ABCAI-moderator filters spam governance proposals so that team does not have to waste their precious time in these boring activities.
Right now, when a scam proposal is published on-chain, the team has to manually removed it, wasting their time. Also it's a big problem for users that could get scammed by those fraudulent proposals and this could also bring a bad reputation for the project itself.
What we do is to make use of the latest technology of ABCI++, Cosmos SDK v0.50 and AI in order to automatically detect scam proposals and delete them just after one block from when they are published.
A little overview of ABCI++ is needed to understand how we code this project.
ABCI++ is the bridge between CometBFT, the consensus protocol, and the application itself, the state machine. Until ABCI++, app developers couldn't implement custom logic in the first phases of the creation of a block, for example using a different transaction inserting algorithm or other more complex stuff.
ABCI++ unlocks all these possibilieties. In particular there are 5 moments where CometBFT and the state machine can communicate in order to implement custm logic:
The proposer can add some data or computation in the block building.
All validators verify that the proposer is behaving properly.
All validators can add extra votes or computation in the block.
All validators verify that other validators have behaved properly during the precedence phase.
The last phase where validators can apply some custom logic.
It all happens in a two-blocks sequence:
-
ExtendVote: validators make an API call to OpenAI that returns a score between 0 from 1 that is then normalized between 0 to 100.
-
VerifyExtendVote: validators some simple sanity checks, like checking if the score is between 0 and 100.
-
PrepareProposal: proposer aggregates all the validators' scores and computes the average.
-
ProcessProposal: validators check if proposer did the work correctly.
-
PreBlocker: validators check if scam score is greater than 90% and if so they reject the proposal.
This is an example of a workflow of a scam proposal that fets published on-chain.
