ETH Strategy: Parallel AI and AskSurf gave opposite assessments of the same DeFi project

Comparison of two AI audits of the DeFi project ETH Strategy yielded nearly opposite results. AskSurf proved more cautious, while Parallel AI was more detailed but made a critical error in basic code verification.

How ETH

Strategy works — it is a DeFi protocol on Ethereum that promises users ETH exposure with leverage, but without the classical liquidation risk. The mechanics are built around Long Bonds: a user deposits USDC, receives a debt token CDT and an NFT option, while the protocol itself effectively borrows capital at zero rate, as it compensates for it through the option structure. The project's treasury grows through bond issuance and placing ETH in lending protocols like Morpho.

At the time of analysis, the STRAT token traded around $0.14, which is 83.8% below its historical maximum.

Market capitalization was estimated at approximately $690 thousand, and TVL at $3.8 million. This is where divergences between the models began.

Both recognized the idea as unusual and understandable, comparing ETH Strategy to a DeFi version of MicroStrategy. But conclusions on tokenomics diverged: Parallel AI noted the risk that 75.5% of tokens would be unlocked in just two months, whereas AskSurf did not find this data.

Where the models diverged

The most important difference concerned not the assessment of the concept, but basic infrastructure verification. AskSurf gave the project a low rating for code openness and directly indicated that only the token is visible in public access. Parallel AI, conversely, gave a high score and reported that ETH Strategy has an open GitHub repository.

Upon manual verification, it turned out that the model referred to the wrong project: the found repository belonged to ETHXR, not STRAT. That is, the system erred already at the level of object identification. This inaccuracy is important not only in itself.

If the model confuses repositories, then further the entire chain of conclusions about transparency, development quality, and risks becomes questionable. In reality, ETH Strategy had no explicit link to a GitHub core protocol in its documentation, and from confirmed external reviews, an audit by Nethermind for the ESPN vault product was mentioned. Audits of the main protocol were promised by the team to be published later, with the launch of permissionless bonding.

Against this backdrop, the more cautious conclusion of AskSurf turned out to be closer to the facts.

"AI audits can be useful as a tool for initial analysis, but still

require manual verification."

What practice showed

The author of the comparison did not limit himself to the models' answers and separately tested the ETH Strategy application. In the interface, STRAT staking and unstaking are available, as well as mint and redeem for ESPN — a separate vault used as a yield product. The stated APR at the time of verification was 18.01%. Visually, the application looked functional and clear, but it was the manual walkthrough of scenarios that revealed several important details that are not visible from beautiful AI summaries.

• Working with the vault requires USDS, not USDC, so an additional swap is needed first The interface does not show the exact amount the user will receive when minting and redeeming In ESPN there is no explicit calculation of how much USDS will arrive upon redemption * Approve is requested for an adequate amount, not for unlimited max Based on the verification results, the interface received a conditional 4 out of 5: it can be used, but transparency at critical points is still lacking. For DeFi, this is not a cosmetic problem. When a user does not see the expected amount on the output, they understand the risk worse, depend more on external calculators, and more often make decisions almost blindly. These are the kinds of details that usually separate a beautiful token presentation from a mature product.

What this means

The comparison of AskSurf and Parallel AI shows a simple thing: AI audits already work for quick initial filtering, but do not replace full verification. If two models give opposite conclusions on the same project, this is not a reason to choose the one that sounds more confident. This is a signal to manually re-verify the tokenomics, the origin of the repository, code audit, and real user scenarios before any decisions about investments or integration.