ELVARA: an intent-based protocol for risk-aware portfolio optimization of a real-world asset proxy basket

1 Introduction

Tokenized real-world assets (RWAs) are increasingly positioned as a bridge between conventional asset ownership and blockchain-based financial infrastructures. In broad terms, tokenization converts rights in a real or financial asset into a digital representation that can be recorded, transferred, and managed on blockchain rails, with the promise of improved transparency, programmability, and market accessibility (Davé et al., 2026; Ballesteros-Rodríguez et al., 2024). Recent work suggests that tokenization is no longer discussed only as a theoretical concept, but as a practical mechanism for representing asset claims across sectors such as carbon markets, securitized assets, and real estate related structures (Ballesteros-Rodríguez et al., 2024; Yang et al., 2025).

Despite this momentum, much of the current RWA literature focuses on issuance, legal design, custody, transparency, compliance, and market structure, rather than on how allocation decisions should be generated and represented in blockchain-native portfolio systems. At the same time, the portfolio optimization literature has long established that investors should evaluate assets at the portfolio level rather than in isolation, beginning with Modern Portfolio Theory (Markowitz, 1952). Subsequent research extended this logic toward downside-risk-aware formulations, for instance, semivariance, conditional value at risk (CVaR), and entropic value at risk (EVaR), which are particularly relevant when asset classes exhibit asymmetry, tail risk, or regime sensitivity (Estrada, 2008; Rockafellar and Uryasev, 2000; Ahmadi-Javid, 2012). Yet these optimization approaches are usually studied as purely financial outputs, ending with a weight vector or rebalancing recommendation, without asking how such outputs should be published, verified, or coordinated in blockchain environments.

A parallel strand of research has examined how blockchain systems can reduce dependence on centralized trust intermediaries in markets for real assets. For example, a blockchain-based mechanism for real-world asset exchange using vehicle history data to mitigate information asymmetry in used-car markets (Notheisen et al., 2017) was proposed earlier. The work was important because it demonstrates that blockchain can support tamper-resistant asset records and partially automate asset transfer logic in markets where trust and provenance are central concerns. More broadly, research in decentralized finance has demonstrated the importance of programmable, machine-readable coordination rules for financial activity, including lending, exchange, collateralization, and adaptive execution (Jensen et al., 2021; Shah et al., 2023; Puschmann and Huang-Sui, 2024).

However, these strands of literature have largely evolved separately. RWA studies often stop at token issuance or market design; portfolio optimization studies stop at allocation; and blockchain protocol studies typically emphasize execution, governance, or infrastructure. This leaves a clear gap at the interface between portfolio analytics and verifiable on-chain publication. In the RWA setting, this gap is especially important because real-asset exposure often depends on different instruments, off-chain valuation, compliance constraints, and operational processes that may not map cleanly to immediate on-chain settlement. In such an environment, the portfolio decision itself should be publishable as a transparent, auditable, and machine-readable declaration of the desired future state, even when execution is deferred, partial, or handled by external systems.

This paper addresses that gap through ELVARA, an intent-based protocol for risk-aware portfolio optimization of tokenized real-world assets. Rather than treating portfolio optimization as the final analytical step, ELVARA treats the optimized rebalance as an intent: a structured declaration of target portfolio weights, validity conditions, metadata, and integrity constraints that can be deterministically generated off-chain and verifiably published on-chain. The protocol therefore links portfolio optimization with blockchain publication without requiring a full execution or settlement layer.

Empirically, the paper evaluates this architecture through a reproducible single-chain proof of concept using a seven-asset RWA proxy basket consisting of gold, silver, oil, real estate, U.S. Treasuries, energy infrastructure, and agriculture. The optimization framework uses risk-adjusted portfolio selection over an RWA-oriented ETF proxy basket while excess-return evaluation is referenced to a separate short-term U.S. Treasury benchmark rather than to the investable Treasury proxy itself. The use of ETF proxies is to preserve exposure to tokenizable real-asset sectors while maintaining sufficient return history and market variability for robust walk-forward testing. In addition, equal weighting is compared against several risk-aware allocation methods, including variance, semivariance, CVaR, EVaR, worst realization, and maximum drawdown. The results show that optimizer-driven intents can provide meaningful portfolio improvements over a simple baseline while remaining compatible with a minimal and auditable blockchain publication workflow.

This study designs and evaluates a reproducible pipeline that links risk-aware portfolio optimization for an RWA-oriented proxy basket to verifiable on-chain intent publication. It has three objectives. The first is to formalize optimizer-driven rebalancing as a deterministic, machine-readable intent. The second is to build and validate a minimal single-chain registry that enforces authorization, ordering, and integrity on published intents. The third is to compare risk-aware allocation rules against an equal-weight benchmark in a fixed walk-forward design. The hypothesis is that, under a fixed and fully reproducible specification, risk-aware optimized intents outperform the equal-weight baseline while remaining compatible with deterministic on-chain publication.

The paper makes four contributions. First, this work frames RWA portfolio rebalancing as an intent-generation problem rather than a pure execution problem. Second, it proposes a deterministic machine-readable artifact that links off-chain optimization to on-chain publication. Third, it implements a minimal single-chain architecture combining an off-chain optimizer with an on-chain registry. Finally, it evaluates the resulting system in a reproducible walk-forward setting, where CVaR outperforms the equal-weight benchmark under the fixed baseline configuration.

2 Related work

2.1 Tokenization of real-world assets

The literature on tokenization generally frames blockchain as infrastructure for digitally representing rights in real or financial assets. In this view, tokenization supports more transparent ownership records, programmable transfer rules, and new forms of market access through divisible and transferable token claims (Xia et al., 2025). Recent work has extended this discussion across multiple application areas, including carbon markets, patent-linked structures, and asset-backed securitization, illustrating that tokenization is increasingly studied as a mechanism for connecting off-chain value with on-chain coordination (Ballesteros-Rodríguez et al., 2024; Peters, 2025; Yang et al., 2025). Within this literature, the main benefits of RWA tokenization are usually described in terms of improved liquidity, fractionalization, transparency, and operational efficiency. At the same time, the literature repeatedly notes practical frictions, including legal enforceability, custody design, valuation reliability, oracle dependence, and the mismatch between on-chain representations and off-chain rights. These issues are particularly visible in real estate and other high-value asset classes, where the token does not automatically remove the legal and operational complexity of the underlying asset, a point emphasized in both conceptual and implementation-oriented discussions of real-estate tokenization (Joshi and Choudhury, 2022; Konashevych, 2020). As a result, tokenization should not be understood as a purely technical transformation, it is an institutional and informational transformation that must coordinate claims, records, compliance, and market access.

For this reason, existing RWA research is strong on representation and market structure, but weaker on portfolio decision-making. That is, prior work often explains how an asset can be tokenized, traded, or governed, yet gives less attention to how a portfolio of various tokenized or tokenizable assets should be optimized over time under explicit risk objectives. ELVARA builds on this literature by focusing not only on issuance, but also on how multi-asset RWA exposure can be translated into optimizer-generated portfolio intents.

2.2 Blockchain-based trust minimization in real-asset markets

A closely related stream of work studies how blockchain can reduce information asymmetry and dependence on centralized intermediaries in markets involving physical or legally grounded assets. Notheisen et al. (2017) provide an influential example by presenting a blockchain-based prototype for real-world asset transfer in the used-car market. Their design uses an immutable vehicle history to reduce “lemon market” frictions and improve trust in asset exchange. This contribution is important because it shows that blockchain systems can add value not merely by recording transfers, but by improving the credibility and accessibility of the information environment surrounding the asset.

Later blockchain studies support this idea. They show that blockchain can help record asset history, make records easier to check, reduce data tampering, and support trust between parties when information is spread across different systems. Evidence from provenance-oriented blockchain studies, particularly in supply-chain settings, shows how immutable records can strengthen verification, traceability, and counterpart trust when asset history matters to market participants (Cui et al., 2019; Montecchi et al., 2019). In the RWA context, these features matter because the quality of portfolio decisions depends not only on return and risk estimates, but also on the credibility of the asset representation itself. A tokenized asset with weak provenance, ambiguous redemption rights, or unreliable off-chain data is fundamentally different from one with strong record integrity and transparent governance.

Nevertheless, the trust-minimization literature mainly addresses asset-level credibility and transaction design, not portfolio-level intent formation. ELVARA extends this direction by treating the output of a portfolio optimizer as an auditable declaration that can itself become part of the trust layer. In other words, the contribution of our work is not to replace legal settlement, but to make the rebalancing decision transparent, reproducible, and verifiable.

2.3 Portfolio optimization and downside-risk modeling

The financial foundation for portfolio construction begins with Modern Portfolio Theory, which formalized the idea that investors should choose portfolios by considering expected return jointly with overall risk rather than evaluating each asset separately (Markowitz, 1952). This idea remains the starting point for systematic allocation. However, mean-variance optimization is often criticized for its sensitivity to estimation error and its limited treatment of downside and tail behavior, especially when returns are non-normal or exhibit crisis-driven asymmetry. Empirical analyses of information flow across global stock markets similarly document nonlinear directional dependence and long-memory effects that variance-based descriptions do not fully capture (Akgüller et al., 2025).

To address these limitations, later research proposed alternative risk measures and objective functions. Semivariance focuses on downside dispersion rather than total variance (Estrada, 2008). CVaR models expected loss in the tail beyond a value-at-risk threshold and has become one of the most widely used coherent downside-risk measures in portfolio selection (Rockafellar and Uryasev, 2000). EVaR further tightens tail-risk control through an entropy-based bound and provides useful convex optimization properties in certain settings (Ahmadi-Javid, 2012). Additional practical approaches evaluate worst realization or maximum drawdown to align optimization more closely with investor concern over severe interim losses rather than only endpoint volatility, with drawdown-based portfolio optimization providing a direct example of this perspective (Chekhlov et al., 2005). These measures are very relevant to RWA-oriented baskets because real-asset exposures can be affected by inflation, interest rates, commodity shocks and liquidity conditions. A portfolio framework that targets only average return or total variance may miss the properties most relevant to capital preservation, namely, asymmetric downside behavior and tail loss. For this reason, ELVARA incorporates CVaR, semivariance, EVaR, worst realization, and maximum drawdown as alternative objectives, formally defined in Section 3.3 and compared empirically in Sections 4.2 and 4.3.

In addition, most research on portfolio optimization stops at the level of the weight vector. They ask which allocation performs better, but not how that allocation should be translated into a verifiable blockchain-native object. ELVARA contributes precisely at this interface by coupling risk-aware optimization with deterministic intent publication.

2.4 DeFi protocols, machine-readable coordination, intent-based design, and DeFi security

Research on decentralized finance (DeFi) has shown that blockchain systems can support complex programmable financial activity without relying on conventional centralized operators. Lending protocols, decentralized exchanges, collateral systems, and automated market mechanisms all depend on machine-readable state transitions enforced by smart contracts (Jensen et al., 2021; Puschmann and Huang-Sui, 2024; Schär, 2021; Shah et al., 2023). More recent work also points toward adaptive and optimization-driven rule systems, suggesting that DeFi protocols can increasingly incorporate algorithmic decision layers rather than solely rely on static contract logic.

DeFi security research shows that smart-contract execution introduces major design risks that are directly relevant to financial system architecture. Vulnerabilities such as reentrancy, oracle manipulation, access-control failures, composability breakdowns, and flash-loan-assisted exploitation can undermine protocol logic, leading to substantial losses. Survey and synthesis studies therefore treat security not as a secondary implementation issue but as a core constraint on DeFi system design. Practitioner-oriented work likewise emphasizes auditing, testing, monitoring, and formal analysis as essential but still incomplete safeguards (Li et al., 2022; Wan et al., 2021; Zhou et al., 2023). Recent work on EVM transaction monitoring further highlights the importance of execution-layer observability by proposing an event-driven approach for flash-loan detection, reinforcing the need for security-aware infrastructure around DeFi-based coordination mechanisms (Ramezany et al., 2023).

Within this broader development, intent-based design has attracted attention because it separates what a user wants to achieve from how execution is carried out, a distinction analyzed explicitly in recent work on intent-based markets (Chitra et al., 2024). This separation is particularly useful when execution may involve routing, batching, auctioning, solver participation, or deferred fulfillment. Although intent-based systems are more commonly discussed in exchange or interoperability settings than in portfolio management (Belchior et al., 2021; Yuan et al., 2025), the architectural principle is highly relevant to RWA portfolios. Real-asset exposure often cannot be rebalanced through a single synchronous smart-contract action. Instead, execution may depend on off-chain valuation, custody windows, broker APIs, legal settlement, token issuance constraints, or compliance procedures that do not map cleanly to immediate on-chain settlement.

Accordingly, an RWA portfolio system benefits from an intermediate layer in which the desired target state is explicitly represented before execution. ELVARA adopts this logic by encoding optimized weights, asset identifiers, epoch metadata, expiry, and integrity information into a deterministic intent artifact. The on-chain layer does not execute trades directly. Instead, it validates and publishes the optimization output as a verifiable object that downstream systems may interpret or act upon. In security terms, this design choice is also important: by limiting the smart contract to validation and publication rather than full execution, ELVARA reduces on-chain complexity and narrows the immediate attack surface associated with trade logic, capital routing, and settlement orchestration. The protocol is therefore better understood as a publication, coordination, and control framework linking RWA-oriented portfolio optimization to verifiable on-chain intent publication, rather than as a complete on-chain trading venue.

2.5 Research gap

The literature therefore reveals a clear gap at three levels. Tokenization research explains how real-world claims can be represented on-chain but gives little attention to dynamic multi-asset portfolio construction. Portfolio optimization research provides risk-aware allocation methods but typically stops at the numerical weight vector. Blockchain protocol research addresses machine-readable coordination and trust minimization but rarely treats portfolio rebalancing as an intent-publication problem for RWA-oriented portfolios.

ELVARA is designed to fill this gap by combining a reproducible off-chain optimizer with a minimal on-chain registry that validates and publishes risk-aware portfolio intents. The protocol is therefore positioned at the intersection of RWA tokenization, portfolio optimization, and blockchain-native coordination. Its novelty lies in connecting tokenization, portfolio theory, and intents into a single reproducible pipeline for risk-aware RWA portfolio intent publication. The result is a system where a portfolio decision is not just computed but also normalized, committed, and independently verifiable, properties that matter increasingly as tokenized asset management moves from concept toward practice.

3 Materials and methods

3.1 Study design and ELVARA workflow

This study was designed as an end-to-end proof-of-concept for publishing risk-aware portfolio rebalancing intents in a single-chain setting. The central objective was to test whether a reproducible off-chain optimizer could be coupled with a minimal on-chain registry so that portfolio decisions are not only computed, but also normalized, hashed, persisted, and publicly committed in a machine-readable form. In contrast to conventional portfolio studies that stop at numerical allocation, ELVARA extends the research pipeline into an auditable publication layer for portfolio intents.

The ELVARA protocol was implemented as a three-layer architecture consisting of a research layer, an off-chain orchestration layer, and an on-chain registration layer. In the research layer, portfolio weights were generated from a fixed multi-asset pool using a walk-forward optimization procedure. The optimization logic was first developed in the research notebook and then exposed through a reusable Python module so that the analytical workflow and the operational pipeline remained consistent. In the off-chain orchestration layer, the selected allocation was transformed into a fixed intent artifact containing the epoch, strategy label, source window, creation timestamp, expiry timestamp, asset identifiers, continuous portfolio weights, basis-point targets, and summary performance metrics.

To ensure artifact integrity, the off-chain layer computed a deterministic hash over the normalized weights payload and a second hash over the broader metadata payload. These hashes provided compact digital fingerprints for verifying that the published intent corresponded to a specific portfolio configuration and research context. The resulting artifact was then persisted in a local storage layer and exposed through both command-line and API interfaces, allowing the same portfolio intent to be reproduced, retrieved, and relayed without manual modification.

The on-chain component was intentionally minimal. Limiting the contract to validation and publication, rather than full execution, reduces on-chain complexity and narrows the attack surface associated with trade logic, capital routing, and settlement orchestration. Rather than executing trades on-chain, the smart contract functioned as a registry for validating and publishing portfolio intents. Before accepting a submission, the contract enforced authorization controls, monotonic epoch progression, non-expired submissions, non-zero asset addresses, and the requirement that the target allocation sum to exactly 10,000 basis points. After validation, the contract recorded the latest epoch and intent hash and emitted an event containing the submitted allocation. This design preserved verifiability and traceability while avoiding the additional complexity of full on-chain execution.

ELVARA runs in two modes. In sample mode, the system reads the fixed optimization outputs from the research notebook, assembles them into a JSON artifact with locked metadata fields (epoch 1, created-at 2026–03-24T00, expiry timestamp of 1893456000), and saves that file to local storage. Because every input is fixed, the same artifact, metadata hash, and intent hash are produced on every run. This is the path used to generate the paper results. In live mode, the system re-downloads market data and recomputes allocations using the same optimization workflow, producing a fresh artifact tied to the current date. This mode is retained for robustness checks and real-world usage but is not used for the paper artifact. Figure 1 summarizes the ELVARA pipeline from market data and walk-forward optimization to deterministic intent generation, hashing, relay, and registry publication.

3.2 Data, asset pool, and walk-forward protocol

The empirical study was conducted on a fixed pool of seven assets constructed from exchange-traded fund proxies representing major real-asset and commodity-linked exposures. The selected assets were gold (GLD), silver (SLV), oil (USO), real estate investment trusts (VNQ), long-duration U.S. Treasuries (TLT), energy infrastructure (AMLP), and agriculture (DBA). This pool was chosen to provide cross-sector diversification across precious metals, energy, agriculture, fixed income, and income-generating real assets, while preserving consistent historical price availability. The full asset pool and its portfolio roles are summarized in Table 1.

TABLE 1. Asset pool, ETF proxies, and economic roles.

Historical market data were collected through Yahoo Finance using adjusted closing prices when available, with a fallback to closing prices where required. The raw market data download window spanned from 1 March 2021 to 3 March 2026. After alignment and removal of incomplete observations, the resulting return matrix contained 1,256 daily observations. Daily asset returns were computed from the cleaned price series and used as the input to all portfolio construction procedures. In parallel, the 13-week U.S. Treasury bill proxy (IRX) was downloaded over the same period and converted to an approximate daily risk-free rate, which was then used within the fold-specific optimization procedure.

To reduce look-ahead bias, the study adopted a walk-forward evaluation design rather than a single in-sample optimization. The baseline protocol used a minimum training window of 504 trading days, a test window of 63 trading days, and a rebalance step of 63 trading days, yielding 11 out-of-sample folds. Under this configuration, the out-of-sample evaluation period extended from 2 March 2023 to 3 December 2025. For each fold, the model was fitted on the training segment and evaluated only on the subsequent test segment. This design ensured that each reported allocation was generated using information available only up to the rebalance date. Because the protocol requires full 63-day test windows, the final 59 trading days of the dataset (extending from December 4, 2025, to March 3, 2026) constituted an incomplete fold and were intentionally excluded from the out-of-sample evaluation.

The portfolio construction process also incorporated practical allocation constraints. Asset weights were bounded between 5% and 35% in the baseline configuration, preventing degenerate solutions with extreme concentration or near-zero holdings in most assets. A proportional transaction cost of 10 basis points was applied to fold-level turnover to represent rebalancing frictions. Portfolio statistics were annualized, assuming 252 trading days. These settings were kept fixed throughout the main experiment so that differences in portfolio performance could be attributed primarily to the optimization method rather than to shifting protocol assumptions.

Because reproducibility was a central goal of the study, the baseline paper result was tied to a fixed historical window and a deterministic artifact generation path. The same underlying configuration was then reused in the off-chain intent pipeline, ensuring that the optimization outputs discussed in the manuscript and the portfolio intents published by the system were generated from the same research specification. The baseline data window, fold design, and implementation parameters are summarized in Table 2.

TABLE 2. Baseline walk-forward configuration parameters.

3.3 Optimization formulation and main algorithm

Portfolio construction was carried out in a walk-forward setting so that each rebalance decision was estimated only from information available up to the rebalance date. Let P t , i denote the price of asset i at time t. Daily simple returns were computed from the cleaned price series and used as the basic input to all subsequent optimization steps. The return of asset i on day t is given by Equation 1.

r t , i = P t , i P t − 1 , i − 1 (1)

The historical sample was partitioned into sequential walk-forward folds. For each fold k, the available observations were divided into a training window T k and a subsequent test window S k. The fold-specific risk-free rate was estimated as the mean daily Treasury bill rate over the training window. This produced a time-local estimate of the opportunity cost of capital rather than imposing a single constant risk-free rate over the full study period. The fold-specific daily risk-free rate is shown in Equation 2.where q t denotes the aligned daily Treasury bill rate and T k is the number of observations in the training segment of fold k.

r f , k = 1 T k ∑ t ∈ T k q t , (2)

For all optimized strategies except the equal-weight benchmark, the portfolio weights were obtained from a constrained mean-risk problem. The objective was to maximize a fold-level risk-adjusted return ratio under full-investment and box constraints. In this study, the risk term varied by strategy and included conditional value at risk (CVaR), semivariance, variance, entropic value at risk (EVaR), worst realization, and maximum drawdown. In implementation, the EVaR configuration used the CLARABEL solver with a constraint-scaling parameter of 10 − 1 , whereas the remaining optimized strategies used the default solver configuration. The generic fold-level optimization problem is shown in Equation 3.where μ k w denotes the fold-level expected portfolio return and ρ k w denotes the selected risk measure under the chosen strategy. The optimization is performed under full-investment and box constraints such that the portfolio weights sum to one, and each asset weight remains within the prescribed lower and upper bounds. In the baseline configuration, these bounds were fixed at 5% and 35%, respectively. For the equal-weight benchmark, each asset was assigned a weight of 1 N, where N denotes the number of assets, and the numerical optimizer was bypassed. To account for realistic rebalancing frictions, turnover in fold k was measured relative to the drifted portfolio entering the fold. Let w ∼ k − 1 denote the carried-forward portfolio weights from the previous test window, with w 0 ∼ = 0 for the first fold, and let λ denote the proportional transaction-cost rate. Fold-level turnover and the associated rebalancing cost were computed using Equations 4, 5, respectively.

w k * = arg max w μ k w − r f , k ρ k w , (3)

T O k = ∑ i = 1 N w k , i * − w ∼ k − 1 , i (4)

c k = λ T O k (5)

In the baseline experiment, λ = 10 10,000, corresponding to a transaction cost of 10 basis points. This cost was applied only once per fold, namely, to the first out-of-sample portfolio return immediately after rebalancing as shown in Equation 6.

r k , 1 n e t = 1 − c k 1 + r k , 1 − 1 (6)

After the test window, the realized asset returns induced portfolio drift. The resulting end-of-window weights were then carried into the next fold and used as the reference allocation for the subsequent turnover computation. The drifted portfolio weight of asset i at the end of fold k was computed using Equation 7.

w ∼ k , i = w k , i * ∏ t ∈ S k 1 + r t , i ∑ j = 1 N w k , j * ∏ t ∈ S k 1 + r t , j (7)

This update rule ensured that turnover was measured against the economically realized portfolio state rather than against the previously targeted allocation. Across folds, the procedure produced the out-of-sample return path, the sequence of rebalance allocations, the average rebalance weights, the final rebalance weights, and the cumulative transaction-cost drag used in the comparative evaluation reported in the Results section. The walk-forward optimization procedure, including fallback reuse of the carried-forward portfolio after post-initialization solver failure, is summarized in Algorithm 1. In Algorithm 1, F denotes the ordered set of walk-forward folds, where each fold contains one training window and one subsequent test window. The series R o o s , denotes the concatenated net out-of-sample portfolio returns produced across all test windows after transaction costs are applied. The metric set M is then computed from R o o s , and includes final value, total return, maximum drawdown, maximum loss, annualized mean return, Sharpe ratio, Sortino ratio, cost drag, and fallback count.

Algorithm 1 Walk-forward portfolio optimization.

3.4 Intent construction, hashing, and on-chain registration

Following walk-forward optimization, the selected allocation was transformed into a fixed portfolio intent artifact. Each artifact was assigned an increasing epoch and an expiry timestamp and recorded the strategy label, source mode, weight-selection mode, as-of date, creation time, source window, asset list, normalized target allocation, and summary strategy metrics. Asset serialization followed a fixed order defined by the study configuration so that equivalent portfolio states produced identical downstream hashes. For each asset, the artifact stored the human-readable name, ETF ticker, fixed on-chain address, continuous portfolio weight, and integer target allocation in basis points. In the fixed sample mode used for paper reproduction, the artifact additionally fixes epoch 1, created-at timestamp 2026–03-24T00, and expiry timestamp of 1893456000 (corresponding to 2030–01-01T00) so that the full serialized artifact can be regenerated exactly across runs.

Artifact integrity was enforced through three sequential hashes. A weights hash was computed as the SHA-256 digest of the ordered (asset address, target basis points) pairs. A metadata hash was then computed as the SHA-256 digest of the full artifact payload, epoch, strategy identifier, timestamps, source window, asset descriptors, basis-point vector, summary metrics, and implementation metadata, updated with the weights hash. The relay layer then derived the on-chain intent hash by applying keccak256 to the ABI-encoded tuple of epoch, asset addresses, basis-point vector, expiry, and metadata hash. Using SHA-256 off-chain and keccak256 on-chain kept artifact validation deterministic while remaining compatible with EVM contract logic.

Once assembled, the artifact was stored in a SQLite intent store with the epoch as the primary key, alongside the strategy, timestamps, status flag, metadata hash, intent hash, and the full serialized payload. New artifacts entered with status created; after on-chain relay, the record was updated to submitted with the registry address and transaction hash, preserving an auditable trail from research artifact to blockchain commitment. The PortfolioIntentRegistry contract, written in Solidity and deployed to a local Foundry (Anvil) EVM environment, accepts submissions only from the owner or an authorized optimizer address.

At submission, it enforces strictly increasing epochs, non-expired intents, non-zero asset addresses, matching array lengths, and an exact allocation sum of 10,000 basis points, then recomputes the intent hash, updates the latest epoch and hash, and emits an IntentSubmitted event. The relay layer converted addresses to checksum form, estimated gas, signed and broadcast the transaction, and waited for finality. A successful submission cross-links the off-chain record, the emitted event, and the on-chain state, completing the publication loop.

3.5 Evaluation metrics and reproducibility

The primary evaluation was performed on the concatenated out-of-sample return path obtained across all walk-forward folds. The reported performance measures were final portfolio value, total return, maximum drawdown, maximum loss, annualized mean return, Sharpe ratio, Sortino ratio, cumulative transaction-cost drag, and fallback count. Final value and maximum loss were reported in dollar terms from an initial portfolio value of $1,000,000 while annualization used 252 trading days. The Sharpe and Sortino ratios were computed using the aligned out-of-sample Treasury bill rate as the risk-free benchmark, with downside deviation used in the Sortino calculation. In addition to these scalar measures, the study retained both the average rebalance weights and the final rebalance weights since the former summarizes the walk-forward allocation profile, and the latter captures the final target state passed into the intent-publication pipeline.

Uncertainty and robustness were assessed in three ways. Sharpe ratio confidence intervals were estimated at the 95% level using 1,000 circular block bootstrap resamples of the out-of-sample return series. The block size was set to 63 trading days, matching the rebalance window exactly, so that the resampling preserves the autocorrelation structure of the evaluation period. A fixed random seed of 42 was used for reproducibility. Walk-forward sensitivity was examined under three alternative fold structures: the baseline 504/63 setting (2 years of training, one-quarter of testing), a monthly 504/21 setting (same training length, 1 month of testing), and a longer-training 756/21 setting (3 years of training). All windows are integer multiples of 252 trading days or 63-day quarters. Constraint sensitivity was evaluated under loose (0%–50%), baseline (5%–35%), and tight (10%–25%) weight bounds. Together these checks assess whether the observed results hold across different retraining frequencies and concentration limits rather than depending on a single experimental configuration.

Reproducibility was handled separately for the research path and the system path. The primary quantitative results were generated on the fixed historical window defined in the preceding subsections and therefore do not depend on later market updates. For the end-to-end publication pipeline, reproducibility was strengthened through a deterministic sample mode that produces an exact artifact with fixed source-window metadata, summary metrics, and basis-point allocations. In this deterministic path, the artifact uses the average rebalance weights rather than a floating live rerun, allowing the same portfolio intent, metadata hash, and intent hash to be regenerated exactly across runs. By contrast, the live path may extend the evaluation window to the present day and was retained only for exploratory reruns rather than for the fixed paper-facing artifact.

Implementation correctness was additionally checked through validation tests at both the off-chain and on-chain layers. The off-chain tests verified that generated intent artifacts sum to exactly 10,000 basis points, that stored intents round-trip correctly through persistence, and that incompatible runtime options are rejected.

4 Results

4.1 Experimental setup

The primary experiment was conducted on the fixed historical window from 1 March 2021 to 3 March 2026 using the seven-asset pool described in Section 3.2. After data alignment and removal of incomplete observations, the return matrix contained 1,256 daily observations. The baseline walk-forward configuration used a minimum training window of 504 trading days, a test window of 63 trading days, and a rebalance step of 63 trading days, which yielded 11 out-of-sample folds. Under this configuration, the out-of-sample evaluation period extended from 2 March 2023 to 3 December 2025. Transaction costs were fixed at 10 basis points per unit of turnover, weight bounds were fixed at 5%–35%, and portfolio statistics were annualized using 252 trading days. These baseline settings are summarized in Table 2.

Within this setup, the study compared six optimized strategies and one non-optimized benchmark. The optimized strategies were CVaR, Semivariance, Variance, EVaR, Worst Realization, and Maximum Drawdown, all estimated through the same walk-forward framework described in Section 3.3. The benchmark was an Equal Weight portfolio that assigned identical weights across all seven assets at each rebalance date. For each method, the experiment recorded the out-of-sample return path, total rebalances, cumulative transaction-cost drag, average rebalance weights, and the full set of scalar performance metrics defined in Section 3.5. The main comparison of this work focuses on the baseline fixed-window experiment while sensitivity analyses under alternative fold structures and weight bounds are reported in Section 4.4.

The results section is organized so that the baseline quantitative comparison is presented first, followed by allocation characteristics, robustness checks, and artifact-verification outcomes. The main baseline outputs are reported as the average rebalance weights table, the $1M walk-forward portfolio performance table, the walk-forward setting sensitivity table, the weight-bounds sensitivity table, and visual comparisons across Figures 2, 3 covering portfolio value, value-versus-loss, allocation heatmap, and risk-return positioning.

4.2 Out-of-sample strategy performance

All strategies were evaluated over the same 693 out-of-sample trading days and 11 rebalance events. Table 3 summarizes the baseline comparison. Among the tested methods, CVaR delivered the highest total return, ending with a portfolio value of approximately $1.64 million and a total return of 63.7%. EVaR ranked second in total return at 63.2%, while Semivariance, Variance, and Worst Realization also outperformed the Equal Weight benchmark. By contrast, the Maximum Drawdown strategy underperformed the Equal Weight portfolio, indicating that minimizing a single downside measure did not necessarily produce the strongest overall out-of-sample result.

TABLE 3. Out-of-sample performance comparison across all strategies.

The performance profile differed across evaluation criteria. CVaR produced the strongest total return whereas EVaR achieved the lowest maximum drawdown at 8.9% and the highest Sharpe ratio at 1.20. Semivariance and variance produced closely clustered outcomes, with total returns of 61.7% and 61.9%, respectively. Worst Realization remained competitive in return performance at 62.5% but incurred the highest transaction-cost drag at 0.29%. Equal Weight finished with a total return of 55.4%, a maximum drawdown of 10.3%, and a Sharpe ratio of 1.08 while Maximum Drawdown finished with a total return of 50.2% and the weakest Sharpe ratio in the comparison. The cumulative transaction-cost drag was modest across methods, ranging from 0.15% for Equal Weight to 0.29% for Worst Realization, with CVaR, Semivariance, Variance, and EVaR each recording 0.23%.

The Sharpe-ratio confidence intervals were broad and partially overlapping, which suggests that the strongest methods were separated more clearly in total return than in interval-based risk-adjusted certainty. CVaR yielded a 95% Sharpe confidence interval of [0.50, 2.12], EVaR yielded [0.52, 2.08], and Equal Weight yielded [0.35, 1.84]. Accordingly, the baseline results support two observations: risk-aware optimization improved on equal weighting under the fixed baseline specification, and different risk measures emphasized different aspects of portfolio quality, with CVaR favoring terminal return and EVaR favoring downside protection. Figure 2 shows the portfolio value trajectories and the final value versus maximum loss across all strategies over the evaluation period.

4.3 Average rebalance weights

The average rebalance weights provide a compact view of how each strategy allocated capital across the seven-asset pool over the 11 walk-forward rebalancing periods. These values, reported in Table 4, summarize the mean target allocation assigned at each rebalance rather than the final end-of-sample portfolio composition. As such, they provide a stable representation of each strategy’s structural preference across the evaluation horizon and are particularly useful for interpreting how different risk measures are translated into distinct portfolio profiles.

TABLE 4. Average rebalance weights per strategy under the baseline configuration.

Two patterns were consistent across the optimized strategies. All methods assigned minimum or near-minimum weight to Silver, REIT, and Treasury while concentrating most heavily in Energy Infrastructure and Agriculture, with Gold and Oil as secondary positions. This suggests that under the baseline constraints, the optimizer consistently favored the higher-returning commodity and infrastructure exposures while still holding the full asset pool. The Equal Weight benchmark, by construction at roughly 14.3% per asset, provides a useful neutral reference for comparison.

CVaR produced an average rebalance allocation of 32.9% to Energy Infrastructure and 24.9% to Agriculture, complemented by 15.4% in Gold and 11.8% in Oil while maintaining only 5% each in Silver, REIT, and Treasury. Semivariance and variance yielded closely related profiles, with slightly lower gold exposure and slightly higher Agriculture exposure than CVaR. EVaR remained similar in overall structure but shifted more weight toward Agriculture and less toward Energy Infrastructure, which is consistent with its stronger downside control. Worst Realization allocated more heavily to Gold and Oil than the other optimized methods while Maximum Drawdown placed the highest average weight on Energy Infrastructure at the 35% upper bound and the lowest average weight on Gold among the optimized group. These differences confirm that although the strategies were evaluated within the same walk-forward and constraint framework, the choice of risk measure materially affected the long-run allocation pattern.

From the standpoint of the ELVARA pipeline, the CVaR average rebalance weights are especially important because they formed the basis of the deterministic paper artifact used in the off-chain intent construction path. In other words, the published portfolio intent is not an abstract strategy label but a concrete serialization of the portfolio structure observed in the baseline walk-forward experiment. The cross-strategy allocation heatmap in Figure 3 further illustrates these patterns and makes clear that the strongest-performing methods shared a broadly similar allocation core while differing mainly in the relative balance among Gold, Oil, Energy Infrastructure, and Agriculture.

4.4 Sensitivity analysis

To assess the robustness of the baseline findings, the study repeated the out-of-sample evaluation under alternative walk-forward settings and portfolio weight bounds. These tests examine whether the baseline ranking depends materially on the experimental design choices used in the main specification. Walk-forward parameters are not neutral. The number of folds, the length of the training window, and the rebalance frequency all shape how much historical data the optimizer sees before each decision, how often allocations are updated, and how many independent out-of-sample periods contribute to the final performance record. Similarly, the permitted weight range directly governs how much the optimizer can deviate from equal weighting. When bounds are tight, all strategies are pushed toward the same allocation regardless of their risk objective. When bounds are loose, the optimizer can express strong conviction, which amplifies both the advantages and the costs of each strategy’s particular risk focus. Testing across a range of these settings is therefore not a secondary robustness exercise but a necessary part of understanding what the baseline result claims.

The sensitivity checks confirm that both the retraining schedule and the weight bounds materially affect strategy rankings. Varying the fold structure across the baseline 504/63, monthly 504/21, and longer-training 756/21 settings produced different winners in each case (Table 5). CVaR led under the baseline and longer-training schedules, while Worst Realization ranked first under the monthly setting at 80.8%. The shift is not surprising. More frequent rebalancing gives every strategy more chances to adapt allocations, which tends to help strategies that benefit from short-term responsiveness more than those that optimize longer-horizon tail risk. Equal Weight also rose substantially under more frequent rebalancing, narrowing the gap with the optimized methods to a much smaller margin than under the baseline schedule. Under the longer-training 756/21 configuration, CVaR remained ahead but only barely, at 54.7% versus 54.6% for Equal Weight, which suggests that with 3 years of training data the optimizer’s edge over the naive benchmark compressed considerably. Varying the weight bounds told a similar story (Table 6). Loose bounds (0%–50%) favored Maximum Drawdown at 74.7%, while the baseline bounds (5%–35%) favored CVaR at 63.7%, and tight bounds (10%–25%) favored Worst Realization at 59.8%.

TABLE 5. Strategy performance sensitivity to walk-forward schedule.

TABLE 6. Strategy performance sensitivity to portfolio weight bounds.

Tighter constraints made the strategies more similar because the optimizer had less room to change the asset weights. This moved the portfolios closer to equal weighting. These results also clarify the main claim of the paper. The paper does not claim that CVaR is always the best strategy across all walk-forward settings. It only shows that CVaR performed best under the fixed baseline setting and was then used in the deterministic artifact and on-chain publication workflow. This is the intended scope of the proof-of-concept study.

4.5 Deterministic intent artifact and on-chain verification

The CVaR baseline output was converted into a fixed paper-reproduction artifact using the sample mode. The artifact stored the strategy label, epoch, source-window metadata, ordered asset list, continuous weights, basis-point targets, summary metrics, and integrity hashes. The basis-point allocation was 1,540 for Gold, 500 for Silver, 1,180 for Oil, 500 for REIT, 500 for Treasury, 3,290 for Energy Infrastructure, and 2,490 for Agriculture, summing to exactly 10,000. Because all inputs were fixed from the average rebalance weights rather than a live rerun, the same artifact, metadata hash, and intent hash are reproduced exactly on every execution. Off-chain tests confirmed that allocations sum correctly, epoch assignment is consistent, and incompatible runtime modes are rejected before any blockchain interaction occurs. This matters for reproducibility: a reader who re-runs the sample mode path from the same repository ¹ should arrive at an identical artifact, not an approximation. The published intent therefore carries the same evidentiary status as the numerical results in the tables; it is a direct, deterministic output of the research pipeline rather than a post hoc representation.

On-chain verification confirmed that the same artifact could be committed to the registry without altering the portfolio semantics. After relay, the contract recomputed the intent hash, updated the latest epoch and intent hash, stored the record, and emitted the corresponding event. Querying the registry returned epoch 1 with a non-zero intent hash, confirming successful publication. Contract tests further verified authorization, epoch monotonicity, expiry validation, non-zero asset addresses, and exact 10,000-basis-point enforcement. The pipeline therefore closes the loop between walk-forward optimization and verifiable on-chain publication. This shows that a portfolio decision can be turned into a machine-readable record and checked publicly on-chain. The record keeps the same weights and metadata from the research output. It does not rewrite, summarize, or explain the allocation. It stores the exact commitment so that other systems can verify it from the contract.

5 Discussion

5.1 Principal findings

The results support three observations. Risk-aware optimization outperformed the equal-weight benchmark under the fixed baseline configuration, with CVaR reaching a 63.7% total return versus 55.4% for Equal Weight, while also maintaining slightly lower maximum drawdown. Most optimized methods beat Equal Weight, so the improvement was not unique to CVaR alone.

The choice of risk measure did affect how that improvement materialized. CVaR produced the strongest terminal wealth, while EVaR achieved the lowest maximum drawdown and highest Sharpe ratio. No single measure dominated across all criteria simultaneously, which is the expected behavior when different objectives weight return, tail risk, and cost differently. CVaR was selected as the paper-facing strategy because it delivered the strongest overall return profile under the fixed configuration while remaining compatible with the deterministic artifact path.

Beyond the backtest, the optimized allocation was normalized to 10,000 basis points, hashed, stored off-chain, and committed to the on-chain registry through the relay layer. A portfolio decision was therefore represented as a reproducible, verifiable publication object rather than a spreadsheet output. That is the main contribution of the systems layer: the optimization and the on-chain record refer to the same portfolio state. This property is harder to achieve than it sounds. In a typical research workflow, a portfolio result lives in a notebook or spreadsheet and the on-chain representation, if any, is constructed separately and may diverge in subtle ways from what was computed. ELVARA closes that gap by making the artifact a deterministic function of the research output rather than a manual translation of it. Any party with access to the registry and the research repository can verify that what was published on-chain matches what the optimizer produced, without relying on the authors to certify the correspondence.

A further important observation is that the relative ranking of strategies was sensitive to the walk-forward schedule and the admissible weight bounds. Tables 5 and 6 show that the identity of the highest-returning method changed under alternative fold structures and concentration limits. This clarifies the scope of the baseline conclusion. The main claim of the paper is not that CVaR is universally optimal, but that under the fixed baseline configuration chosen for the experiment, CVaR provided the strongest overall result and could be carried through a fully reproducible off-chain and on-chain publication workflow.

5.2 Comparison with prior work

Prior work related to ELVARA has generally addressed one component at a time. Tokenization studies focus on representing and governing off-chain assets on-chain, portfolio optimization studies stop at the numerical weight vector, and intent-based blockchain research addresses execution and routing rather than portfolio construction. ELVARA differs by combining all three into a single workflow: the optimizer produces an allocation, that allocation is serialized into a deterministic artifact, and the artifact is committed on-chain as a verifiable portfolio intent.

Across all three reference points, ELVARA’s contribution is the same: it adds a layer the prior work leaves out. Against tokenization research, it adds a portfolio decision layer that derives a rebalancing target from formal optimization rather than simply representing a pre-chosen asset on-chain. Against portfolio optimization research, it adds a systems extension that turns the weight vector into an auditable, serializable intent object. Against blockchain coordination research, it shifts the role of the intent from trade execution to portfolio-state publication.

The distinction is also visible against representative live protocols. Ondo ² , OpenEden ³ , Backed ⁴ , and Centrifuge ⁵ focus on tokenized exposure, vault structure, issuance, redemption, and compliance rather than on a walk-forward optimizer publishing epoch-ordered portfolio intents (see notes 1–4). CoW ⁶ Protocol and UniswapX ⁷ demonstrate that signed intents work well for swap settlement and routing, but their architecture is not designed for multi-asset RWA portfolio construction (see notes 5–6). ELVARA occupies a different position: the intent is a portfolio-level commitment from a risk-aware optimization pipeline, not a single-trade execution message. Its novelty lies in linking optimization, artifact generation, and on-chain commitment into one reproducible framework, as summarized in Table 7.

TABLE 7. Comparison of ELVARA with prior academic work and representative live protocols.

5.3 Practical and research implications

ELVARA suggests that portfolio management and blockchain publication can be linked more tightly than current tokenized-asset workflows typically allow. In most existing settings, asset selection, risk management, and on-chain representation are handled as separate stages. The registry design shows that a rebalancing decision can instead be published with explicit epoch ordering, expiry constraints, and deterministic hashes, making downstream verification easier than off-chain reporting alone.

The results also indicate that adding a risk-aware optimization layer can produce meaningful improvements over naive equal-weight allocation in a constrained backtesting environment.

Tokenized treasury, equity, commodity, or multi-asset products therefore do not need to remain static wrappers around underlying assets. They can be designed as programmable portfolio systems whose target compositions are periodically recomputed, serialized, and committed transparently. In that sense, ELVARA is relevant not only for academic portfolio studies, but also for protocol designers, asset issuers, and treasury managers interested in combining tokenization infrastructure with reproducible allocation logic.

Two research directions follow from this work. RWA tokenization research would benefit from moving beyond asset representation toward how portfolio-level decision rules can be formalized and verified on-chain. Intent-based architectures are also worth examining outside the trade execution context: in ELVARA, the intent carries a full portfolio state rather than a single swap, which suggests a broader role for signed intents as coordination objects for rebalancing, treasury reallocation, and multi-agent oversight.

5.4 Stakeholders, blockchain rationale, governance, and compliance in deployment

ELVARA is a proof-of-concept publication layer, and the easiest way to see its practical value is to look at who would use it in production. The publisher is whoever makes the allocation decision, such as a tokenized fund issuer, an RWA vault manager, a DAO treasury, and in the current implementation this is the contract owner or an authorized optimizer key. Verifiers are the people who need to trust that an allocation is authentic, correctly ordered, and unmodified, such as auditors, governance participants, or token holders, while actors are the systems that execute it, such as custodians, brokers, or vault contracts. The registry checks form and authorization rather than economic soundness, so responsibility for the content stays with the publisher. The gas cost depends on which blockchain it runs on, and a protocol could also choose to sponsor it, while the cost stays low because publishing an intent is just one small update per rebalance. Publishing on-chain rather than sharing a signed off-chain file matters because it adds a third-party-verifiable timestamp, a tamper-evident order, and public non-repudiation, and because other smart contracts can read the commitment directly to automate tasks like vault rebalancing. Epoch ordering, expiry validation, and a recomputed intent hash let anyone confirm the on-chain record matches the off-chain file.

Because weights are generated off-chain, on-chain validation cannot certify model quality, so model governance stays the responsibility of the institution or the protocol’s developers, though every artifact embeds the strategy, data window, parameters, and metrics so any change is verifiable after the fact. ELVARA does not execute trades, settle assets, manage custody, or enforce compliance, and the study treats it as research and coordination infrastructure rather than a regulated product. In a research or treasury context an intent is simply a transparent declaration of target state, but in a regulated fund it could count as a model portfolio or rebalance instruction, triggering obligations around suitability, disclosure, and record-keeping that would need to sit inside existing compliance systems. This decoupling is also what lets ELVARA handle illiquid real-world assets, because the intent declares a target state with an explicit expiry instead of triggering immediate settlement, so execution can be deferred, partial, or routed through custody windows and broker APIs.

6 Conclusion

This study presented ELVARA as a single-chain proof of concept that links walk-forward, risk-aware portfolio optimization for an RWA-oriented ETF proxy basket with deterministic artifact generation and minimal on-chain publication. Under the fixed baseline configuration, the CVaR strategy outperformed the Equal Weight benchmark in terminal wealth while remaining competitive on downside risk and transaction-cost drag. More importantly, the study showed that the resulting baseline portfolio representation can be normalized into basis-point targets, serialized into a fixed paper-reproduction artifact, hashed, stored off-chain, and committed on-chain through a registry contract.

The significance of this work is integrative. The optimizer, artifact model, persistence layer, relay path, and registry contract are each simple on their own, but together they form a coherent bridge between portfolio research and blockchain-native verification. Rather than treating tokenized asset infrastructure and quantitative allocation as separate concerns, ELVARA shows that a portfolio state can be computed, represented, and published in a form that other systems can inspect and reproduce. Within the scope of this proof of concept, that is a practical foundation for verifiable RWA-oriented portfolio coordination. The sensitivity analyses further reinforce this framing: strategy rankings shifted across different walk-forward schedules and weight bounds, which is expected when comparing methods that optimize different aspects of portfolio risk. The baseline result is not that CVaR is a universally superior risk measure. It is that, under one well-defined and fully reproducible specification, CVaR delivered the strongest outcome and proved compatible with a complete deterministic publication pipeline.

A reproducible system must be evaluated under fixed conditions, and this protocol documents exactly what those conditions were. More broadly, the work suggests that the gap between quantitative portfolio research and blockchain-native infrastructure is narrower than it appears when these fields are treated separately, and that bridging it requires less architectural complexity than is often assumed.

From an economic policy view, the results have two implications. First, verifiable intent publication gives supervisors, auditors, and investors a low-cost way to make tokenized asset products transparent, since allocation decisions become independently timestamped, ordered, and reproducible, which narrows the information gap between operators, investors, and regulators. Second, because the registry keeps decision publication separate from execution, policymakers can judge disclosure-focused uses of blockchain on their own terms, without grouping them together with custody, settlement, or trading rules. These points apply within the scope of the proof of concept, and real-world adoption would still need the governance and compliance controls covered in Sections 5.4 and 6.1.

6.1 Limitations

Several limitations apply. The asset pool uses ETF proxies rather than natively tokenized instruments, so the empirical results should be read as an architecture demonstration rather than performance evidence on live RWA products. The on-chain component is minimal by design: it records and validates intents but does not handle execution, settlement, custody, or compliance. Performance figures are specific to the fixed historical window and baseline walk-forward configuration, and the sensitivity analyses confirm that strategy rankings shift under alternative settings. The live recomputation path depends on external market-data downloads and is therefore less reproducible than the deterministic sample path.

The protocol also does not settle how regulators would classify published intents. Whether an intent counts as a model portfolio, an investment recommendation, or an execution instruction depends on where and how it is used, and the related legal duties around trust, suitability, disclosure, and investor protection fall outside the scope of this paper. Model governance is likewise left to the institution rather than enforced by the protocol, since the registry only checks authorization and artifact integrity, not whether the underlying optimization is economically sound. In the same way, gas-cost benchmarking, latency profiling, and production-chain deployment all sit outside the scope of this proof of concept.

6.2 Future research

Future work can extend ELVARA in three directions: asset realism, methodological depth, and systems expansion. Replacing ETF proxies with live tokenized instruments would allow evaluation under production-oriented settings. Additional risk measures, adaptive constraint schemes, and broader statistical testing could clarify when specific optimization objectives are most appropriate. Future versions could also support signed intents, richer governance, direct execution modules, multi-chain deployment, and formal benchmarking of gas consumption and confirmation latency.

Across these directions suggest that ELVARA should be viewed as a foundation rather than a finished production protocol. Its value lies in showing that risk-aware portfolio outputs can be transformed into deterministic, machine-readable, and on-chain-verifiable intent objects. Extending that principle to broader asset classes and more mature deployment settings is a promising next step for both RWA systems research and blockchain-based portfolio infrastructure.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. The authors declare that generative AI was used for language refinement during manuscript preparation. OpenAI GPT-5.2 (OpenAI, December 2025) was used for drafting support and language editing. The authors reviewed and edited all output, verified the accuracy of the final text, and take full responsibility for the content of the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fbloc.2026.1896673/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Marcelle MGM Kress von Wendland, Bancstreet Capital Partners Ltd., United Kingdom

Reviewed by: Krisanth Sivarajh, RealAssets, United States

Ioan Batrancea, Babeş-Bolyai University, Romania

Footnotes

https://github.com/0xsha/elvara.

https://ondo.finance

https://openeden.com

https://backed.fi

https://centrifuge.io

https://cow.fi

https://blog.uniswap.org/uniswapx-protocol