Limits and what Sentinel misses
Read this before you rely on a grade.
What it is good at
- Known bug classes with a recognisable shape: reentrancy, unchecked calls, missing access control, unsafe arithmetic in old compilers,
tx.originauth, unbounded loops. - Second opinions: the static analyser and the models check each other, and criticals need two of three models to agree.
- Speed and cost: minutes and cents, so it can run on every version you deploy.
What it will miss
- Intent. Sentinel does not know what your contract is supposed to do. A fee that goes to the wrong address on purpose is not a bug to it.
- Economic and cross-protocol attacks. Manipulation that needs a model of other protocols' liquidity, oracles or incentives is mostly out of reach.
- Off-chain parts. Keepers, multisig signers, front ends, deployment scripts and key management are not audited.
- Upgrades. An audit covers the code that was submitted. A proxy that is upgraded later is a different contract; audit the new implementation.
- Long or unusual code. Files over 900 lines are split per contract, and models see each part with the Slither results that touch it, not the whole system at once.
- Model mistakes. Models produce false positives and miss real issues. The benchmark numbers on the landing page are published as measured.
Hard limits in v0.1
| Limit | Value |
|---|---|
| Upload size | 2 MB total |
| Lines of code | 600 / 3,000 / 10,000 by tier |
| Languages | Solidity. Vyper is rejected. |
| Compilation | Must compile with the pragma's solc. A compile error ends the audit. |
| Address input | Verified source on Etherscan or Blockscout only |
| Network | Base Sepolia attestations; mainnet is out of scope for v0.1 |
| Refunds | Manual, by the operator, for jobs that fail after payment |
When a part was not analysed
If a model fails to answer in the required format twice, that part of the code is marked not analysed by a model in the report, with the error. Slither still ran on it. The grade does not hide this; read the warning at the top of the report.
What to do with a good grade
Ship your testnet launch. Before mainnet with real money, get a manual audit from people who read your design, and keep Sentinel in your deploy pipeline for every change after.
Owner powers are graded as high
A contract whose deployer keeps powers to mint, block transfers, raise the tax to 100% or move holders' balances is found reliably, and quorum confirms those findings, but they are usually rated high, not critical. Four such highs cost 20 points, so a deliberately malicious token can still grade B (the RugPull.sol demo on the landing does). Read the findings, not only the letter, and treat any centralization finding that lets one key take user funds as disqualifying.