Three questions, not one label
exit window: decided by someone elseWhat does “L2” tell you about where your money sits?
Almost nothing, and the industry is not being careful about this. The label covers an optimistic rollup with a live fraud-proof system, a zk rollup with a proven verifier, and a chain that is, in every operational sense, a single company running a database and posting a hash to Ethereum once in a while. All three call themselves L2. Two of them will let you withdraw without asking anyone’s permission. One will not.
So I stopped using the label as a safety rating and started asking one question instead.
If the operator, the sequencer, and the security council all vanished tomorrow — no announcement, no warning, no cooperation — could you get your money back using only what is on Ethereum?
That is the walkaway test, and it is the only question in this space whose answer does not depend on trusting the party you are evaluating.

Diagnosis: Three Places Trust Hides
When I look at a rollup now, I do not read the marketing page. I look for three specific facts, because trust can only hide in three places.
1) Where is the state root, and can you prove it wrong? A rollup posts state commitments to L1. The question is what happens when a commitment is false. With a validity proof, the commitment cannot be false — the proof is checked, and an invalid one is rejected by the L1 contract. With a fraud proof, the commitment is assumed true until somebody challenges it, which means the security property is “at least one honest watcher is online and funded.”
2) Where is the data? Not “where is the data advertised to be.” Where can you actually recover the state from, if the operator’s servers are gone? On-chain blob or calldata means you can rebuild it. An external DA layer or a “data availability committee” means you are trusting a second party that nobody made you evaluate.
3) Who can change the rules without your consent? Every rollup has an upgrade path. The question is whether that path runs through a contract with a delay, or through a multisig that can execute in the next block.
withdrawable = state root on L1 × data available × no privileged upgrade without an exit window
Miss any one of those three and the walkaway test fails — and it fails in a way that only becomes visible on the day you need it.
The Stages Framework, Read Carefully
L2Beat’s stages are the closest thing the ecosystem has to a standard, and they map almost exactly onto those three questions. Here is what each stage actually requires, stripped of framing:
| Stage 0 | Stage 1 | Stage 2 | |
|---|---|---|---|
| Self-identifies as rollup | Required | Required | Required |
| State roots on L1 | Required | Required | Required |
| Data availability on L1 | Required | Required | Required |
| Working proof system | Required | Required | Required |
| Who can submit a fraud proof | ≥ 5 external actors | ≥ 5 external actors | Permissionless, anyone |
| Security council | None required | ≥ 8 members, 75% threshold | On-chain-provable bugs only |
| Challenge period (optimistic) | — | ≥ 5 days | ≥ 5 days |
| Exit window for upgrades | — | ≥ 5 days if initiated outside the council | ≥ 30 days in all cases |
Now read row six and row eight together, because that is where the headline version of this framework goes wrong.
Stage 1 allows the security council to upgrade instantly. The council is a multisig of at least 8 members with a 75% threshold — and if those 75% agree, they can push a change in the next block, with no exit window at all. That is not a loophole; it is the design, on the theory that a council which can act fast can also stop a bug before it drains the bridge.
The consequence is uncomfortable: a Stage 1 rollup can have an exit window of literally “none” and still be Stage 1. OP Mainnet’s own L2Beat page lists exactly that — Stage 1, fraud proofs, on-chain data availability, exit window: none, because the security council has instant upgrade power. The protection you get from Stage 1 is not “you always have time to leave.” It is “the only way to block your withdrawal indefinitely is to compromise roughly three quarters of a known, named, jurisdiction-diverse set of people.”

The Sequencer Problem the Stages Do Not Measure
There is a fourth trust assumption that the stage framework only partially touches, and it is the one you interact with every day.
Almost every live L2 runs a single, permissioned sequencer. Arbitrum runs on Offchain Labs’ sequencer; OP Mainnet on the OP Foundation’s; Base on Coinbase’s; zkSync Era on Matter Labs’. One operator decides the order of your transactions, whether to include them at all, and how long to make you wait. All of those teams have had multi-hour sequencer outages.
A sequencer cannot steal your funds — that is what the proof system and the bridge contract are for — but it can censor you, reorder you for profit, or simply stop. The mitigation is a force-include path: submit your transaction directly to the L1 contract and it gets queued regardless of what the sequencer thinks. That caps the worst case at “you wait, and you pay L1 prices to prove a point.”
Which is the honest description of the user experience on a Stage 1 rollup in 2026: cheap and fast when the operator is online and honest, slow and expensive when you need to route around it, and provably survivable in the case where neither holds.

What the Label Is Worth
Here is my working summary, and it is deliberately unflattering to the category.
Stage 0 is “trust the operator, but the operator posts receipts.” Stage 1 is “trust 75% of a named council, plus at least one honest watcher.” Stage 2 is “trust the code.” As of now, no major rollup is at Stage 2. The biggest ones — Arbitrum One, Base, OP Mainnet — are Stage 1, which is a genuine and recent achievement, and is also a statement that three quarters of a known group of people could, if they agreed, change the rules on you tonight.
That is not a scandal. It is a stage, and stages are supposed to be climbed. But it does mean that “it’s an L2” is a statement about architecture, not about your risk — and the two have been conflated for long enough that people are surprised when a council upgrade moves their funds faster than they can bridge out.
Stage 1: ≥8 council members at 75%, ≥5-day challenge period, ≥5-day exit window only for upgrades initiated outside the council · Stage 2: permissionless fraud proofs, ≥30-day exit, council limited to on-chain-detectable bugs
So when the next chain launches and calls itself an L2 with a big number and a token, ask three questions:
- Where is the state root, and can you prove it wrong? If the answer is “a fraud proof that someone is watching,” ask who is watching and whether they are paid to.
- Where is the data, and can you rebuild the state from it alone? Not “is there a DA layer,” but “could I reconstruct this chain with the operator’s servers unplugged.”
- Who can change the rules while you sleep? If the answer is “a multisig that can act in the next block,” then you are not using a trustless system. You are using a very good custodial system with a public audit trail — and the difference between those two things only shows up on the day you need it, which is precisely the day you cannot check.
(The End)


