/yields/pools` endpoint | 12,845+ pools across 489 protocols. Comprehensive yield benchmarking. Pro API for historical data. | Yield service discovery: external benchmark integration | | **Superform Universal Yield** | Superform -- superform.xyz | Permissionless ERC-4626 vault listing. Cross-chain distribution via LayerZero/Hyperlane/Wormhole. | Yield service discovery: cross-chain vault distribution | | **A2A + Ledger-Anchored Identities** | arXiv:2507.19550 | Combining A2A with on-chain identity and x402 micropayments for agent commerce. | D-083, D-085: A2A + ERC-8004 + x402 integration rationale | | **Google AP2** | Google Cloud blog | Agent Payments Protocol for agent-to-agent payments. | Ecosystem context for agent commerce infrastructure | *** ## Agent Learning & Strategy Economy (17-learning-economy.md) Papers informing the episode capture system, Reflexion/ExpeL loops, ZKML verification, federated learning, and strategy marketplace design. ### RL for Uniswap LP | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------------ | ------------- | ------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------- | | "Efficient LP with Deep Reinforcement Learning" | Xu, Brini | AAAI 2025 Workshop; arXiv:2501.07508 | PPO agent dynamically adjusts concentrated liquidity ranges; outperforms heuristics in 7 of 11 out-of-sample windows; rolling window essential for regime shifts | Validates RL-based LP strategy learning within PolicyCage; backbone of strategy-optimizer outer loop | | "DDQN with Dueling Architecture for ETH/USDC LP" | Zhang et al. | arXiv (2023) | Dueling DDQN achieves superior risk-adjusted returns vs. passive LP strategies on ETH/USDC; action space is LP range adjustment | Baseline RL architecture for lp\_optimization strategy family | | "Mamba + Dueling DDQN for Sequential DeFi Data" | Multiple | arXiv (2025) | Mamba state-space model outperforms LSTM for sequential DeFi data due to selective state retention; combined with dueling DDQN for LP optimization | SSM as alternative to LSTM for episode sequence modeling | | "PPO with Sharpe Reward for Crypto (2020–2024)" | Bandarupalli | arXiv (2025) | PPO with Sharpe ratio reward achieves 1.23 Sharpe on 4-year backtest across BTC, ETH, and altcoins (2020–2024) | Validates Sharpe-as-reward for vault strategy optimization | | "FQL: Federated Q-Learning for DeFi" | Multiple | MDPI Electronics (2025) | Federated Q-learning framework achieves 12.72% annualized return, Sharpe 1.12, 10.3% max drawdown on crypto portfolios | Validates federated learning for cross-vault strategy improvement | | "PER + HER: Extended Trading DQN" | Takara et al. | arXiv (2024) | Prioritized Experience Replay + Hindsight Experience Replay: 1.46×–7.13× improvement over vanilla DQN for crypto trading | PER for episode prioritization in ExpeL outer loop | ### Data Marketplace & Knowledge Economics | Paper | Authors | Venue | Key Finding | Application | | --------------------------------------------------------------------- | ---------------- | ----------------- | -------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------- | | "Model-Based Pricing for Machine Learning" | Chen et al. | SIGMOD (2019) | Gaussian noise injection creates arbitrage-free lower pricing tiers; provides formal basis for differential-privacy-compatible tiered data pricing | Lower price tiers for basic signals; higher tiers for full strategy params | | "Dealer: End-to-End Data Marketplace with Shapley Value Compensation" | Multiple | VLDB (2021) | Shapley value compensation for data contributors; privacy sensitivity adjustment; incentive-compatible payment for data quality | Shapley-weighted federated regime detector (§9); contribution weight computation | | "Alpha Decay in Systematic Strategies" | Maven Securities | Industry research | 9.9% alpha decay per unit of delay in European markets; 5.6% in US markets; intraday signals decay within hours | Calibration baseline for DECAY\_CONSTANTS table in §11 | | "Commit-Reveal² for Privacy-Preserving Dispute Resolution" | Multiple | arXiv:2504.03936 | 80% gas savings over on-chain baseline for commit-reveal voting; preserves vote privacy until resolution | StrategyEscrow dispute resolution mechanism | ### Federated Learning | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------------------------------ | ------------------ | ----------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | | "FPPO: Federated Proximal Policy Optimization for Trade Execution" | Multiple | arXiv (2025) | Federated PPO for trade execution achieves 3–15% return improvement vs. isolated training; agents benefit from shared market regime knowledge without sharing positions | Validates federated regime detection as first collaborative feature | | "MHESA: Multi-Key Homomorphic Encryption for Secure Aggregation" | Multiple | MDPI Electronics (2025) | Random masking + multi-key CKKS achieves adequate accuracy in only 2 communication rounds; enables federated learning without trusted aggregator | Future federated strategy aggregation beyond regime vectors | | "GTG-Shapley: Cumulative Shapley Values for Federated Learning" | Multiple | ICML (2022) | GTG-Shapley over cumulative rounds is more accurate than per-round Shapley for contribution evaluation in non-i.i.d. settings | Shapley-weighted federated detector; contribution weight persistence across rounds | | "Knowledge Distillation for Privacy Amplification" | Tsantekidis et al. | Neural Networks (2021) | Knowledge distillation from ensemble to single model provides privacy amplification; distilled model leaks less about individual contributors | Future: distilling federated strategy ensemble to shareable compact model | ### Regime Detection | Paper | Authors | Venue | Key Finding | Application | | ------------------------------------------------- | ----------- | --------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------ | | "4-State HMM for Cryptocurrency Regime Detection" | Koki et al. | Multiple (2024) | 4-state Hidden Markov Model (bull/bear × low/high volatility) achieves 60%+ directional accuracy on ETH price series; outperforms 2-state and 3-state variants | Regime classifier in VaultStrategy (§5); federated regime detection (§9) | | "Generalizable Episodic Memory for Deep RL" | Hu et al. | NeurIPS (2021) | k-NN retrieval over episodic memory enables faster learning in new environments; geometric mean of stored returns as value estimate | EpisodeStore semantic retrieval design; k-NN over embedded episodes | ### ZKML | Paper / Source | Details | Key Finding | Application | | ------------------------ | ------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------ | | **EZKL library** | Trail of Bits audit (2024); @ezkljs/engine npm package | Only production-ready TypeScript/WASM binding for ONNX→Halo2 proof generation; 521–671× faster verification vs. proof generation; \~0.5 second on-chain verification | TypeScript-native ZKML verification layer (§8); tiered performance attestation | | **Halo2 proving system** | zcash/halo2 (2021+) | Recursive zk-SNARK without trusted setup; EVM-verifiable via exported Solidity verifier contract | `verifyPerformanceProof()` via deployed Halo2 verifier on Base |