Research & Reports

How the Execution-Quality Numbers Actually Work (There's No Single Score)

Ask ClearTrace for "the execution-quality score" and there isn't one, on purpose. This is the formula-level companion to the interactive explainer: the quote-gap decomposition (fee, MEV, slippage, gas) spelled out symbol by symbol, and the rated-versus-preliminary confidence gate (≥30 realized samples over a ≥7-day span) that decides when a number is stable enough to cite.

Prefer to explore the data first? The interactive execution-quality explainer is the companion chart to this piece.

The $540.85M of Base DEX Volume Nobody Can Attribute

In the 7 days to September 21, 2026, 9.09% of tracked Base DEX volume routed through contracts that no public label, no registry, and none of our forensic passes can tie to a named origin. A further 18.2% resolves to an operator or a role rather than a frontend, and on the strict reading the figure is 27.31% — unchanged from the previous label date. That pairing is the story: 13 contracts holding $52.76M left the unattributed pool when the 2026-09-23 sync re-resolved names, all 13 to one operator's automated trading fleet, so the headline fell by $52.76M while the strict figure did not move at all. Nothing was placed; volume moved from "no name at all" to "an operator we can point to", which is a real result and still not a product anyone could name. The head of the pool did not trade differently. The three largest moved $199M between them, and the two largest still open ($92.1M and $82.0M) are characterised but not identified. We publish what we know about each one and the dead ends. This is an open request for labels, with credit for verified contributions.

What it covers:

• The seven contracts we took off the open list ($597M in a week), the two largest still open at that point ($92.1M and $82.0M), and every attribution vector the named ones survived.
• Why the unattributable slice matters: growth-number integrity for incentive programs, and unknown winners no ecosystem team has on its radar.
• How to contribute a label, and the verification bar it has to clear before it ships.

Three of the Four Tokenized Stocks We Measured Were Fakes

Coinbase issued tokenized Apple, Nvidia, Meta and Alphabet shares on Base in August 2026. Our first measurement of how they trade was built on Base's own documented address convention, and for three of the four tickers it returned an impersonator instead of Coinbase's token. This note is about why, because the answer is not carelessness: the ticker is free to claim, one impostor's token name is byte-identical to the real one, the address prefix is shared by 7,532 tokens, and every contract in that namespace returns the same single byte of code. We report our own error and what survives it.

What it covers:

• Each signal someone might use to tell a real tokenized share from a fake, tested against Coinbase's NVDAc and the impostor that outranked it, and why six of the seven fail.
• Why "18 decimals and a round billion supply" is not the tell it looks like: that is the ordinary default in the namespace, and it is Coinbase's tokens that are unusual.
• The fleet's shape: 13 contracts, ten of them alive between 28.9 and 30.0 minutes, 635 taker-slots at $401 a head, starting on launch day.
• The one check that works, why it is an unsatisfying answer, and what we changed in our own pipeline.

One Token Is Not One Share

Tokenized stocks on Solana do not keep their share count in the balance. Both large issuers hold it in a separate on-chain multiplier, which is how dividends and stock splits are applied without minting or burning anything. Trade data reports the raw token amount, which is correct and is not a share count. Divide a trade's dollar value by it and you get a price per token: for 424 of 1,101 xStocks that is not the share price, and for Netflix it is ten times it.

What it covers:

• The census across every mint both issuers publish: 424 of 1,101 xStocks and 242 of 451 Ondo Global Markets tokens carry a multiplier other than 1, ranging from 0.0667 to 10.0261.
• Netflix as the visible case, checked against an independent perp oracle, and why the small multipliers are the dangerous ones.
• The second trap: the field you read is not the value in force, which returns a superseded number for 416 of 1,101 xStocks including the Netflix split.
• Why this is not a defect in anyone's data, and the three steps that get a share price right.

A Neutral Verifier for DeFi-Execution Agents

If you are building an agent that trades on-chain, how do you score whether it executed well? The quote is the obvious signal and the wrong one. A quote is a promise, and the realized fill is a different number. ClearTrace already runs a verifier for that gap: it fork-simulates a proposed route against live chain state and decomposes what the trade actually lost to fees, MEV, and slippage, scored the same way for every venue against a neutral baseline. This note describes the mechanism, shows a worked example on real data, and asks what a verifier like this would need to be to pull into an agent loop.

What it covers:

• The step function (propose calldata → realized fill) and the rubric that decomposes the quote-vs-fill gap into fee, MEV, and slippage in basis points.
• The rated-cell trust gate (≥30 realized samples, ≥7-day span) that keeps a burst of correlated samples from masquerading as an eval signal.
• Three ways it plugs into agent work: offline benchmark, online RL reward, and neutral referee, plus the limits (Ethereum-only, ~27 days, RFQ venues unrealized).

The Quote You Are Shown vs. the Price You Get

Every aggregator shows you a quote before you sign. ClearTrace's quote sampler captures that quote and fork-simulates the fill, so the gap between promise and delivery is measurable rather than assumed. Only cells rated at 30+ realized samples over a 7+ day span are claimed. This window, that gap is near zero for almost every venue, with one clear exception.

Key findings (Ethereum, rated quote-sampler cells):

• Seven of eight rated cells quote within about 2 bps of their realized fill: 1inch, Bebop, SushiSwap and the direct-Uniswap baseline at 0.0 bps, Odos at 0.21, ParaSwap at 1.0, KyberSwap at 2.02. Median across the rated aggregators: 0.21 bps. For the sampled book, the quote most aggregators display is a reliable predictor of the fill.
• OpenOcean is the outlier: its median realized fill ran 11.3 bps below the price it quoted, drawn from 1,154 realized samples (99% of its sampled quotes realized, so this is not a thin-sample artifact). That is roughly 5.6× the next-widest rated gap and the only double-digit gap in the set.
• Quote accuracy is a separate axis from execution cost. SushiSwap quotes almost perfectly (0.0 bps gap) yet posts the widest realized on-chain slippage in the rated set at 25.86 bps: an honest quote for an expensive fill. 1inch is tight on both (0.0 bps gap, 5.49 bps slippage, inside the 10.75 bps direct-Uniswap baseline). OpenOcean is wide on both.
• Honest bound: these are fork-simulated forward captures on a fixed set of pairs and sizes, not a census of every trade. The claim is that on the sampled book, OpenOcean's realized fills sat measurably below its quotes while its peers' did not.

Want to explore the same sampler data yourself? The interactive execution-quality explainer lets you isolate any aggregator across four trade sizes.

What Execution Quality Costs You on Base (Interactive)

We requested the identical trade from nine aggregators on Base and replayed each against a Base fork. The headline is good news: most venues fill within a basis point of their quote and failures are rare. The exceptions are the finding: one router's failure rate quadruples as the order grows, one venue's costs are dominated by its own disclosed fee and it often declines to quote at all, and exactly one aggregator routes any flow through off-chain RFQ makers. Explore reliability and realized cost across three trade sizes.

What Execution Quality Actually Costs You (Interactive)

The quote a DEX aggregator shows you is not what you get. We requested the identical trade from seven aggregators and replayed each against a mainnet fork to see what actually fills. The cheapest quotes were often the least reliable, reliability can degrade sharply with order size, and the one that almost never fails charges roughly 50x the rest. Explore reliability and realized cost across four trade sizes.

The Tightest Fill On-Chain Is a PMM, Not an AMM

Aggregate execution tables reward size: the venues with the most trades set the mental baseline, and a deep constant-product pool gets treated as about as tight as on-chain execution gets. Ranking sampled venues by realized slippage instead of by volume flips that. The tightest fill in the dataset is not the largest AMM or any aggregator. It is a proactive market maker (PMM).

Key findings (all chains, on-chain execution data):

• DODO posts the tightest median on-chain slippage of any venue sampled: 1.37 bps across 8,423 trades, roughly 8.3× inside the direct-Uniswap baseline (11.43 bps, 1,283,914 trades).
• It is tighter than every aggregator in the set; the best-executing aggregators this window are Bitget's DEX aggregator (4.79 bps) and 1inch (5.73 bps), both several basis points wider.
• The three tightest venues are all non-constant-product designs: DODO (1.37 bps) and woofi (2.67 bps) are PMMs, fluid (3.68 bps) is CLOB/smart-debt adjacent. A PMM prices from professional market makers around an external reference rather than walking a fixed curve, which collapses slippage on liquid pairs. Execution quality tracks routing topology.
• Honest bound: DODO's figure comes from a much smaller book (8,423 trades vs. Uniswap's 1.28M), so it reflects the pairs and sizes it actually served this window, not Uniswap's breadth. The claim is that the tightest fills are structural, not that a PMM beats an AMM on every pair.

RFQ at Size: One Aggregator Routes 86% of $1M Flow Off-Chain

The standing objection to off-chain RFQ routing is quote fade: a market maker shows a firm-looking price and delivers something worse by the time the fill lands, and the bigger the trade, the bigger the temptation. ClearTrace's quote sampler now has enough rated data on Ethereum to put a number on that fear, because exactly one sampled aggregator actually leans on RFQ at size.

Key findings (Ethereum, rated quote-sampler cells + on-chain data):

• At $1M trade size, KyberSwap routes 86.3% of sampled flow through off-chain RFQ makers (214 of 248 samples). Every other sampled aggregator (1inch, Bebop, Odos, ParaSwap, OpenOcean, Sushi, and the direct-Uniswap baseline) shows 0% RFQ share at that size. Across all sizes, its RFQ share is 20.6%.
• Quote accuracy does not degrade with size. The $1M tier realizes within 1.12 bps of quote (186 rated samples), tighter than the mostly-AMM $10k and $100k tiers (4.60 and 8.31 bps), precisely where routing flips to RFQ.
• Isolating the RFQ-routed $1M fills, the median gap is 2.69 bps (155 samples) versus roughly 0 bps for the thin AMM remainder (31 samples, indicative only). Committing 86% of $1M flow to off-chain makers costs a couple of basis points of quote fidelity, not the tens of basis points the quote-fade objection imagines. KyberSwap's overall rated gap is 2.61 bps.
• A fork simulation cannot replay an off-chain RFQ commitment, so the citable reliability measure is on-chain: KyberSwap's revert rate is 2.66%, below the direct-Uniswap baseline's 3.83%. The RFQ-heavy routing is not showing up as failed fills.

Update (2026-07-14): This report is withdrawn. Its RFQ-routing shares are a detection artifact. ClearTrace's quote sampler only parsed the RFQ flag from KyberSwap's API format, so KyberSwap read ~86% while every other aggregator (including heavy RFQ users like 1inch) read exactly 0.0%. A metric where most venues are exactly zero is unmeasured, not zero, so the RFQ metric has been removed from all published ClearTrace surfaces. See the methodology page.

The Uniswap Default: What Aggregators Are Actually Beating

ClearTrace's leaderboard flags one row as a baseline: Uniswap, traded directly, no aggregator in the path. That gives a clean yardstick for the question aggregators exist to answer: does routing through one actually beat just swapping on Uniswap?

Key findings (Ethereum, on-chain execution data):

• Direct Uniswap swaps cost a median 10.75 bps against the minute-level off-chain reference price (278,010 trades, 3.83% revert rate). That is the bar every aggregator below is measured against.
• Only two of eight sampled aggregators beat that baseline on execution cost: 1inch (5.49 bps) and Bebop (9.30 bps, on a thinner 1,259-trade sample).
• Odos ties the baseline on cost (10.42 bps); its 17.41% headline revert rate is a documented bot-spam residual: address-rotating spam the v5 sender filter cannot fully separate dominates that cell, and the revert rate genuine Odos users experience is ≈0.2% (see the methodology page). CoW Protocol has the lowest revert rate in the sample (0.04%) but higher execution cost (15.19 bps), so cost and reliability don't move together.
• KyberSwap routes 87.2% of simulated $1M flow through off-chain RFQ desks (vs. 20.5% averaged across all sampled sizes), yet holds a below-baseline 2.66% revert rate, so its heavy RFQ routing at size doesn't appear to cost it reliability.

Update (2026-07-14): The off-chain-RFQ-routing figure above is withdrawn. ClearTrace's sampler only parsed the RFQ flag from KyberSwap's quote format, so KyberSwap read high while every other venue read exactly 0.0%, a detection artifact rather than a measurement. The RFQ metric has been removed from all published ClearTrace surfaces; this report's other findings are unaffected. See the methodology page.

DEX Aggregator Execution Scorecard: Edition 3 (September 2026)

The third rated edition. It decomposes every headline revert rate into the part that is the user's own failed swap and the part that is not, and reports the two side by side for every venue. Revert rates are ranked only within an execution model; batch-auction and intent venues are shown but flagged not-comparable, never ranked against a router.

Edition 3 findings (Ethereum, on-chain 7-day window to September 8, 2026):

• Most of what a headline revert rate counts is not the user's own failed swap. Across the 9 pooled routers, the non-user term accounts for more than half the headline difference in 31 of 36 venue pairs.
• User-only (v6) revert rates span 0.80 percentage points; headline rates span 4.06. Across the 8 routers where the rate is comparable and uncontaminated the v6 rates run 0.23% (1inch) to 1.03% (OKX), and 7 of the 8 route measurably more reliably on that rate than the direct-Uniswap default (1.04%). OKX is lower but within measurement error. The v6 rate is over senders whose own transactions succeeded at least 90% of the time over the window, a threshold on outcomes rather than a classification of senders, bounded under 10% by construction.
• Most aggregators do not beat the venue default. Against the direct-Uniswap execution baseline (10.93 bps median), 5 of the 11 venues we measure on-chain deliver cheaper realized execution: Bitget DEX (4.14 bps), 1inch (7.45 bps), OpenOcean (9.34 bps), CoW Protocol (9.66 bps), Bebop (10.74 bps). Three venues carry no on-chain execution-cost data this edition (Odos, LI.FI, DODO X) and are listed unmeasured rather than ranked.

DEX Aggregator Execution Scorecard: Edition 2 (August 2026)

The second rated edition, and the first to tag every venue with its execution model and report the v6 user-only revert rate beside every headline (senders whose personal revert rate is under 10%: a threshold on outcomes, not a classification of senders). Revert rates are ranked only within an execution model; batch-auction and intent venues are shown but flagged not-comparable, never ranked against a router.

Edition 2 findings (Ethereum, on-chain 7-day window to August 5, 2026):

• Most of what a headline revert rate counts is not the user's own failed swap: across the 7 pooled routers, the non-user term accounts for more than half the headline gap in 18 of 21 venue pairs. The 4.5× headline gap between OpenOcean (5.87%) and Bitget DEX (1.31%) is substantially a difference in traffic mix; on the user-only rate the same two are 0.76% and 0.47%, a 1.6× gap.
• User-only (v6) revert rates span 0.71 percentage points; headline rates span 4.56. Across the 6 comparable routers the v6 rates run 0.29% (SushiSwap) to 1.00% (KyberSwap); 4 of the 6 beat the direct-Uniswap default (1.05%) at p < 0.05, and two are lower but within measurement error.
• Only 3 of the 8 venues with an Ethereum execution-cost measurement beat the direct-Uniswap baseline (8.82 bps): Bitget DEX (2.24), Bebop (6.66), and 1inch (6.90).

This edition carries the dated errata for Edition 1: the withdrawn cross-model reliability finding, the move from revert methodology v4 to v5+v6, and the removal of the RFQ-at-$1M column. Edition 1's snapshot and hash are unchanged.

DEX Aggregator Execution Scorecard: Edition 1 (July 2026)

The first rated edition of ClearTrace's recurring execution scorecard: five aggregators with rated quote-accuracy cells joined to on-chain execution, revert, and RFQ-routing data: frozen, dated, and SHA-256-hashed so the record can't be back-dated. The neutral, citeable alternative to self-reported "best price" and reliability claims.

Edition 1 findings (Ethereum, frozen July 4, 2026):

• Odos posts a rated 0.0 bps median quote gap, and a 23.78% on-chain revert rate (~1 in 4 routing transactions fails, 14.8× the direct-Uniswap baseline). Quote accuracy does not imply reliability.
• KyberSwap routes 86% of sampled $1M flow via off-chain RFQ desks, yet posts the widest rated quote gap (2.87 bps).
• Only three aggregators beat the direct-Uniswap execution baseline (10.54 bps): Bitget DEX (2.32), 1inch (7.02), and Bebop (9.73).

Update (2026-07-05): A post-publication audit found the Odos revert figure is dominated by address-rotating solver bots, not failing user transactions (user-only rate ≈0.2%). The Edition 1 record stays frozen per its integrity model and carries a dated note; Edition 2 will use revert methodology v5. The other findings are unaffected. See the methodology page.

Update (2026-07-14): The RFQ-routing finding above is withdrawn. Its per-venue RFQ share was a detection artifact. The sampler only parsed the RFQ flag from KyberSwap's format, so every other venue read exactly 0.0%. Edition 1's frozen snapshot stays unchanged per its integrity model; Edition 2 carries the dated correction and drops the RFQ column. See the methodology page.

Quote Accuracy Results: Near-Perfect Quotes, Mixed Execution

One week after launching ClearTrace's fork-simulation quote sampler, the first rated cells are in: 3,557 realized samples across six aggregators over seven calendar days. Quote accuracy is higher than expected: five of the six rated aggregators deliver near-perfect quotes. The more important finding is what that accuracy doesn't predict.

Key findings (Ethereum, rated quote sampler cells + on-chain data):

• Most rated aggregators post ≤1 bps median quote gap: 1inch at 0.0 bps (598 samples), Odos at 0.03 bps (528 samples), ParaSwap at 1.0 bps (623 samples). Most aggregators quote almost exactly what they fill.
• KyberSwap is the outlier: 3.78 bps quote gap, widest among rated aggregators, despite routing 88.5% of $1M flow via off-chain RFQ desks (Hashflow / Native RFQ).
• The sharpest paradox: Odos posts 0.03 bps quote accuracy (528 rated samples) but a 19.86% on-chain revert rate, roughly 1 in 5 transactions fails before filling. Quote accuracy and execution reliability are separate dimensions.
• Among rated aggregators, only 1inch beats the direct-Uniswap execution baseline on cost (7.02 bps vs. 10.54 bps baseline) while also holding 0.0 bps quote accuracy and a 2.62% revert rate.

Update (2026-07-05): A sender-concentration audit found the Odos revert figure above was dominated by address-rotating solver bots, not user transactions failing. The user-only rate (senders whose personal revert rate is under 10%) is ≈0.2%. Revert methodology moved to v5 (stronger sender-level bot filter) on 2026-07-05; see the methodology page. The quote-accuracy findings are unaffected.

Update (2026-07-14): The off-chain-RFQ-routing figure above is withdrawn. ClearTrace's sampler only parsed the RFQ flag from KyberSwap's quote format, so KyberSwap read high while every other venue read exactly 0.0%, a detection artifact rather than a measurement. The RFQ metric has been removed from all published ClearTrace surfaces; the quote-accuracy findings are unaffected. See the methodology page.

Aggregator Execution Quality: Why Lowest Revert Rate ≠ Best Outcome

A common shortcut for evaluating DEX aggregators is to compare revert rates, on the assumption that lower is safer. The on-chain data tells a more complicated story: the aggregators with the best revert rates post the worst execution costs, and vice versa. Using ClearTrace's latest Ethereum data and our newly launched quote-accuracy sampler, we map the full trade-off landscape across 14 aggregators.

Key findings (Ethereum, on-chain data):

• Tokenlon posts the lowest revert rate of any aggregator in our sample (0.07%, 2,339 trades), yet its median execution cost is 64.48 bps vs. the minute-level off-chain reference price, more than 6× the direct-Uniswap baseline.
• Only two aggregators beat swapping on Uniswap directly (10.54 bps baseline): Bitget DEX at 2.32 bps and 1inch at 7.02 bps, both at the cost of higher revert rates.
• Odos and OpenOcean show revert rates of 19.86% and 11.69% respectively: roughly 1 in 5 or 1 in 9 transactions fails before filling.
• At $1M order size, KyberSwap routes 93.8% of flow through off-chain RFQ desks (Hashflow / Native RFQ), compared to 0% for 1inch, Odos, and ParaSwap, a routing topology shift that aggregate statistics miss.
• Quote accuracy (quoted vs. realized) is accumulating via our new fork-simulation sampler; no cells are rated yet. First rated results expected around June 30.

Update (2026-07-05): The Odos (19.86%) and OpenOcean (11.69%) figures above were computed under revert methodology v4 and are dominated by solver-bot spam, not failing user transactions. A sender-concentration audit puts the user-only revert rate (senders with a personal revert rate under 10%) near 0.2–0.4% for both. Methodology v5 (2026-07-05) filters these senders; the trade-off thesis of this report stands, but those two headline rates should not be cited. See the methodology page.

Update (2026-07-14): The off-chain-RFQ-routing figure above is withdrawn. ClearTrace's sampler only parsed the RFQ flag from KyberSwap's quote format, so KyberSwap read high while every other venue read exactly 0.0%, a detection artifact rather than a measurement. The RFQ metric has been removed from all published ClearTrace surfaces; this report's trade-off thesis is unaffected. See the methodology page.

Update (2026-09-10): The Tokenlon execution-cost figure above is withdrawn, and with it the "6× the direct-Uniswap baseline" comparison. Most of it was never a cost anyone paid. Dune's Tokenlon spellbook reads the signed order's quoted amount as the fill, so the number ClearTrace measured against was a quote no transfer in the transaction matches. Decoding every AMM-route fill of a recent seven-day window from its receipt, takers received a median 47.4 bps more than the source records; recomputed against what they actually received, the venue reads about 25 bps rather than the 66 bps then published. Roughly 30 bps of what remains is Tokenlon's own disclosed fee, leaving a few bps of genuine routing cost. Tokenlon's execution cost has been withdrawn from every ClearTrace surface pending a source that records settled amounts; its revert rate, the other half of this report's finding, is unaffected and stands. The trade-off thesis stands for the report's other venues, but Tokenlon should no longer be cited as its illustration.

DEX Aggregator Execution Quality Benchmark: June 2026

Cross-frontend comparison of execution across major DEXs and aggregators on Ethereum, Base, Arbitrum, and Optimism. We measure realized slippage versus a minute-level off-chain reference price (Dune prices.usd, an interpolated 5-minute exchange-aggregate feed and not a VWAP) and transaction revert rates directly from on-chain trace data, reported as separate metrics, not a single blended score.

Findings from the latest sync (figures are live on the dashboard):

• Revert rates vary widely by aggregator: CoW Protocol records the lowest (~0.1%), versus KyberSwap (~3.8%) and OpenOcean (~11.7%).
• On slippage, 1inch shows among the tightest median execution of the major aggregators (~7.5 bps), versus KyberSwap (~13 bps) and CoW Protocol (~15 bps).
• Lowest revert and lowest slippage aren't the same protocol, because execution quality is multi-dimensional, which is why we publish each metric separately.
• Roughly half of analyzed routing volume traces to identified frontends; the remainder flows through unattributed proxy contracts we are actively resolving.

Cross-Frontend Attribution: Uncovering Hidden DEX Volume

Technical deep-dive into ClearTrace's 4-vector attribution engine: calldata suffix trapping, proxy-router hunting, multi-hop origin tracing, and fee-recipient attribution. Learn how we identify hidden mobile apps, institutional smart wallets, and anonymous frontends that traditional analytics miss.

Techniques Covered:

• Parsing raw calldata for hidden hex fingerprints
• Detecting proxy contracts via tx_to mismatches
• Window functions for behavioral sequence analysis
• Fee-recipient clustering for anonymous interface attribution