On the toxicity classification of MEV transactions

Given a FB bundle containing some MEV (eg: https://flashbots-explorer.marto.lol/?block=15537389), do we have a proven way to analyze whether the MVE was toxic (sandwich/frontrun) or non-toxic (backrun)? This would be very useful to feed all the discussions on the toxicity of MEV. Essentially what we need is a way to classify a bundle into one of the following categories:

  • frontrun
  • backrun
  • sandwich
  • liquidation
  • something else

I think consensus is that it’s hard to detect frontruns if someone is trying to hide them. Current data is mostly heuristics. @taarushv should know

Pmcgoohan may have more to say - check out his zeromev.org project if you don’t know it already.


In order to identify MEV as being toxic, you first need to admit that validators reordering is what makes it toxic.

Flashbots have not seemed able to do this so far, perhaps because their business is maximizing the ability of validators to reorder transactions for profit.

But it’s not hard to grasp. Uniswap, for example, is completely time order dependent. If Alice and Bob are trying to buy the same asset, the amount they get differs depending on who executes first. There is no way around this, it is literally how Ethereum/Uniswap were designed.

It is also the reason why Flashbots continued assertion that time order is not objectively fair must be false in the case of L1 Uniswap.

Grabbing the latest example as I’m writing this:

Here you can see 2 sandwiches where the victim has been delayed by 11 secs and 1 min 22 secs respectively. Hover over arrival times and check the network latency box and you’ll see the p2p latency was around 798 ms in this case. This is so much less than 11 secs that there can be no doubt these transactions were frontrun.

By looking for reordering beyond measured network latency in this way, we can identify to a very high degree of accuracy whether txs have been frontrun or not- it’s not difficult or ambiguous.


How can one hide a frontrun?

I see your point, different definitions. I qualify a MEV transaction as toxic if it is a frontrun/sandwich while I qualify it as non-toxic if it is a backrun/liquidation. Is your fork of mev-inspect open source @pmcgoohan ?

I am trying to formalize all of this in mathematical terms, I should have something clean to share soon.

1 Like

Yes, we’re pretty much saying the same thing.

The backrun is non-toxic because it follows after the target tx. However, if that backrun was to be re-ordered so that it also frontruns another swap, that would become a frontrun. I’m interested in finding out how much that happens with arbs, for example.

A liquidation is also non-toxic, as long as it doesn’t involve frontrunning someone elses attempt to liquidate, or censoring a tx attempting to recollateralize.

Will be interested in your formalization!

Counter to my defining toxic mev as reordering for profit, I define fair order as send time order.

I do understand your preference for time order as the fairest. But what is the argument against gas order (economically rational)? Do you expect in a strict gas-order world, that proposers wrap their gas fees around a potential MEVable transaction?

Yes I see your points, because as a user, even if you do not get a worse price by being sandwiched/front ran, your tx is still delayed because some MEV txs got included in front of you. I am very much reasoning in terms of toxic/non-toxic in a world where I accept that MEV is happening and I try to see which form is worse than the other.

Work In Progress but taking some time. I have actually submitted a FRP, I think it is worth the time investment to create a framework where we formalize everything and on which we can build further research.
Who has never dreamt to see the “toxic” vs “non-toxic” MEV split on the flashbots explorer? (cc @bert)

1 Like

I just realised that there may be a partially overlapping effort: mev-research/FRP-1.md at main · flashbots/mev-research · GitHub

@alex @fiiiu - has the new taxonomy been published?

No, as mentionned in the FRP, we want to build on this paper and further develop the concepts for specific set of transactions (i.e. bundles), to have a more precise taxonomy + corresponding formulas for sandwiches/backruns/frontruns…

Good point, I didn’t notice that “Quantifying Realized Extractable Value” was the output of that FRP. But I thought that there was a more detailed MEV taxonomy about to be published by Flashbots - isn’t that right @alex ?

Using a different EOA to the backrun. Potentially even inserting some other irrelevant transactions in between the frontrun and the target.