This paper is a companion to Open vs. Sealed: Auction Format Choice for MEV and studies a distinct but related problem: even a format-optimal auction can be undermined by builder defection ex post. We model a builder who, after observing submitted bids and payloads, defects from the honest auction outcome with probability \varepsilon and replicates a type-specific fraction \gamma(\tau) of the winning opportunity. Searchers respond by choosing between a risky first-price bid and a safe deterrence bid, yielding a piecewise equilibrium. Using the same libmev dataset, we estimate \gamma(\tau) from right-tail bribe plateaus and decompose observed auction revenue against the surplus a defecting builder could capture.
TLDR
-
Builder commitment is summarized by \varepsilon \in [0,1]: the share of opportunities on which the builder defects by exploiting observed bid and payload information. This channel is active regardless of auction format.
-
Searchers respond by choosing between a risky first-price bid and a safe deterrence bid b = \gamma(\tau)v, yielding a piecewise equilibrium with a type-specific cutoff v^*(\varepsilon, \tau).
-
Estimated foregone frontrun surplus is $49.4M, which equals to 48.8% of observed tips, with naked arbitrage and liquidations most exposed.
-
Sandwiches, already near the low-extractability regime identified in the companion paper, show almost no additional surplus exposed to defection.
-
Credible MEV auctions require constraints on the builder’s ability to use observed bid and payload information ex post, not just a better auction format.
1. The commitment problem
The companion paper shows that English and SPSB auction formats strictly dominate sealed-bid formats under affiliated MEV valuations, with a linkage gap of 14–28% at moderate affiliation. That analysis takes builder honesty as given. This paper asks what happens when it is not.
A builder who has observed all submitted bundles and their transaction payloads can replicate the winning searcher’s strategy, replace their transaction, and capture the underlying opportunity directly. Standard revert protection (MEV-Share, Flashbots) means the searcher’s transaction does not execute and no bid payment is collected, so the builder’s gain from defecting is \gamma(\tau) v_{(1)} rather than \gamma(\tau) v_{(1)} - b_{(1)}. The builder defects whenever \gamma(\tau) v_{(1)} > b_{(1)}.
The replicability fraction \gamma(\tau) varies sharply by MEV type. Sandwich attacks are mechanically reproducible given the victim transaction (with \gamma \approx 1). Complex liquidations require proprietary infrastructure (with \gamma \ll 1). This type-specificity is the key structural difference from the auctioneer-corruption literature, which treats the manipulation gain as format-agnostic.
2. Data
The dataset is identical to the companion paper: 2.2 million libmev transactions on Ethereum from September 2024 to August 2025, totalling $168.5M in extracted value and $101.3M in tips. Variable definitions, the log-normal fit (\hat{\mu} = 1.102, \hat{\sigma} = 2.524), and summary statistics by MEV type are reported there and not reproduced here.
The new variable this paper introduces is the potential defection gain: the positive part of \hat{\gamma}(\tau)v_i - b_i for each transaction, which measures the incremental revenue available to the builder from frontrunning rather than honoring the outcome. Five major builders, beaverbuild, Titan, BuilderNet (Beaver), bobTheBuilder, and BuilderNet (Flashbots), account for 93% of transactions.
3. Model
The environment and signal structure follow the companion paper directly: n risk-neutral searchers, log-normal affiliated valuations v_i = \exp(\mu + \sigma z_i) with z_i = \sqrt{\rho}Z + \sqrt{1-\rho}\,u_i, and MEV type captured through (n(\tau), \rho(\tau)).
The new element is the builder’s ex-post decision. After observing all bids (b_1, \dots, b_n) and the winning payload, the builder defects with probability \varepsilon \in [0,1]. Upon defection, the builder captures \gamma(\tau) \in [0,1] of the winner’s opportunity; revert protection ensures the searcher pays nothing. The timing is:
-
The builder’s defection rate \varepsilon becomes known.
-
Searchers decide whether to enter; n entrants are realized.
-
Each entrant i submits a sealed bid b_i.
-
With probability \varepsilon, the builder frontruns the winner; with probability 1-\varepsilon, it honors the outcome.
Searcher i's payoff is:
In the risky regime (b_i < \gamma v_i), defection zeroes out the surplus with probability \varepsilon. In the safe regime (b_i \ge \gamma v_i), the bid exceeds the frontrun value and the builder weakly prefers collecting the bid.
4. Equilibrium
Risky regime
In the risky regime, (1-\varepsilon) scales the payoff uniformly and drops out of the first-order condition. The equilibrium bid function is therefore independent of \varepsilon and satisfies the same affiliated first-price ODE as in the companion paper:
Piecewise Nash equilibrium
Comparing the conditional payoffs in each regime, \hat{\pi}(v_i) = (1-\varepsilon)(v_i - \beta(v_i)) in the risky regime, \tilde{\pi}(v_i) = (1-\gamma)v_i in the safe regime, defines the indifference threshold:
Since proportional shading falls with competition, \bar{\varepsilon}(v_i) is strictly decreasing in v_i. For any fixed \varepsilon, there is a unique cutoff v^* defined by \bar{\varepsilon}(v^*) = \varepsilon. The symmetric Bayesian Nash equilibrium is:
Optimal builder defection rate
The builder’s expected revenue is:
Taking the derivative with respect to \varepsilon (via Leibniz’s rule) and using the indifference condition to simplify the boundary term:
Since \frac{dv^*}{d\varepsilon} < 0, the boundary term is always non-negative. The optimal defection rate is therefore always a corner:
-
High-extractability (\gamma(\tau)v > \beta(v) for all v): integrand strictly positive \Rightarrow \varepsilon^* = 1.
-
Low-extractability (\gamma(\tau)v < \beta(v) for all v): frontrunning threat never binds \Rightarrow \varepsilon^* = 0.
There is no interior optimum. Whether a builder ever defects depends entirely on whether \gamma(\tau) is large enough relative to the competitive bid \beta(v) across the value support.
5. Empirical analysis
The empirical analysis proceeds in three steps: estimate the type-specific recoverable share \hat{\gamma}(\tau) from the right tail of the bribe schedule, decompose observed revenue against the surplus a defecting builder could capture, and check the maintained assumptions.
Bribe schedule and \hat{\gamma}(\tau). Under the piecewise equilibrium, low-value searchers bid the competitive schedule \beta(v_i), while above the cutoff v^*(\varepsilon,\tau) they switch to the deterrence bid \gamma(\tau)v_i. The empirical signature is a rising bribe share that flattens into a right-tail plateau, and the plateau identifies \hat{\gamma}(\tau).
Sandwiches are nearly flat at 90-95%: competition already absorbs almost the full recoverable value, so \gamma(\tau)v > \beta(v) rarely binds and there is no structural break to detect. Naked arbitrage shows the clearest transition, rising from approximately 45% at low values to a plateau at \hat{\gamma} = 0.74. Backruns plateau near \hat{\gamma} = 0.70, and liquidations, which are noisier because the sample is small, rise to \hat{\gamma} = 0.88. The rising-then-flat shape is the direct empirical content of the piecewise equilibrium: above the cutoff, high-value searchers bid up to the deterrence level \gamma(\tau)v to eliminate the frontrunning threat.
Revenue decomposition. For each transaction the potential defection gain is the positive part of \hat{\gamma}(\tau)v_i - b_i. Aggregated by type, this is the surplus a defecting builder could capture by replacing the winner rather than honoring the outcome.
The total foregone frontrun surplus is $49.4M, or 48.8% of observed tips. Naked arbitrage contributes the most, with $24.3M foregone against $18.3M of tips, because \hat{\gamma} = 0.74 exceeds the 67% mean bribe share over a large part of the distribution. Liquidations show the same mechanism on a smaller, noisier sample, with $12.4M foregone on $7.0M of tips. Sandwiches sit at the opposite extreme, with $7.3M foregone on $57.8M of tips, because competitive bids already extract close to full value. Backruns are intermediate, with $5.5M on $18.3M.
Affiliation and concentration. Two maintained assumptions hold in the data. Within each block and type, the largest and second-largest log extracted values are positively related, supporting the affiliated-values assumption; and extracted MEV is highly concentrated at the searcher level, with Gini Index equals to 0.93-0.97 across types, which supports the right-tail interpretation of the revenue decomposition, which is a small number of high-value opportunities drive most of the foregone surplus.
Honest-disclosure benchmark. As a reference point, Bergemann et al. (2022) characterize the revenue-maximizing disclosure policy of an honest seller, giving a threshold \varepsilon^{\text{Berg}}(n) that depends on the effective number of bidders: thicker markets tolerate more information release, thin markets require more pooling of the upper tail.
The empirical implied \varepsilon by type is type-specific through n(\tau), so the honest benchmark itself is type-specific rather than a single universal prescription. \varepsilon^{\text{Berg}}(n) should be read as an upper bound on user-beneficial disclosure, not as the builder’s optimal defection rate. Once the builder can act on observed information, the admissible exposure must sit below the honest-disclosure benchmark wherever the defection gain is positive.
6. Cross-builder evidence
The five major builders account for 93% of transactions. Disaggregating the bribe schedule by builder shows that the type-specific pattern holds within each builder, but with meaningful cross-builder dispersion.
Sandwich bribes cluster at 95-100% regardless of builder, consistent with \hat{\gamma} \approx 1 and saturated competition. Naked arbitrage and liquidation show much wider cross-builder spread, consistent with the type-specific \gamma(\tau) structure and with differing searcher pools across builders. Titan’s bribe standard deviation is roughly 3 times beaverbuild’s, consistent with a more heterogeneous searcher base. The same heterogeneity appears on the demand side: builders attracting more unique searchers show higher mean bribe shares and lower bribe volatility in commoditized types, while specialized types show fewer searchers and more dispersion. This cross-builder heterogeneity is what the next section exploits: the same nominal defection gain implies very different incentive constraints depending on a builder’s revenue base and order flow.
7. Builders’ defection and incentive compatibility
The result from Section 4, that \varepsilon^* depends on whether \gamma(\tau)v exceeds or falls below the competitive bid \beta(v), leaves open the question of whether commitment can be sustained at all when the auction is repeated. We close that gap by embedding the static auction in a repeated game between the builder and the searcher pool.
The IC condition
Defection in period t yields the one-shot surplus \Delta_t = \sum_{i \in \text{round } t} \max\{\gamma(\tau_i)v_i - b_i,\, 0\}, where the \max captures selective defection (the builder frontruns only bundles for which \gamma v_j > b_j). With probability p the defection is detected and the searcher pool migrates away, so the builder loses its continuation revenue from t+1 onward; with probability 1-p play continues. Comparing the two paths at per-period discount factor \delta:
Subtracting gives \text{NPV}_d - \text{NPV}_h = \Delta - p\frac{\delta}{1-\delta}\pi_h, so honesty is incentive-compatible iff
Honesty is sustainable only when the builder is patient enough that the discounted future business stream exceeds the one-shot temptation. As \delta \to 0 (the myopic builder), the threshold m/(m+p) \to 0 and any positive \Delta produces defection, recovering the high-extractability branch of Section 4; interior \delta produces interior \varepsilon^*.
Per-builder thresholds
For each builder we compute the panel-total defection gain \Delta_b, the mean monthly tips \pi_h over active months, the months-equivalent m = \Delta_b/\pi_h, and the annual IC threshold \delta^*_{yr} = (m/(m+p))^{12} stands for the minimal annual discount factor for honesty.
| Builder | \pi_h ($M/mo) | \Delta_b ($M) | m | \delta^*_{yr} (p=1) | \delta^*_{yr} (p=0.1) | \delta^*_{yr} (p=0.01) |
|---|---|---|---|---|---|---|
| beaverbuild | 3.74 | 10.10 | 2.70 | 0.02 | 0.65 | 0.96 |
| Titan | 2.25 | 9.01 | 4.00 | 0.07 | 0.74 | 0.97 |
| bobTheBuilder | 0.47 | 4.14 | 8.88 | 0.28 | 0.87 | 0.99 |
| rsync-builder | 0.36 | 0.61 | 1.69 | 0.00 | 0.50 | 0.93 |
| BuilderNet | 0.42 | 23.00 | 54.71 | 0.80 | 0.98 | 1.00 |
| BuildAI | 0.15 | 2.06 | 14.00 | 0.44 | 0.92 | 0.99 |
| Ty For The Block | 0.49 | 0.07 | 0.15 | 0.00 | 0.00 | 0.47 |
At p = 1 (case of certain detection) only BuilderNet has a threshold above plausible operator discount factors. As p decreases the binding region expands: at p = 0.1 the threshold for beaverbuild and Titan is 0.65-0.74, and at p = 0.01 every major builder has \delta^*_{yr} > 0.95, meaning IC fails for any operator with even modest impatience. Since replication-based frontrunning has no clean on-chain fingerprint, a revert looks identical from the searcher side to a lost auction, plausible p is in the lower part of this range. Within that range, the IC constraint binds for every major non-TEE builder in the panel.
Type-level decomposition
Decomposing m by type with total monthly revenue in the denominator (detection on any one type costs all future business, not just that type’s) gives m^{\text{tot}}_\tau = \Delta_\tau / \pi_h:
| Builder | sandwich | naked arb | backrun | liquidation |
|---|---|---|---|---|
| beaverbuild | 0.27 | 1.22 | 0.57 | 0.65 |
| Titan | 0.78 | 0.53 | 1.19 | 1.50 |
| bobTheBuilder | 8.65 | 0.00 | 0.23 | — |
| BuilderNet | 0.35 | 38.42 | 0.76 | 15.17 |
| rsync-builder | 0.95 | 0.34 | 0.38 | 0.01 |
Cells with m^{\text{tot}}_\tau > 1 indicate that single-type defection alone exceeds one month of total honest revenue, so even at moderate p defection on that type alone breaks IC. The first type to defect on varies systematically across builders: naked arb for beaverbuild, liquidation and backrun for Titan, sandwich for bobTheBuilder, and naked arb + liquidation for BuilderNet, which together account for 98% of its counterfactual exposure.
BuilderNet as a natural experiment
BuilderNet, rbuilder running in TEEs, enforces \varepsilon \equiv 0 architecturally rather than reputationally: the operator literally cannot observe individual bundle payloads in a form usable for replication-based defection. This gives a clean cross-builder comparison. BuilderNet has the lowest value-weighted bribe share of any major builder, equal to 12.3%, which reflects the absence of the deterrence premium \gamma(\tau)v that searchers post at non-TEE builders, so bidding sits closer to the competitive level \beta(v); yet by far the largest counterfactual defection exposure (with m = 54.7), concentrated in naked arb and liquidation. The TEE architecture is doing precisely the work the IC constraint would otherwise require of reputation: searchers route the most replication-vulnerable flow to the one operator that cannot defect, precisely because the same surplus would be exposed elsewhere.
Two architectural responses close the IC gap, both by driving \varepsilon \to 0 structurally rather than reputationally: TEE-shielded execution (payloads not observable in usable form, so defection is cryptographically precluded), and sub-slot composable settlement (bundles settled by an upstream auctioneer and delivered to the builder as already-committed obligations, so the builder never holds the payload ex post and p becomes irrelevant; that is, composability adds a second efficiency channel).
8. Design implications
The companion paper’s recommendation, switch from FPSB to English or SPSB and capture 14-28% more revenue, is valid but incomplete. A builder who can observe bids and payloads before finalizing the block can undermine any format’s honest outcome. The commitment channel is quantitatively comparable to the format-choice channel and operates independently.
The relevant safeguards are not format changes but constraints on the builder’s information advantage at the moment of allocation: payload encryption until after block finalization, cryptographic commitment schemes, or slashing conditions that make defection observable and punishable. Delayed or aggregate information about competition can preserve some of the linkage benefit identified in the companion paper while reducing the ex-post manipulation surface.
The appropriate design therefore has two layers: first, choose a format that exploits affiliation (English or SPSB for thin, moderately correlated markets); second, constrain the payload and identity information available to the builder so that the chosen format’s honest outcome is actually implemented.