The Free Option Problem in ePBS

tl;dr: we argue that the “free option” problem is potentially substantial in the current specification of EIP-7732: Enshrined Proposer-Builder Separation

joint work with @0xSybil and @BrunoMr
Special thanks to @dataalways for extensive discussions and to Fei Wu and @sui414 for shared data. Here is Part II.

Imagine you are at the roulette table in a casino and somebody makes you an offer: you get the option to revert the next spin of the roulette wheel if you don’t like the outcome. Reverting here means that all placed bets are ignored and every gambler has the same amount of chips as before the spin of the wheel. How much would you be willing to pay for that option? How much would you bet? And how likely is it that you will exercise the option?

You should value it a lot. You get an expected return of almost 50% with zero downside risk. You would like to bet as much as the bank allows you to bet if you had the option. And you would exercise the option every time you lose, i.e. a bit more than half of the cases.

The “free option” problem is the following: a builder that commits to an execution payload and wins the block auction, can wait to invalidate the execution payload until shortly before the payload timeliness committee (PTC) deadline (for the details see below). In case new information arrives after the commitment but before the PTC deadline and that information makes it preferable to publish an empty block rather than publishing the block committed to, the builder can choose to do so.

We argue that the current specifications of EIP-7732 (8s to exercise the option) could make it a valuable option to builders and a large fraction of blocks could end up empty, with more empty blocks in times of high volatility and around important events. We moreover, argue that the problem becomes much more severe in a potential future where more and more liquidity and financial activity moves to Ethereum.

Casinos, Binance and Ethereum

We can quite easily modify the above casino story to see the intuition:

Imagine you are trading BTC or ETH (spot and futures) in a highly liquid market such as Binance and someone offers you the option to revert all of your trades after 8 seconds if you don’t like how the market has moved within these 8 second. How much would you be willing to pay for that option? And how likely is it that you will exercise it?

Again, you get a positive expected return with zero downside risk. Liquidity determines the size of your position and how much value you can extract: at current market conditions (as of writing this) you can trade ~10 BTC on the smallest tick (1 cent for the spot resp. 10 cents for futures which means less than a dollar of slippage cost). Let’s ignore Binance fees for the moment (not because they aren’t substantial, but because we want to figure out what would happen if the same amount of liquidity would be traded on chain where hopefully there are no rent seeking middle men). So the casino allows you to bet around a million $ per spin. The average value of a spin now is ~1-3 bps (see below) instead of close to 50%. Thus, we are at 100-300$ option value on average per spin. On the other hand, Binance is a casino with several (correlated) roulette wheels: major token pairs have significant liquidity (although less than BTC) which should give us a factor 3-5 in comparison to only trading BTC. 300-1500$ for 8s, not bad. If you only trade BTC, you should exercise the option about half of the time (over 8 seconds it’s almost a coin toss whether BTC goes up or down). If you trade other things as well, then you get less exercises of the option (diversification reduces risk), but still significant chances of exercise.

Ethereum block building is of course different:

Simplifying, but not oversimplifying too much, if you are an (integrated) builder, your value for the block 8 second after having committed to it is a positive constant term (whatever block value you have at commitment from “non-arbitrage” MEV) + a random term (you directional risk if you take on a position e.g. in ETH by trading on chain). Exercising the option in case of a bad realization of the random term will also make you lose the positive constant value. But as in the casino example and the BTC example you can amplify the random component as much as (on-chain) liquidity allows you to, thereby increasing the overall return. Thus, you might exercise the option less than half of the time, but more often than we like, provided there is enough liquidity on chain.

Put differently, having the option to invalidate the block after eight seconds changes the kind of blocks that builders will build and the option will exercised much more often than naive extrapolation from current data would suggest.

Some napkin math & preliminary conclusions

Suppose ETH returns (log prices) are normal (over short short time intervals they really really aren’t, but it’s an approximation good enough for napkin math standards and ceteris paribus everything would be much worse with fat tailed returns). A napkin math approximation of the option value with 0 mean is

V\approx E[|\text{returns}|\cdot L/2]\approx\tfrac{1}{\sqrt{2\pi}}\sigma_\tau \cdot L\approx 0.4\cdot\sigma_\tau\cdot L

where L is notional of my position and \sigma_\tau is the volatility over \tau seconds. In other words, and most relevant for ePBS design considerations:

The value of the option is increasing in volatility and increasing in liquidity. As volatility (and also liquidity to some degree) is increasing in time, its value is increasing in the time to expiration.

We can get a more sophisticated version of napkin math if we look at the positive constant term as well. Let’s call it \mu for MEV. If we factor it in we get the absolute value of a normal with positive mean as the option value. Thus, the value of the option per unit of MEV is

\frac{V}{\mu}\approx0.4\frac{\sigma_\tau L}{\mu}\exp\{-\tfrac{\mu^2}{2\sigma^2_{\tau}L^2}\}-(1-2\Phi(-\tfrac{\mu}{\sigma_{\tau} L}))

and we get a corresponding probability of option exercise of \Phi(-\frac{\mu}{\sigma_\tau L}) where \Phi is the cumulative distribution of a standard normal. Thus, whether or not we have a problem depends on the ratio \rho:=\tfrac{\mu}{\sigma_{\tau}L}.

In other words, the napkin math tells us that everything else being equal, the free option problem gets worse if we reduce (non-arbitrage) MEV per unit of liquidity on-chain. As reducing on-chain MEV is a goal that many solutions and applications work towards (e.g. through better AMM design, middle-layer infrastructure, encrypted mempools and so on and so forth), paradoxically, the free option problem gets exacerbated the better we make the on-chain trading experience.

Next let’s go to the ultimate piece of napkin math: suppose you are an arbitrageur who wants to make use of the CEX-DEX price discrepancy at the end of the slot. One (suboptimal) way of doing this as an integrated searcher-builder would be, at the beginning of the slot, to place a top-of-block DEX swap in a block and cancel the block bid close to the end of the slot if the price moves against you on the CEX in the meantime. If it moves in your favor you would hedge on the CEX. This gives you half of the value of an optimally executed arbitrage trade at the end of the slot, as you are only trading in one direction and prices go up and down with equal likelihood. In other words:

:robot: The free option should get you about half the pure CEX-DEX arbitrage gains you could make in a world with \tau seconds length slots. If CEX-DEX arbitrage accounts for x% of block value, then \rho=\frac{\mu_{12sec}}{0.5/0.4\tfrac{x\%}{1-x\%}\mu_\tau}=\frac{4(1-x\%)}{5x\%}\frac{\mu_{12sec}}{\mu_\tau} where \mu_{\tau} is the MEV we could capture per block if slot times were \tau seconds.

How arbitrage gains scale with slot time is a contested question. But the factor should be between \sqrt{\tfrac{12s}{\tau}}\leq \rho<\tfrac{12s}{\tau}, i.e. for \tau=8s, we have 1.2247<\rho<1.5 . This is because volatility scales with the sqrt of time in theory (assuming a Gaussian process) and practically, in short time intervals, between the sqrt and linearly in time. Thus, the probability of executing the option in case of 55.5% arbitrage MEV is between \Phi(-1.5)=6.68\% and \Phi(-1.2247)=11\%. The following graph, illustrates hypothetical exercise probabilities with sqrt resp. linear scaling of MEV in time. While the estimates might be quantitatively off, qualitatively we should expect this behavior. The longer the free option window relative to the slot time, the worse the problem.

There are a bunch of immediate questions that are best answered empirically (as we partially do below):

  • How does volatility and therefore value scale with \tau?
  • When is \tau too long?
  • What is the expected cost of exercising the option \mu as a function of liquidity and slot time?

Time to expiration in ePBS

In current specs of ePBS [1,2,3], the validator-builder handshake starts with a commitment at t=0s, the builder signs an ExecutionPayloadHeader and the bid is directly removed from the builder’s beacon chain balance and credit to the beacon block proposer, but the delivery of the blocks and blobs now passes two checkpoints enforced by the 512-member Payload-Timeliness Committe (PTC). At t=4s there is the payload deadline, if not enough PTC votes attest that the full execution payload has been propagated, the slot is flagged empty and there wont be any state transition. At t=10s, a second PTC pass requires the blobs are available. In case sufficient PTC votes, the slot is re-labelled as empty, and the Payload is in-validated. Since the builder has already paid in the moment of taking the decision to release or not release the blobs, the decision remains in the builder’s hands, and so, the builder enjoys a time-bounded free-option of approximatly 8s (assuming that 2s are needed to safely propagate blocks among PTC members).

Figure 1. Dual-deadline PTC vote in ePBS (Adapted from Dual-deadline PTC vote). The diagram shows that the payload must be available early in the slot (≈ 4 s) to give clients time to execute it, whereas blobs enjoy a longer propagation window until ≈ 10 s. Red vertical markers highlight the last moment when a strategic builder can safely release the payload, and later the blobs, to almost guarantee inclusion.

Some empirics

Let’s look at the actual performance. Here is a histogram of returns for one million notational traded every 8sec in ETH-USDC based on the last trading day (21 July 2025):

We choose, one million notational because that is about the size of liquidity we can access on Binance for ETH-USDC/T pairs (with close to no slippage cost). We get a realized average value of the option of ~117$ (~1.26 Million $ per day). Whether you choose to always go long or always go short (or randomize) does not really matter for the realized average value and also does not for the realized frequency of exercising the option (around 48.6%). Of course a sophisticated trader might be able to extract more value, e.g. if she has a signal that allows her to choose with better than 50% the direction of the price over the next 8 seconds. The point is that even an unsophisticated trader makes a lot of money.

Let’s also look at how arbitrage gains scale with time. We can observe at least two things: first that the napkin math model that arbitrage gains are proportional to volatility and that volatility is proportional to the sqrt of time is not too much off. Second that the option value and therefore, arguably, the free option problem is substantial for modestly long time to expiration.

Caveats and Qualifications

We want to caveat that the “free option” might not be used by builders even if highly profitable, as there might be social slashing and reputation effects at play. If validators blocklist builders who have missed payload delivery, this can make exercising the option for short term profit not attractive. Similarly, if order flow providers, searchers and other transaction originators, avoid builders who have failed payload delivery repeatedly in the past, builders might refrain from using the option. While we hope that these “good” equilibria would play out in reality, it is important to note that there are also other equilibria where social slashing does not happen and reputation effects are not strong enough to enforce “good” builder behavior. Protocol design should arguably not rely on good equilibria being played, but should provide robust incentives in a variety of scenarios. Hence, we think that the analysis of whether the option is valuable to exercise absent of repeated games effects is important.

Second, the free option problem is a problem if we want to use L1 as execution layer in the future. If the L1 is primarily used for data availability and settlement and we primarily use it to post blobs or settlement transaction, the free option problem becomes less important.

Conclusion

Whether the option value of not releasing a block matters depends on several factors: volatility, liquidity, spreads, on-chain MEV, and the expiration time of the option. Yet as long as Ethereum supports real economic activity, the ePBS free option will keep a non‑negligible value with the current parametrization, and builders have an incentive to exercise it while entry to the ePBS market stays permissionless. Each time they do, users suffer: their trades remain pending and trading venues widen spreads. The number of empty blocks can vary, but the moments when, arguably, users most need to trade, namely periods of high volatility, are exactly when empty blocks spike. Honest builders who always release payloads and blobs on time then can become less competitive. That dynamic can trigger a vicious spiral in which validators stop block‑listing the withholding builders to earn higher fees, until almost every builder adopts the same strategy.

We see three groups of economic mitigations in the current ePBS design: block‑listing, shortening the option window, and making the option effectively more costly. Block‑listing builders that fail to release the payload or the blobs on time can work for a while, but it is fragile. Any validator can disable the list with a simple configuration change and is tempted to do so, because allowing the full set of builders to bid, including those that exercise the option, boosts competition and therefore expected auction revenue; once one staking pool unblocks them, the rest will potentially follow.

Shortening the release window for blobs and payloads, reduces the option’s value quickly. Cutting the free option window from eight seconds to six, or even better three seconds would sharply lower the probability that builders exercise the option. One must, however, balance this against the fact that windows that are too short can increase the risk of re‑orgs.

Finally, exercising the “free option” can be made more costly. In today’s ePBS spec, releasing or not releasing costs the builder exactly the same: the bid paid in the ePBS auction. Not releasing is therefore a sunk cost. Introducing a penalty for failing to release either the payload or the blob (possibly with different penalties, because blobs give the builder more late‑MEV value than payloads) can strongly disincentivize the strategy. For any fixed cost there will exist rare events where deviation is still profitable; the goal is not to eliminate the strategy in every edge case but to make it unprofitable in almost all practical situations, thereby leaving only negligible negative externalities while preserving ePBS’s scalability benefits. The penalty could be static, for example one ETH per unreleased block, or dynamic, similar to EIP‑1559 but keyed to the share of unreleased payloads, blobs and previous MEV bids. The fee must be large enough to make the option essentially worthless for 99.9 percent of the time, yet small enough to keep entry barriers low and MEV bids competitive.

In short, today’s ePBS design may solve scaling, but it can also distort the block‑building market and hurt users. Tightening PTC deadlines and/or slashing builders on non‑release, look like the most promising fixes if we want to implement ePBS.

16 Likes

This is a very nice analysis. I want to point out two things that were said at the call.

There are three objects that are being broadcast during a slot, the block, the payload and the payload data (blobs). Today the deadline to see those three is the same one. When we scale to maximally utilize the slot each of these objects requires a different deadline. The data processing is only about broadcasting it, validation is free, thus, we want to set the deadline for its availability as late as the end of the slot. The payload does not have this luxury: it needs to be executed today and it needs to be ZK proven in the future. Thus, the optimal deadline for the payload is some time before the slot. On any system that has these different deadlines, the builder will have an option to withhold some private data. This includes, APS, Mev Boost, DE and slot auctions in ePBS (as opposed to block auctions) any system that has different deadlines for the payload than the data suffers from this option. So this is an inherent tradeoff to scaling.

Another observation is that there are two issues with this option: one is the value to the builder. Another one is the effect of missing blocks on L1. I’d argue that the value to the builder is actually higher than the naive approach that you only exercise it when the trade has gone against your bet. The issue is that the builder can run a secondary auction after the commitment and take bids for both winning and losing parties. This should make at the same time the option itself more valuable and also produce less missing blocks on L1.

An interesting point is that the transactions are fully visible during this option, it’s not like the builder is the only one with this information. Analyzing how the market can react with this knowledge, that is, that the post state is either A) The payload is executed or B) the payload is not executed as opposed to any arbitrary post-state, seems like an interesting problem in its own.

Perhaps the most important argument attenuating this problem is the above mentioned one on private order flows and searchers, given that these agents are the ones that would lose the most on missing payloads.

Finally one last comment is that builders that are willing to play this game will be detectable before they actually manage to withhold a payload. The reason being that the only way to prepare for such a withholding is by first introducing a private blob that has not been seen by anyone on the chain. Any large node on the chain should be able to immediately recognize this behavior. Not that it makes any difference in this analysis of course.

6 Likes

On the first point: after this has been explained to me like five times in the last days by various people including you in the call yesterday, it has clicked with me. :slight_smile: Yes, this, seems to be a fundamental problem with any design that has full “pipelining”, modulo some other clever solutions or re-thinking what we do with blobs in general. So I don’t have the answers, but I acknowledge the problem.

Regarding the second point: yes a builder could monetize the option differently, which is weakly increasing its value. I think in reality though, and maybe I’m wrong on this, the free option is attractive not so much to established builders with private orderflow, who tend to have skin in the game, but for a searcher who plays the roulette on/off if they received a good signal and there is high EV for the next spin. So they spin up a builder just for these occasions, don’t even need orderflow and are relatively sure that their signal is the best.

On the point that your plan to use the option can be revealed by introducing a private blob, that certainly helps if we want to socially slash or withhold order flow from you. But it doesn’t change the incentives fundamentally, in particular for the case of the searcher using the option to only place their high EV bundles.

Agree, on the suggestions and the fact that more research can and will and should be done.

2 Likes

Nice post that solves my confusion :slight_smile:

Before we answer the question “what’s the probability of withholding if I wait”, should we first ask the question “should I wait and keep the option, or should I just kill this option”?

So for an integrated searcher-builder, they already have their DEX position upon winning the block. Assuming that the arbitrage opportunity is profitable at the time of commitment, they have the choice to immediately hedge on CEX to neutralize the position, and then reveal the valid payload before the deadline to realize the profit.

Should builders even “gamble” on the “wait and optionality” policy?

The builders prob need to decide which policy to use at the time of bidding. How would this change their bid?

Would love to know your insights.

Usually, the best policy is to wait than actually hedging immediately. Suppose searcher has a CEX-DEX arbitrage opportunity. At time t=0 she can buy a quantity q of Eth at price P_0 on the DEX and sell it at price K_0 on the CEX. If she hedges instantly, the payoff is q(K_0-P_0). If, instead, she waits until the option’s expiration time t=8s, the expected payoff is \mathbb E[q(P_0-K_t)_+]. When (K_t)_{t\geq0} behaves like a martingale (which is a reasonable assumption when t is sufficiently small), this expected payoff increases with time. So, the searcher will have incentives to wait for the CEX price movement at time t=8s rather than hedging immediately.

3 Likes

Not sure whether I got your question correctly, in addition to what @0xsybil said: Even if you arbitrage away the CEX-DEX difference and trade the opposite lag immediately, the option has still the same value. If you went short on the CEX and long on the DEX, you make money if you can cancel the long trade if the price goes down on the CEX.

Also you can take additional position and gamble on price going up. So you can and might want to trade a larger quantity than if you don’t have the option.

2 Likes

We observed empirically that waiting till the end is optimal for almost all trading pairs and time periods, in the context of TimeBoost, which covers at least the same CEX-DEX part of the value that you discuss in your analysis: [2410.10797] MEV Capture Through Time-Advantaged Arbitrage

2 Likes

Thanks a lot for the write up!

I just want to highlight several aspects based on empirical data collected for our paper on CEX-DEX arbs that make withholding a worse strategy than it seems.

  • Looking at data we find that the median builder’s bid is ≈ 6 bp of block notional, and the bid is paid to the proposer no matter what so it’s a sunk cost.
  • We also find that the median builder profit margin only represents ≈ 3 bp, so much smaller than the bid.
  • Immediately after winning integrated searcher-builders can hedge the DEX leg on a liquid CEX for less than 1 bp (taker fees + spread). This is a bit less clean for searchers that are not integrated with builders but a first intuition could be that builders would be incentivize to reveal to searchers as soon as possible so searchers can hedge, otherwise searchers would shade their bids and the builder would capture less value.
  • The residual 8-s noise for ETH is around σ₆s ≲ 0.8 bp historically.

So given the numbers above you could compute the reveal vs withhold payoffs:

Without hedging:

  • Reveal: +3 bp margin − ≤ 0.8 bp residual ≈ +2.2 bp
  • Withhold:−6 bp sunk bid

With hedging you just subtract 1 bp in both conditions but it doesn’t change the point.

So in most realistic price paths, revealing outperforms withholding by ~8 bp. Of course, you also lose all the other MEV in the block that’s not associated with this particular asset price action. And endogeneity doesn’t really help because the price movement becomes Δ which can be immediately hedged.

1 Like

The issue is that the builder can run a secondary auction after the commitment and take bids for both winning and losing parties. This should make at the same time the option itself more valuable and also produce less missing blocks on L1.

Wouldn’t this mean it’s back to 50-50% chance and you there is no point in playing this game for searchers? As in there would be 50% chance your bid on the secondary auction matches the outcome you look for.

This means the secondary auction systems wouldn’t drive the value of the block, so a pure integrated searcher/builder doing it would more easily win the blind auction (as it has 100% chance having the desired outcome).

Thanks for adding the data for context. I understand from your argument that on average the option wouldn’t be exercised. That is also true in our napkin math :smiley: But we would need to look at the quantiles to know how often it would be exercised. But without knowing that in detail, I would believe you that we currently don’t have the liquidity on chain to make this option as useful as it would be in other regimes. But also averages on the volatility don’t give you the full picture. By chance we picked a higher vola day for our very preliminary data analysis. But hopefully, we don’t only want to produce >99% non-empty blocks on average per year, but >99% on average per day.

Your point that if you consider exercising the option than you need to coordinate with the searcher or be the searcher I agree 100%. The fact that you can trade out of your position with <1bps trading cost seems, however, independent of whether you do it after 1 sec or 8 sec (or 3s or 15s or whatever). Holding a position for longer might have risk but doesn’t reduce EV (with the usual caveats). And if you have the free option, your value for it and the decision whether to exercise it for the DEX trade should be independent of any CEX trades you do.

1 Like

Also there is a common misconception I also tend to have once or twice per day: It is not ! true that the builder forgoes the arb profit of 3bp if he withholds the block. Suppose I long ETH on the DEX and short ETH on the CEX at t=0 and the price differential between the two is 9bp. Suppose the price goes down on the CEX by 9bp until t=8. I cancel my trade on DEX. Then I gained 9pb on CEX. I still made 9-6=3bp.

Even more so: suppose the only value in the block is CEX-DEX arbitrage. Then you pay the same 6bps independently of whether you exercise the option or not. So you lose 6bps in both scenarios on the bid. So the only cost of exercising the option is the non-arbitrage MEV in the block.

1 Like

Thanks for the interesting write-up! At Shutter, we’ve recently looked into the free option problem in ePBS as well and believe there’s a nice cryptographic solution for it.

The idea is for the builder to send an encryption of the block along with the commitment and (optionally) a zero-knowledge proof that the encrypted block matches the commitment and is valid. The ZK proof might not even be necessary if the builder can be slashed for submitting an invalid encryption. If the builder now refuses to release the block, the encrypted payload can simply be decrypted.

To enable this, we could use silent threshold encryption, which allows the builder to encrypt the block to a committee (potentially the payload timeliness committee). This committee can later decrypt the block in case the builder does not release the payload. The nice thing about silent threshold encryption is that the committee can be formed dynamically (i.e., it does not consist of a fixed set of parties) and committee members do not need to interact with each other. This approach definitely introduces some latency, but within a what we believe to be a reasonable range (max a few hundred ms if no ZK proof is used).

We described this approach in a recent blog post [1]. Paradigm actually proposed a similar idea in a post last year, though not specifically in the context of ePBS or the free option problem [2] (Can’t add links to the posts as this is a new account).

Would be great to hear your thoughts on this idea!

[1] Shutter’s Perspective on Glamsterdam - and a Cryptographic Fix for ePBS’s Free Option Problem
[2] How to Remove the Relay - Paradigm