Four real exploits. 48 graded fixes.
Each case is a contract that was exploited on chain, frozen one block before the attack. A fix counts only if the exploit fails and every hidden security and legitimate-use obligation passes. We attack our own verifier with 48 test fixes, the original contracts included as baselines: 15 of them stop the exploit and are still not repairs.
MCAI
The tax wallet could move anyone’s tokens without their approval.
GoldReserve
Profit followed the NFT. The record of what was paid followed the address.
Bitallx
A batch payout never checked what it paid against what was funded.
NGP
Every sell pulled fees out of the pool and re-priced it mid-trade.
Losses as listed by DeFiHackLabs.
A case ships only when it can catch a near-miss.
Four exploit classes today, all 2025 incidents from DeFiHackLabs, on verified contract source. Each clears the same bar before any fix is graded against it.
The exploit reproduces offline
Recorded from the chain once, then replayed with the network off until the exploit and every check run without missing state. The recording is the only data source.
Security obligations state the class
Each fails on the original contract by construction, so it can tell a repair from a non-repair, including the near-misses the exploit alone would pass.
Legitimate use is written down
Hand-written tests of what ordinary holders, traders and operators do, plus generated checks that the original functions survive.
The controls grade as intended
The reference repair and an independently written alternative must both be solved. Incomplete and destructive fixes must fail for the right reason; a build failure must be inconclusive.
The harness is locked
The exploit test and every hidden suite are hash-locked and re-verified before each run. A patch may only touch the contract source.
The grade reproduces
Each case’s reference and original grades hash identically on the local backend and in the no-network container, and every control re-grades to the same hashes on GitHub’s Linux runner, on every push.
Every verdict has a receipt.
All 48 grades with their checks, patches and hashes, and the method behind them.