II. Market Risk (continued)
2.10 RWA Market Volatility
The portfolio-level drivers of mWIN’s drawdown risk are the sector mix, the credit-quality distribution, and the duration profile. The sector and rating tables below are the modeled target.
Portfolio Characteristics
The model portfolio is short-duration and investment-grade: effective duration 0.79 years, spread duration 2.04 years, average credit quality A+. To first order, a 100 basis-point rate move costs roughly 0.8% of value and a 100 basis-point spread widening roughly 2%. The live book, as reported by the issuer as of September 15, 2026, carries an effective duration of 0.6 years, a spread duration of 2.6 years, and a market yield of 5.52%. On that spread duration the same widening costs roughly 2.6%. The issuer states that the portfolio ramp-up is complete and the book sits close to the model portfolio.
On August 26, 2026, the live book held 94 positions worth USD 24.7 million, an average position of roughly USD 263,000, spread across 16 collateralized loan obligation managers with 26 deals and 44 corporate bonds from 44 issuers. Minimum lot sizes of USD 100,000 to 250,000 keep the position count low while the fund is small, with positions expected at 0% to 2% each as the fund grows. Two guardrails bind in advance: a 5% cap on any single issuer or issuing trust of structured securities, government paper excepted, and a 0-to-2-year band on portfolio effective duration. Neither rule limits how sensitive the book can be to widening credit spreads. That exposure is set by the manager’s day-to-day positioning choices rather than by any mandate limit. The mandate also permits instruments the model book does not hold, repo transactions, to-be-announced forwards, and interest-only, principal-only, and inverse-floater mortgage tranches, so the mandate allows more rate sensitivity than the model shows.
Sector Exposure
The modeled sector allocation is built from three sleeves: a 55% Structured Finance and Floating-Rate Income sleeve, a 25% High-Quality Institutional Credit sleeve, and a 20% Liquidity and Defensive Anchors sleeve. The investment guidelines cap no sector, so the weights below are the modeled composition.
Source: Documents shared under NDA, June 2026
The single largest modeled sector is CLOs at 40%, almost entirely A-rated. Structured and securitized product taken together is roughly 75% of the model book and 75.1% of the live book, so the dominant risk factor is securitized-credit spread, not corporate or sovereign risk. The forward-looking stress analysis credits a United States rates and agency-RMBS allocation as a flight-to-quality offset.
The live book is positioned more heavily in collateralized loan obligations than the model which is confirmed by the issuer to be an active-management choice: Wellington’s strategic targets have moved since the model portfolio was struck. The asset-backed securities are 56.0% of market value, all of them collateralized loan obligations backed by broadly syndicated bank loans, mortgage paper 19.1% across passthroughs and collateralized mortgage obligations, corporates 20.1%, and a single Treasury position 4.8%. It is expected that the asset-backed sleeve broadens over time into traditional and esoteric consumer and commercial deals such as autos, credit cards, and equipment.
The leverage prohibition is settlement-based, the account counts as levered only if overdrawn on a settlement basis, while the permitted futures, swaps, options, and repo can carry economic leverage without one. And compliance is checked only at purchase. A limit exceeded through market moves, cash flows, or rating changes is not a breach, Wellington may continue to hold, must notify within 30 days, and no cure obligation or forced sale follows. This has two consequences. The book is never forced to sell below fundamental value to cure a passive breach, and nothing forces it back inside its limits.
Sector Stress Behaviour
No daily return series exists for the book, so each sleeve’s stress behaviour is read from a daily-priced fund over the same paper. The figures are exchange prices.
In the model book, asset-backed securities and non-agency residential mortgage paper, roughly 20% of market value, have no public proxy. On the live book the asset-backed line is held entirely in collateralized loan obligations, so the CLO proxies reach it and the uncovered residual is the non-agency mortgage paper alone. The corporate and mortgage proxies carry ten years of history including March 2020, while the CLO proxies list only from October 2020 and January 2022, so the reference for 56.0% of the live book has never traded through a liquidity event of that kind. The chart carries the episodes, with each proxy’s worst one-day and five-day prints, the figures the parameter work uses, in the legend. Its deepest prints are −12.86% on the corporate proxy over ten trading days in March 2020, and the 2022 rate shock on the collateralized loan obligation proxies, −3.96% at AAA against −13.65% at BBB, the spread the credit-quality discussion leans on.
Source: NASDAQ, August 18, 2026
The sleeves don’t have high correlation and how they couple depends on the shock. In the 2022 rate shock the rate-sensitive sleeves moved together, a 0.88 daily-return correlation between the corporate and agency-mortgage proxies, while the floating-rate collateralized loan obligation proxies decoupled from both at correlations near 0.3. In the March 2020 liquidity event daily correlations were low, 0.14 to 0.66 across the three proxies with 2020 history, yet all three printed their worst days inside the same eight trading days. The low daily correlations offer no real protection as in a liquidity event all of the sleeves fall in the same week, meaning that longer horizon correlation is the driving factor.
Geographic Exposure
The reviewed materials do not break the portfolio out by country or region. The sector composition is United States-centric, built on United States agency and non-agency mortgage paper, United States CLOs and ABS, and United States investment-grade corporates, and the stress analysis references North American credit and United States rates exposures.
The portfolio is United States-dollar-denominated, non-dollar securities are capped at 10% of the book and must be hedged 100% back to the base currency within a 99%-to-101% band, and active currency management is not permitted. Currency risk is therefore bounded to a small, fully hedged residual.
Credit Quality
The modeled book is entirely investment grade. The rating distribution from the modeled snapshot is as follows.
| Rating | Market value % |
|---|---|
| AAA | 5% |
| AA | 26% |
| A | 47% |
| BBB | 22% |
Source: Documents shared under NDA, June 2026
More than 75% of the modeled portfolio is rated A or higher, and there is no sub-investment-grade or unrated allocation in the snapshot. Every modeled position carries an external rating. This snapshot is more conservative than the mandate permits. The investment guidelines require only an investment-grade weighted-average rating and allow up to 50% in BBB or Baa securities, up to 10% rated below investment grade, and up to 10% unrated. A+ average is expected by the issuer for the live book once ramping completes.The live book is rated higher than the model. As reported by the issuer as of September 11, 2026, it is 55% AAA, 24% AA, 8% A and 13% BBB, so 79% sits in the two top rating bands against 31% in the modeled snapshot, and 87% is rated A or better against 78%.
What the rating mix does to volatility and tail risk is evidenced in the stress behaviour subsection, where the AAA and BBB collateralized loan obligation proxies bound the book’s blend: annualized volatility runs near 2% at pure AAA against 5.4% at pure BBB, and the worst observed episodes near −4% against −13.7%. The modeled composition, 78% rated A or better, sits between those bounds, closer to the AAA end and the live composition, with 79% in AAA and AA, sits closer still.
The Palmer Square CLO Senior Debt Index, daily since 2015 and covering AAA-senior collateralized loan obligation debt, drew down 8.33% peak-to-trough into March 24, 2020, over 19 trading days with 115 trading days underwater, with a worst single day of −1.57%. The mWIN model book, holding its CLO exposure four to five rating notches below that index alongside non-agency mortgage and asset-backed paper, backtested a 13.32% drawdown in the same window, roughly 1.6 times the senior index, and roughly 17% carried to the live book’s 2.6-year spread duration. The senior index is a floor on this composition’s stress and no exchange-traded fund exists at the model book’s A-rated centre of gravity, the AAA and BBB funds in the sector-stress table bracketing it from either side.
Originator / Issuer Concentration
Source: Wellington, August 19, 2026, shared under NDA
The book spans 16 collateralized loan obligation managers across 26 deals, 2 non-agency mortgage deals from distinct managers, and 44 corporate bonds from 44 issuers. The investment guidelines’ 5% structured-securities cap applies per deal, each deal being a distinct legal entity, and every deal sits inside it. No cap reaches manager-level exposure or the corporate sleeve.
The levels themselves are low: no manager above 7.0%, no single issuer above 3.6%, no corporate name above 1.0%. Current concentration is acceptable.
Liquidity Profile
There is no maturity-bucket distribution is available. Wellington’s simulation of the live portfolio, shared under NDA, puts 63% of the portfolio saleable in one day under normal market conditions on its expected path, 76% on its 95th-percentile path, and 100% of every sector’s market value within one week. The simulated cost of selling the whole book is 0.1% of market value, or 0.28% if the whole book is fire-sold in a day, with collateralized mortgage obligations the widest sector at 0.75% and 2.24%. Under the simulation’s stressed conditions an orderly sale reaches roughly 87% of market value in one week and 99% in one month, and a same-day fire sale of the whole book is priced at roughly 3.3%.
2.11 Peer / Sibling Case Study
mWIN’s weeks of live history contain no stress event, so the relevant precedent for how its tokenization infrastructure behaves under stress comes from the other products on the Midas platform. Its closest architectural sibling is mGLOBAL, with which it shares the mTokenPermissioned base contract and the same on-chain ProxyAdmin, TimelockController, and access-control singletons. mGLOBAL is itself a recently structured product with no observed stress event of its own, so the only Midas token that has passed through a credit-stress episode with an on-chain oracle write-down is mF-ONE, a token backed by the Fasanara F-ONE trade-receivables fund that runs on the same Midas issuance, redemption, and oracle framework mWIN inherits.
Event Timeline
mF-ONE was deployed on June 11, 2025. The stress sequence ran as follows.
| Date | Event |
|---|---|
| September 24-28, 2025 | First Brands Group files Chapter 11 bankruptcy (factoring-fraud allegations, roughly $24M Fasanara exposure). Oracle price unchanged. |
| November 3-6, 2025 | The Stream Finance collapse triggers rapid redemptions. TVL falls from $80.5M to $24M at 100% utilization, with 14.3M mF-ONE burned through the redemption vault. |
| December 3, 2025 | Oracle write-down: base NAV -1.96%, discounted oracle -2.07%. First on-chain signal, roughly 66 days after the bankruptcy filing. |
| December 8, 2025 | A second exit wave follows the write-down; TVL falls to $45.8M. |
| February 27, 2026 | NAV recovers above the pre-write-down level; collateral declines to $15.4M with zero liquidations throughout. |
Observations
The dominant feature of the episode was the lag between the credit event and its appearance in the oracle. The First Brands bankruptcy was filed in late September 2025 and the oracle wrote the loss down on December 3, 2025, a gap of about 66 days produced by the amortized-cost, monthly valuation of an illiquid private-credit book. The lag opened a stale-price redemption window: roughly $14 million of mF-ONE was redeemed about 30 days before the write-down, exiting at a net asset value that did not yet carry the loss and leaving it with the holders who remained, while new capital kept entering at the same stale price.
No forced liquidations occurred despite the write-down and a spell of 100% utilization. That outcome rested on voluntary deleveraging by a small number of sophisticated borrowers, a single address repaying 71% of the USDC, not on any structural backstop. The 1.96% write-down itself was rebuilt in roughly 90 days by the fund’s receivables yield.
Relevance to mWIN
The infrastructure carries over but the loss model does not hold for mWIN. The 66-day lag and the stale-price window were products of an amortized-cost monthly valuation that mWIN does not share, and a realized single-name write-down is not mWIN’s dominant loss channel, as the valuation review above establishes. Three facts survive the translation:
- The manual submission step between the published net asset value and the on-chain price is the same, and this episode is the observed behaviour of that step under stress.
- The zero-liquidation outcome was decided by the voluntary behaviour of one dominant address, and roughly 73% of mWIN’s supply sits with its three largest investors, so the precedent of a few actors deciding the stress outcome carries over.
- mWIN publishes a single daily feed with no analogue of mF-ONE’s discounted mint-and-redeem pair, so the 11 basis-point oracle discount visible in the write-down has no counterpart in mWIN’s design.
The comparison comes down to four points. On valuation cadence mWIN is materially safer: a daily mark-to-market strike would have compressed mF-ONE’s 66-day recognition lag to about one business day, and the stale-price redemption window with it. On backing mWIN is safer again: a marketable investment-grade book, simulated as roughly two thirds saleable in one day and fully within a week, can be sold to meet exits where mF-ONE’s private receivables could not, and a diversified securitized book has no analogue of the single-name fraud that produced the write-down. One design point runs the other way: mF-ONE’s discounted oracle pair built a small buffer into exits that mWIN’s single feed does not have. The manual submission step and the concentration of supply in a few holders are common to both and favor neither. The verdict is that an analogous episode should resolve significantly more safely on mWIN, on the strength of the valuation cadence and the asset backing.
2.12 Stress Testing
mWIN’s underlying is a marketable multi-sector fixed-income book, so the loss that matters for collateral is mark-to-market drawdown, not borrower default, and the stress analysis below sizes that drawdown.
Manager’s Forward Scenario Analysis
The forward-looking analysis applies a set of macro scenarios to the model portfolio and reports a first-order price impact and a carry-driven recovery time for each.
| Scenario | Modeled price impact | Recovery time |
|---|---|---|
| Credit sell-off | −2.54% | 5.67 months |
| Risk-off | −0.42% | within 6 months |
| Severe recession | −0.37% | within 6 months |
| Base | −0.21% | within 6 months |
| Recession | +0.10% | n/a |
| Bullish | +0.33% | n/a |
Source: Documents shared under NDA, June 2026
The worst modeled outcome, the 2.54% credit sell-off, is mild. As the issuer describes it, the scenario widens investment-grade and high-yield credit spreads by roughly 190 basis points and A-rated CLO spreads by roughly 320, and offsets part of that loss with a 200 to 250 basis-point fall in front-end interest rates. The 2.54% is therefore a net figure. The blended widening of roughly 230 basis points costs the model book about 4.7% through its 2.04-year spread duration, and the rally gives back about 1.6 to 2.0 percentage points through its 0.79-year effective duration. The offset is limited by the book’s short rate duration. At the live book’s 0.6-year effective duration a 200 basis-point rally recovers roughly 1.2 percentage points, which does not neutralize a spread event of crisis size. The same scenario applied to a pure AAA CLO portfolio gives a 5.5% drawdown, while the Palmer Square CLO Senior Debt Index fell 8.33% in March 2020, so the scenario is milder than that episode. The recovery times are conservative in one specific sense, they assume carry accrual alone with no spread retracement. Applied to the crisis drawdowns this section sizes, carry at the model book’s 4.63% net-of-fee yield would need roughly three years to rebuild a 12% to 13% drawdown and nearly four years to rebuild the 17% the live book’s spread duration implies. Those multi-year figures describe the slowest possible recovery, where prices never bounce back and interest income has to earn the whole loss back on its own. Actual crisis losses have recovered far faster, because most of the loss is prices marking down and prices rebound once spreads normalize: in March 2020 the book was down 13.32% at the worst point and still finished the same year up 4.98%.
The strongest tail evidence is the model’s own backtest. In March 2020 the book would have lost 13.32%, more than five times the manager’s worst forward scenario, and on the live book’s higher spread sensitivity the same episode is roughly a 17% loss. Because the net asset value is marked to market daily, the decline reaches the price feed day by day as it happens, so a borrower is liquidated on the first daily print past the trigger.
Independent Cross-Validation
3 independent checks were run against the manager’s figures.
The first recomputes the drawdown directly from duration, at the model portfolio’s 2.04-year spread duration and at the live book’s 2.6 years, applying broad credit-spread shocks of the magnitude seen in past dislocations.
| Spread shock | Model book, 2.04y | Live book, 2.6y |
|---|---|---|
| +250 bps (moderate crisis) | 5.1% | 6.5% |
| +400 bps (severe crisis) | 8.2% | 10.4% |
| +600 bps (GFC-scale investment-grade peak) | 12.2% | 15.6% |
| +600 bps with a 200 to 250 bps rates rally | 10.5% | 14.2% |
Source: LlamaRisk, on durations shared under NDA, June and September 2026
The second replays the last decade through the sector proxies. The five proxies, covering 80% of the model book, are combined at the model’s sleeve weights, renormalized as each proxy lists, and the composite’s drawdown path is read against the manager’s loss points. On the live book the same proxies reach every sector except the non-agency mortgage paper. The replay troughs at −9.5% in March 2020 and −9.8% in the 2022 rate cycle. Both sit far below the −2.54% worst forward scenario, which the composite in fact breaches even in ordinary years, and both stay inside the −13.32% model backtest trough, which remains the deepest print on the page.
Source: LlamaRisk, August 2026
The linear extrapolation understates the securitized book, where collateralized loan obligation and non-agency mortgage spreads gapped wider than corporates in both 2008 and 2020, so the observed −13.32% stays the better crisis anchor for the model composition. Rate risk is secondary: at the model’s 0.79-year effective duration the 2022 hiking cycle produced only a −1.57% year, and the guidelines hold effective duration under two years. Combining both legs at crisis magnitude, 600 basis points of spread widening and a 200 basis-point rate rise, gives roughly 13.8% first order on the model book, in line with the −13.32% it printed, and roughly 17% on the live book’s durations. The same widening with the flight-to-quality rally Wellington’s scenario assumes, a 200 to 250 basis-point fall in rates, nets to roughly 10.5% on the model book and 14% on the live book. The rates gain stays near two points at most, because effective duration is under one year and rates can only fall so far, while the spread loss keeps growing with the widening. The 2022 cycle showed the offset can also run the other way, with rates rising as spreads widened.
A third check reads the same proxies at daily resolution, since a daily-strike collateral feed is exposed to what one publication can move.
| Proxy | Coverage | Worst 1-day | Worst 3-day | Worst 5-day |
|---|---|---|---|---|
| VCSH, 1 to 5 year investment-grade corporates | Aug 2016 to Aug 2026 | −3.43% (2020-03-19) | −6.37% | −8.74% (2020-03-12 to 2020-03-19) |
| VCSH excluding 2020 | −0.92% | −1.90% | −2.24% (June 2022) | |
| JAAA, AAA collateralized loan obligations | Oct 2020 to Aug 2026 | −1.03% (2025-04-10) | −1.25% | −1.60% |
| JBBB, BBB collateralized loan obligations | Jan 2022 to Aug 2026 | −2.26% | −3.61% | −4.99% (Sep 2022) |
Source: Nasdaq historical quotes, fetched August 18, 2026
The panel carries three gaps. The figures are exchange prices rather than net asset values, and in March 2020 short-duration bond fund prices traded below net asset value, so the 2020 rows overstate what an evaluated-price net asset value would have printed. VCSH carries a spread duration near 2.6 years, above the model portfolio’s 2.04 and equal to the live book’s 2.6, so the panel overstates the model book’s single-day sensitivity and matches the live book’s. The two collateralized loan obligation funds are floating-rate like the CLO sleeve, 40% of the model book and the whole of the live book’s 56.0% asset-backed line, JBBB sits one notch below the modeled sleeve’s A-rated bulk, and neither existed in March 2020.
Crisis Extrapolation and Default Overlay
Because the dominant loss channel is mark-to-market, the headline crisis figure is the drawdown extrapolation above: on the model book, on the order of 5% in a moderate crisis and 12% at a GFC-scale spread peak, roughly 6.5% and 16% on the live book’s spread duration, and the −13.32% the model book actually printed in March 2020, roughly 17% carried to the live book. Actual defaults add little even in the worst case. Take the riskiest book the mandate allows, up to 50% BBB, 10% below investment grade, and 10% unrated, and let crisis-peak default rates hit it with zero recovery: the loss is on the order of 2.5%, roughly 1% at a realistic 60% recovery, almost all of it from the below-investment-grade and unrated slices. Against crisis price drawdowns near 13% on the model book and near 17% on the live book, that is a fifth the size or less, so falling prices, not failing borrowers, remain the important loss channel.
Assessment
The conservatism verdict is that the manager’s scenario set understates the tail. The worst forward scenario, −2.54%, sits below the composite proxy replay’s near-10% crisis troughs and far below the portfolio’s own realized −13.32% drawdown. This is not a defect in the manager’s model, which measures a carry-driven recovery under moderate shocks and is sound for that purpose. The mitigating structural features remain real and are quantified elsewhere:
- The effective-duration cap limits rate risk. The guidelines set it at zero to two years and the issuer states it has been tightened to one year, with the model portfolio at 0.79 years and the live book at 0.6 years inside it.
- The mandate holds the weighted-average rating at investment grade, with up to 50% BBB and up to 10% below investment grade permitted, and the 5% single-issuer cap limits single-name risk within the structured book.
- The daily mark-to-market NAV recognizes a drawdown within one business day, not after months of smoothing.
- A mark-to-market loss retraces with spreads where a defaulted private-credit loss does not.
Limitations
Three limitations stand:
- The manager supplies scenarios, not a full loss distribution, so the only tail evidence is the March 2020 backtest, and losses have no reason to stop at what one episode produced.
- The analysis is struck on the model portfolio while the live holdings are known only through manager-provided aggregates. The live book has a spread duration of 2.6 years against the model’s 2.04, driven by a heavier collateralized loan obligation allocation. CLOs cluster at three to four years of spread duration with new issues at six to eight. The issuer states that Wellington has agreed limits of one year on interest-rate duration and four years on spread duration, with the objective of staying below both. The spread-widening losses shown in this review scale up by roughly a quarter on the live book, and by roughly 95% at the four-year limit.
- The mandate checks its limits only when a position is bought, with no duty to sell if the book drifts past them afterwards, so the live book can move further from the analysed composition over time, and the model book is already more conservative than the mandate requires.
Therefore, as we underwrite the asset initially, the recommended parameters are set more conservatively comparing to what operational evidence would allow.
III. Technological Risk
3.1 On-chain footprint
mWIN is deployed on Ethereum mainnet only, with no bridge or remote deployment. The token is an ERC-20 built on the permissioned Midas variant (mTokenPermissionedMinBalance), carrying three on-chain gates: a permissioned-transfer greenlist enforced at the ERC-20 transfer level, a sanctions blacklist, and a global pause. Minting and redemption route through dedicated vaults that read price from a NAV feed chain anchored to a single administrator-submitted NAV source. The cluster shares its role registry, upgrade admin, and timelock with other deployments on the Midas platform.
| Component | Address | Role |
|---|---|---|
| mWIN token (proxy) | 0x4E72025984424E52838cf8953E2863eFf036B67A |
ERC-20 (mTokenPermissionedMinBalance), 18 decimals, with permissioned-transfer greenlist, sanctions blacklist, and global pause |
| NAV custom-aggregator feed (proxy) | 0x1725A66D810C0775f6B3B0FD85646D371dA19517 |
Manual NAV oracle (mWIN/USD, 8 decimals), administrator-submitted |
| Data feed wrapper (proxy) | 0xa27c1658730e4FAFb7fB8B257a64BbB6A0ea4077 |
Chainlink-compatible (AggregatorV3) wrapper the vaults consume |
| Chainlink mWIN NAV feed (proxy) | 0x3EC0233530c548Ab984eb06BEE8a2E404aeA7557 |
MWIN NAV, 8 decimals, reported by a Chainlink oracle network through an OCR2 aggregator |
| Deposit vault (proxy) | 0xF7F1b944FCDe7805F6Ef3088817145d2eB667db4 |
USDC or PYUSD to mWIN instant mint path |
| Redemption vault with swapper (proxy) | 0x605704d7b36d1677a8d242ded68eD505523c7924 |
mWIN to USDC or PYUSD redemption on the MSL path, routes through the mTBILL USTB vault via an external liquidity provider |
| Redemption vault, Internal Liquidity Sleeve (proxy) | 0x14fECa41FB9541Fd8f61a6bA6304c5b706709fca |
mWIN to USDC redemption funded from the sleeve, redeems its own mTBILL through the mTBILL USTB vault when short of USDC |
| MidasAccessControl (proxy, shared) | 0x0312A9D1Ff2372DDEdCBB21e4B6389aFc919aC4B |
Role registry granting and revoking every M_WIN_* role |
| ProxyAdmin (shared) | 0xbf25b58cB8DfaD688F7BcB2b87D71C23A6600AaC |
Upgrade admin for every mWIN proxy above |
| TimelockController (shared) | 0xE3EEe3e0D2398799C884a47FC40C029C8e241852 |
Owner of the ProxyAdmin, 48-hour minimum delay |
| MSL liquidity provider | 0xee382068364Ed944D72183CdF83DE5630DB59ccB |
Off-portfolio account funding instant redemptions on the swapper path only |
mWIN is built on the Midas permissioned token variant, so the greenlist gate sits at the transfer level on top of the deposit and redemption gates present in earlier Midas tokens. Three features of the cluster carry through the rest of the technological and pricing analysis.
- Two redemption vaults and no separate standard redemption vault. The swapper draws on an external MSL liquidity provider, and the sleeve vault holds the compartment’s own mTBILL. Both are built on the same Midas redemption base and both sit under the shared ProxyAdmin and its 48-hour timelock. Both burn mWIN, the swapper lists USDC and PayPal USD (PYUSD) as payout tokens, and the sleeve vault lists USDC only.
- A single NAV feed. The deposit vault mints at NAV and the redemption vault settles at NAV less the instant fee, so the discount sits in the vault fee.
- Instant redemption liquidity routed through the mTBILL USTB redemption vault, with no standalone credit line. The swapper’s
liquidityProviderandmTbillRedemptionVaulttargets are non-timelocked and can change immediately, as is the sleeve vault’sredemptionVaulttarget, which its vault admin can reset in one transaction.
3.2 Upgrade architecture
7 functional contracts are upgradeable transparent proxies.
| Proxy component | Implementation |
|---|---|
| mWIN token | 0x3421478BdE3ce905d85fE24682FA8CafdFF3E44a |
| NAV custom-aggregator feed | 0x7156b140AaD6999F48767Db6687F46e8Df331C4C |
| Data feed wrapper | 0x1D0CB5685791F6E9ABc1B876E3b9017F8aa1807c |
| Deposit vault | 0x30eA22780397d82116b905E7471cCA458aaf6053 |
| Redemption vault with swapper | 0xeb21dB42a06Dc73353C6fa2956f5D8dAdCC0879c |
| Redemption vault, Internal Liquidity Sleeve | 0x765F5c63961325102BFcbC92a9f3fd0193dD4D90 |
| MidasAccessControl | 0xDd5a54bA2aB379A5e642c58F98aD793A183960E2 |
Every admin slot resolves to the ProxyAdmin, which is shared across the Midas platform deployments. Its owner() returns the TimelockController, which is therefore the only address that can authorise an upgrade or admin action on any of these proxies. The TimelockController carries a minimum delay of 48 hours and self-administers its TIMELOCK_ADMIN_ROLE, the standard OpenZeppelin pattern under which changes to the timelock’s own role membership must queue through the same 48-hour delay, with no privileged escape path.
The cluster was deployed on June 3, 2026 and first seeded the following day. The token implementation has been upgraded twice since then, on June 22, 2026 and on July 28, 2026, and the NAV custom-aggregator implementation once, on July 28, 2026. The deposit vault and the redemption vault with swapper were each upgraded once, on June 15, 2026. The Internal Liquidity Sleeve vault was deployed on July 2, 2026 and has not been upgraded. The data feed wrapper has not changed. Every upgrade ran through the ProxyAdmin-to-Timelock chain and its 48-hour delay. The code now running in the transfer path and in the priced path dates from July 28, 2026.
The 48-hour timelock does not extend to role grants and revocations, to vault parameter changes such as fees and daily limits, or to the manual NAV price submission. Those paths execute immediately under the access-control roles.
3.3 Audit history and bytecode lineage
The mWIN token contract extends mTokenPermissionedMinBalance, the shared Midas permissioned-token base.
Source: LlamaRisk, August 27, 2026
The mWIN-specific contracts (mWIN, MWinDepositVault, MWinRedemptionVaultWithSwapper, MWinDataFeed, MWinCustomAggregatorFeed, and MWinMidasAccessControlRoles) are thin product subclasses that set only names, symbols, and role identifiers over the shared base.
Four public audits cover the broader Midas codebase that the inheritance chain reuses.
| Audit | Date | Coverage | Outcome |
|---|---|---|---|
| Hacken (first) | Sep 2023 | Original mTBILL vaults | No open findings after remediation |
| Hacken (second) | Jan 2024 | mTBILL vault, December 2023 scope | 1 High and 2 Medium resolved; 1 Low (burn-from-holder without prior notice) accepted by design |
| Sherlock (first) | May 2024 | mTBILL upgraded contracts | Storage-gap inconsistency fixed; pause-authority documentation added |
| Sherlock (second) | Aug 2024 | Minter / Redeemer and BUIDL variant | BUIDL-path denial-of-service fixed; further storage-gap adjustments |
The transfer gate itself was reviewed privately. Côme du Crest reviewed the mTokenPermissioned base at pull request 194 of midas-apps/contracts, commit ff4d994, reporting no issues, and the Information Memorandum lists them with an internal reviewer at Gnosis as the architecture’s private reviewers. Both reviews are private.
Every implementation the cluster runs postdates every attestation. The four public audits are two to three years old and cover the mTBILL-era vaults and the Minter/Redeemer family, not the mWIN contracts. The live vault implementations date from June 15 and July 2, 2026, and the token and NAV aggregator implementations from July 28, 2026. That July 28 token implementation added the minimum-balance enforcement to the transfer path, newer than even the private review.
The Midas bug bounty is live on Sherlock and Cantina, paying up to $500,000 in USDC for a critical vulnerability across smart contracts, configuration, and front-end. No unfixed critical findings are reported across the public audits or the bounty.
3.4 Access control model
Role management is centralised in MidasAccessControl, shared with other Midas deployments. The registry is OpenZeppelin’s AccessControlUpgradeable with three changes:
- Deployer bundle. At initialisation one account received the top-level administrator, vault-admin, list-operator, and token-operator roles in a single call. The roles were then distributed to the operating wallets.
- Re-parented list roles. The vault-level greenlist and blacklist membership roles are administered by their operator roles, so list changes do not need the default administrator. The token-level greenlist that gates ERC-20 transfers (
M_WIN_GREENLISTED_ROLE) is not re-parented: every grant and revocation of it runs through the default administrator, and all 59 grants on the deployed registry were made by it. - No renunciation. A call to
renounceRolereverts withMAC: Forbidden. A compromised signer cannot self-clear, and only the role’s administrator can move a role, by granting then revoking.
Role changes are not timelocked. The 48-hour delay covers contract upgrades only, so the only safety on a role change is the off-chain control at the role holder.
The mWIN-namespaced roles are unreachable from another Midas deployment’s admin roles. The blacklist is the exception. BLACKLISTED_ROLE and BLACKLIST_OPERATOR_ROLE are Midas-wide, the token’s transfer hook enforces them on both the sender and the recipient of every transfer, and any of the role’s 52 holders, all externally-owned accounts onboarded across the whole product line, can freeze an address’s mWIN in a single transaction.
| Role | Purpose | Holder (on-chain) |
|---|---|---|
DEFAULT_ADMIN_ROLE |
Grant and revoke every role in the registry, including the token-level greenlist | Externally-owned account 0xd4195CF4df289a4748C1A7B6dDBE770e27bA1227 and the Midas Proxy and ACL Admin Safe 0xB60842E9DaBCd1C52e354ac30E82a97661cB7E89, a nominal 1-of-3 whose owners are two multi-party-computation systems at issuer-stated 4-of-7 quorums and a Team Signer Safe at an on-chain-verified 3-of-7 |
BLACKLIST_OPERATOR_ROLE |
Manage blacklist membership, generic Midas-wide role, not namespaced to mWIN | 52 externally-owned accounts across the Midas product line, among them the token role-manager EOA 0x20D4CeD0EFac28517C1b0a06F98B1180F28f5125 |
GREENLIST_OPERATOR_ROLE |
Manage vault-level greenlist membership, the gate no vault enforces, generic Midas-wide role | 52 externally-owned accounts, among them the token role-manager EOA 0x20D4CeD0EFac28517C1b0a06F98B1180F28f5125 |
BLACKLISTED_ROLE |
Revokes transfer and redemption rights, generic Midas-wide role administered by BLACKLIST_OPERATOR_ROLE |
(membership gate) |
GREENLISTED_ROLE |
Vault-level gate, not enforced: the vaults read greenlistEnabled as false |
(membership gate) |
M_WIN_GREENLISTED_ROLE |
Token-level gate: required for ERC-20 transfers, administered by DEFAULT_ADMIN_ROLE |
(membership gate) |
M_WIN_MINT_OPERATOR_ROLE |
Mint mWIN | Deposit vault and the token role-manager EOA 0x20D4CeD0EFac28517C1b0a06F98B1180F28f5125 |
M_WIN_BURN_OPERATOR_ROLE |
Burn mWIN from any holder, no allowance required | Both redemption vaults and the token role-manager EOA 0x20D4CeD0EFac28517C1b0a06F98B1180F28f5125 |
M_WIN_PAUSE_OPERATOR_ROLE |
Global token pause | Token role-manager EOA |
M_WIN_DEPOSIT_VAULT_ADMIN_ROLE |
Configure and pause the deposit vault | Vault role-manager EOA |
M_WIN_REDEMPTION_VAULT_ADMIN_ROLE |
Configure and pause the redemption vault and swapper | Vault role-manager EOA |
M_WIN_CUSTOM_AGGREGATOR_FEED_ADMIN_ROLE |
Manage the NAV custom-aggregator feed | Oracle role-manager EOA |
Two powers matter most for a Horizon listing, because either one freezes the Aave Pool’s mWIN balance in a single transaction with no on-chain delay.
- Blacklisting the Pool, open to 52 keys, the token role-manager wallet among them.
- Revoking the Pool’s token-level greenlist, a default-administrator power. The default administrator can also grant or revoke the mint, burn, pause, vault-admin, feed-admin, and minimum-balance-exemption roles, none of it timelocked.
The split of the two freeze powers across different holders is not a segregation control. It places the second freeze power at the root of the registry.
Off-chain the two freeze paths carry different protection.
| Holder | Freeze power | Off-chain control |
|---|---|---|
| Token role-manager | Blacklist the Pool, pause the token | Fordefi MPC wallet, 4-of-8 quorum, one-signer fast path for blacklist and pause, 3 signers to reverse |
| 51 further blacklist-operator accounts | Blacklist the Pool | None disclosed |
| Default administrator, externally-owned account | Revoke the Pool’s greenlist, grant or revoke every role | None disclosed, a plain key |
| Default administrator, Safe | Revoke the Pool’s greenlist, grant or revoke every role | Nominal 1-of-3 whose owners are two multi-party-computation systems at issuer-stated 4-of-7 quorums and a Team Signer Safe at an on-chain-verified 3-of-7 |
Source: Documents shared under NDA, August 2026, and MidasAccessControl, August 17, 2026
The greenlist freeze path and the registry root therefore sit behind a weaker control than the token role-manager wallet, which itself clears a blacklisting on one signer. The blacklist freeze is additionally reachable from 51 further keys for which no control is described at all.
3.5 Pause surface
Two pause mechanisms coexist.
- Token-level pause. Halts all ERC-20 balance state changes, which means transfers, mints, burns, and any liquidation against the token. Its authority is
M_WIN_PAUSE_OPERATOR_ROLE, held on-chain by the token role-manager wallet. - Vault-level pause. Adds a per-selector mask, so a single entry point such as instant redemption can be paused while other paths on the same vault keep operating. Its authority is the relevant vault-admin role, held by the vault role-manager wallet.
The token is not paused at the time of this assessment. A token-level pause is the more consequential of the two for a collateral position, because while it is active a liquidation cannot transfer the seized mWIN and the position cannot be unwound on-chain until the pause is lifted. Like every role action, a pause executes immediately with no on-chain delay.
3.6 Token-level semantics
The defining behaviour of mWIN at the token level is the permissioned transfer. The _beforeTokenTransfer hook of mTokenPermissionedMinBalance requires M_WIN_GREENLISTED_ROLE on the recipient of every transfer or mint, on cross-wallet transfers on the sender as well, and enforces the blacklist on both parties on top. The combined behaviour:
- Transfers between two greenlisted wallets that are not blacklisted: pass.
- Transfers or mints to a wallet without
M_WIN_GREENLISTED_ROLE: revert. - Burns: skip the greenlist check, so burn-from-any-holder remains functional.
- Blacklisted addresses cannot send or receive.
- Every transfer, mint, or burn must leave both sides holding either nothing or at least one whole mWIN, roughly $130,966 at the live net asset value, unless the address holds
M_WIN_MIN_BALANCE_EXEMPT_ROLE.
The minimum balance reaches partial liquidation directly. A seizure that would strand a sub-token remainder on either side reverts unless the address concerned is exempt, so the smallest position a non-exempt wallet can carry, and the smallest remainder a partial liquidation can leave, is one note. Six addresses hold the exemption, the swapper’s external liquidity provider among them and none of the three mWIN vaults, and the exemption changes by a role action that executes immediately.
The burn function takes an explicit holder address and does not check allowances, so the burn-operator role can destroy any holder’s mWIN without approval, flagged in Hacken’s December 2023 audit as low severity and accepted by design. Metadata setting is restricted to the top-level administrator role.
3.7 Oracle architecture
mWIN prices off a two-contract feed chain. A single network-asset-value custom aggregator holds the manually submitted mWIN/USD price, and a Chainlink-compatible data-feed wrapper reads that aggregator and exposes the base-18 price the vaults consume. Chainlink has deployed an mWIN NAV feed on Ethereum mainnet, reported by a Chainlink oracle network through an OCR2 aggregator, which at the time of this assessment carries the same net asset value as the Midas aggregator. There is no separately deployed mint feed at a positive offset and no redemption feed at a negative offset, so the deposit vault mints at the net asset value the aggregator reports and the redemption vault settles at that same value less the instant fee.
The custom aggregator is an AggregatorV3-style contract whose answer the oracle role-manager wallet holding M_WIN_CUSTOM_AGGREGATOR_FEED_ADMIN_ROLE submits each round, striking one round each business day. Three on-chain circuit-breakers bound a submission:
- a per-round deviation band of plus-or-minus 27%,
- a floor of $90,000 and a ceiling of $140,000 on the answer,
- a staleness tolerance of 30 days carried on the data-feed wrapper.
The attestation pipeline is a Chainlink Cross-Chain Runtime Environment (CRE) proof-of-reserve that has confirmed overcollateralization since August 3, 2026, computed from claims Midas provides using Northern Trust’s net asset value. The vLayer-notarised email flow described in the offering materials is not the deployed pipeline, and Northern Trust application programming interface (API) access that would let the workflow pull the net asset value directly is pending.
3.8 Minting and redemption vaults
Three vaults bracket the token. A deposit vault handles stablecoin-in (USDC or PayPal USD) and mWIN-out. Two redemption vaults, the swapper and the Internal Liquidity Sleeve vault, handle mWIN-in and stablecoin-out.
Source: LlamaRisk, September 17, 2026
The deposit vault exposes an instant entry point that settles in one transaction and a request-based entry point that queues for back-office processing. Every entry passes three checks: the caller must not hold the blacklist role, must clear the Chainalysis sanctions oracle, and the relevant selector must not be paused. The instant mint limit is 155 mWIN a day, roughly $20.2 million,
The redemption vault with swapper settles an instant redemption from its own balance of the output asset where that balance is sufficient, and otherwise activates the swapper path. The redeemer’s mWIN is transferred to the external liquidity provider, the provider supplies mTBILL in return at the two oracle rates, and the vault redeems that mTBILL for USDC through mTBILL’s redemption vault, so the externally sourced USDC is raised by redeeming short-dated treasury-bill collateral.
The blacklist and sanctions checks apply on both branches. The vault’s own greenlist check greenlistEnabled is false. A redeemer must still be greenlisted in practice, because only a wallet holding M_WIN_GREENLISTED_ROLE can hold mWIN at all, and the swapper branch re-tests both parties at the token level when the mWIN moves to the provider. The gap is the payout destination. The entry point that names a separate recipient no longer screens that recipient, so redemption proceeds can be paid to a wallet that has never been greenlisted, subject only to the blacklist and sanctions checks.
On the deployed contract the payment tokens are USDC and PayPal USD, the instant fee is 0.5%, the daily limit is 4 mWIN, roughly $524,000 at the live net asset value, and a fiat settlement adds 0.1% plus a flat $30. The configured minimum of roughly $1.00 of mWIN is not the operative floor, because the token leaves any redeemer holding either nothing or at least one whole mWIN. The slippage bound is 0.2%, compared in base-18 USD-equivalent units on both sides of the swap, so the denomination error Sherlock flagged in the BUIDL variant does not apply. Request-based redemptions settle through the dedicated request-redeemer.
The sleeve vault settles without an external provider. Its verified implementation MWinRedemptionVaultWithMToken burns the redeemed mWIN and always pays USDC, never delivering mTBILL in kind. Where its own USDC is short it redeems its mTBILL through the same mTBILL redemption vault inside the transaction, a fee-exempt hop. That vault raises its USDC by redeeming Superstate USTB and carries its own daily limit of 10 million mTBILL, roughly $10.7 million, so it does not bind ahead of the two mWIN vaults’ 4 mWIN limits. The vault holds the compartment’s mTBILL, roughly $1.9 million at the live read, screens against the same Chainalysis sanctions oracle as the deposit vault, shares the request-redeemer with the swapper, and sends collected fees to the fee receiver.
The daily limits, the fees, the slippage bounds, the greenlist flags, the swapper’s liquidity provider, and both mTBILL swap targets change by a vault-admin role action that takes effect immediately, outside the 48-hour upgrade timelock, and all of them are carried in the monitoring inventory shipped with this report. At the live total value locked the redemption vaults’ 4 mWIN daily limits are the binding throttle on an exit. That immediate-change power stands to lift the instant daily limits on short notice under redemption need
IV. Parametrization Recommendations
4.1 Market Parameters
Base Analysis (Without Buffer)
Two distances govern the parameter set, both measured from a fully drawn position at the maximum loan-to-value ratio: the net asset value decline that reaches the liquidation trigger, and the decline at which the collateral no longer covers the debt plus the 8% liquidation bonus.
| Distance | Base (67% LTV / 72% LT) | GHO E-Mode (69% / 74%) |
|---|---|---|
| To the liquidation trigger | 6.94% | 6.76% |
| To bad debt | 27.6% | 25.5% |
The deepest single-day print anywhere in the ten-year proxy record is 3.43%, from the March 2020 short-duration credit dislocation, struck on a proxy carrying a 2.6-year spread duration. The distance covers that raw print roughly twice over, and covers it about 2.1 times at the live book’s 2.6-year spread duration, which matches the proxy’s, once diluted by the sleeve, roughly 3.3% as published. At the four-year spread-duration limit the same print is roughly 5.0% as published, still inside the trigger distance. Measured against the ordinary daily distribution the margin is far wider again, since the proxy record’s daily standard deviation is near 0.2% and its worst day outside 2020 is under 1%.
Multi-day drawdowns are not sized into the spread, because they do not arrive in one print. A crisis-scale slide reaches the feed as a sequence of daily strikes, positions delever or liquidate progressively as it runs. The deepest drawdown the model book has printed, the 13.32% of March 2020, is roughly 17% carried to the live book’s spread duration.
Buffer Adjustment
The Internal Liquidity Sleeve dilutes the portfolio’s mark-to-market volatility, because the published net asset value blends the Wellington book with the sleeve’s short-dated treasury-bill holding. At the 5% cap every shock reaches the feed at 95% of its size, which leaves the manager’s worst forward scenario inside the trigger at roughly 3.1% carried to the live book’s spread duration and trims the scaled March 2020 drawdown to roughly 16.1% as published, roughly 11.5 percentage points inside the bad-debt distance. At the four-year spread-duration limit the same drawdown is roughly 24.8% as published, still inside the bad-debt distance. Episode-scale shocks still cross the trigger and liquidate progressively across the prints that carry them. The sleeve also carries a yield drag. Its mTBILL holding earned 3.27% annualized over the trailing 91 days, against the book’s 4.63% net-of-fee yield, which costs roughly 7 basis points of blended accrual and about 20 basis points of looped return at the 67% loan-to-value ratio. The looping figures below are struck on that blended basis.
Looping Profitability
The looping case turns on the spread between mWIN’s blended net asset value accrual, roughly 4.56% on the model portfolio’s market yield, and the Horizon borrow rate. At the recommended 67% loan-to-value ratio and the June 30, 2026 rates, looping mWIN against RLUSD returns roughly 7.0% on equity and against USDC roughly 3.7%. The USDC loop returns below the unlevered yield, so at these rates levering against USDC reduces the return on equity rather than raising it.
| Borrow asset | Borrow rate | Net looping APY at 67% LTV (blended yield) |
|---|---|---|
| RLUSD | 3.34% | ~7.0% |
| USDC | 4.98% | ~3.7% |
| Unlevered | n/a | ~4.56% |
Source: Aave, August 2026
The economics remain viable for the cheaper borrowable and negative for USDC at present rates. The figures carry the sleeve drag measured above, roughly 20 basis points of looped return, which the leverage partially offsets.
Supply Cap
The initial cap is 80 mWIN, roughly $10.5 million at the live net asset value of $131,064 per token, sized against redemption capacity and the expected lined-up demand. A filled cap clears through the standard fee-free route within one to two business days of the portfolio-sale settlement. Any increase is set on the same basis at the then-current supply.
General Market Parameters
Aave Horizon is an Aave v3.3 instance, and its live listings carry their collateral parameters in the base reserve configuration with a GHO e-mode adding a small collateral-efficiency uplift.
| Parameter | LlamaRisk recommendation |
|---|---|
| LTV | 67% |
| Liquidation Threshold | 72% |
| Liquidation Bonus | 8% |
| Reserve Factor | - |
| Optimal Utilization | - |
| Base Variable Borrow Rate | - |
| Variable Slope1 | - |
| Variable Slope2 | - |
| Supply Cap | 80 mWIN |
| Borrow Cap | - |
| Borrowing Enabled | No |
| Usage as Collateral | Yes |
| Isolation Mode | No |
| Siloed Borrowing | No |
| Borrowable in Isolation | No |
| E-Mode | mWIN GHO |
| Flashloan Enabled | No |
GHO E-Mode
| Asset | mWIN | GHO |
|---|---|---|
| Collateral | Yes | No |
| Borrowable | No | Yes |
| LTV | 69% | - |
| LT | 74% | - |
| Liquidation Bonus | 8% | - |
V. Pricing Setup
mWIN prices off the single administrator-submitted NAV feed read on-chain as stated in the technological review. The setup is the standard one used across the Horizon listings: the daily NAV struck by Northern Trust is published through the Chainlink mWIN NAV feed the Aave price source reads, and wrapped in LlamaGuard safety bounds parameterized to the asset. The feed carries the same value as the issuer’s administrator-submitted aggregator. The setup does not verify the source, because the aggregator round is an administrator submission and the Chainlink CRE proof-of-reserve runs on Midas-provided claims, so the bounds and the freshness window constrain how far a published value may move and how stale it may be. The analysis values a Pool-held or liquidator-held position at published NAV.
Two design features simplify the setup:
- mWIN mints atomically at the full reported NAV with no mint fee and redeems the full fee-adjusted value in one transfer, so no entry or exit haircut bears on how collateral is valued while a position is open.
- NAV publishes daily, so upward moves are small increments of accrued yield that are safe to release gradually, downward moves are losses that must reflect promptly, and the cadence already compresses the recognition lag the design has to defend.
5.1 Feed Behaviour and Bounds
The oracle steps to each newly published NAV when it is written and holds flat until the next write. The published NAV is net of the management fee and blends the Wellington book with the 5% mTBILL sleeve, so the feed accrues at roughly 4.56%, an average daily accrual of roughly 1.8 basis points. The full daily accrual lands in one block at each strike, but that block is under 2 basis points, so the step-function arbitrage the bounds must contain is negligible.
LlamaGuard wraps the feed with the parameters below, each sized to mWIN’s daily step and the drawdown magnitudes:
| Parameter | Value |
|---|---|
maxExpectedApy |
800 bps (8%) |
upperBoundTolerance |
50 bps (0.5%) |
maxDiscount |
250 bps (2.5%) |
lowerBoundTolerance |
100 bps (1%) |
lookbackWindowSize |
4 |
Disclaimer
This review was independently prepared by LlamaRisk, a DeFi risk service provider funded in part by the Aave DAO. LlamaRisk is not directly affiliated with the protocol(s) reviewed in this assessment and did not receive any compensation from the protocol(s) or their affiliated entities for this work.
The information provided should not be construed as legal, financial, tax, or professional advice.