Method
A verdict is only worth what its procedure is worth.
Uniswap’s documentation warns that an auction can be configured with malicious parameters and tells every reader to validate them. This is how that validation runs here, in the order it runs — the order is the safety property.
The chain is the only source
Every figure is fetched from a Robinhood Chain RPC directly — logs, contract code, balances, storage reads. No indexer and no price API is treated as evidence. When a node cannot be reached, that is an error state, never an empty result: unknown is not the same as safe.
Deterministic checks set a ceiling
The auction parameters are decoded and checked: the LP allocation schedule bracket by bracket, the position plan, the committed pool, the recipient's code and history, and the currency's identity against the issuer registry. Whatever those facts allow becomes the best verdict this launch can receive.
A judge reasons below the ceiling
An agent takes the factsheet and classifies what a script cannot settle: whether a custom hook's economics look hostile, what the deployer's provenance suggests, how much of the picture is genuinely unresolved. It may lower a verdict. The code that applies its answer cannot raise one.
The verdict is published with its evidence
Three states, no scores: PASS, FLAG, FAIL. Each verdict carries the findings it rests on and the transaction it came from, so a reader can disagree with the reasoning without having to trust the reader.
Context is published, but it does not warn
Most launches pay an ordinary wallet and price the auction in ETH. Those facts are worth stating and are not warnings, so they are recorded as context and cannot move a verdict. Warnings are kept for what is genuinely unusual against the whole record — a recipient with almost no history, a raise that keeps most of itself back from liquidity. A verdict layer earns attention by being quiet.
What is guaranteed
- The ceiling can never be raised
- A hard failure caps the verdict at FAIL and a warning caps it at FLAG. This is enforced in code and tested exhaustively, so a model having a bad day — or one steered by a hostile token name — cannot talk a launch up.
- The model writes none of the words you read
- It returns a verdict, a set of our own finding keys, and classification labels from a closed list. Every sentence on a card or a page is ours. There is no path for injected copy to reach a reader.
- Claims of officialdom are checked, not believed
- A currency is an official Robinhood stock token only if its address is in the issuer's registry. A name that claims the pattern from an unregistered address is called an impostor — and only when the registry was actually readable.
- A warning has to be rare to be worth reading
- Severity is set against the whole record, not against intuition. Paying an ordinary wallet is what most launches do, so it is published as context and cannot move a verdict; a wallet with almost no history is unusual, so it warns. Thresholds are read off the measured distribution and change when the distribution does.
- GAVEL never touches funds
- No custody, no keys, no wallet connection anywhere in the system. If every GAVEL process stopped right now, no one's money would be affected, because there is no path from this software to money.
What GAVEL cannot see: intent. It reads parameters and provenance. A launch that passes every check can still fail, and a flagged launch is not an accusation of fraud — it is a statement that something in the configuration deserves a second look before money moves.