Failed Arbitrage Attempts on Base: Bot Competition or Isolated Search?

Arbitrage bots on fast, cheap rollups send many transactions that revert or do nothing. A recent study left open whether these failed attempts come from bots competing for the same opportunity or from each bot searching on its own. I test this on Base with execution traces of sampled blocks. Among non-probe transactions of bot contracts, at equal entity and position in the block, the revert rate over 48 weeks of 2025 is 7.75 percentage points higher (95% CI 6.56 to 9.28) when another entity swapped earlier in the block on one of the same pools. In 48 weeks of 2024 and 2026 that I had not seen, registered with an external timestamp, the gap is 13.6 points (11.4 to 15.8). A prior swap on pools the transaction does not touch shows no detectable change (+0.1 points, -0.5 to 0.7), and a prior swap by the same entity is followed by a lower rate, which I did not predict. A later swap on the same pools goes with a revert rate 1.1 points higher (-0.1 to 2.2). That upper limit is above the 2-point margin fixed for calling it null, so the joint reading fixed in advance is inconclusive: the data do not separate a race for pool state from a shock common to the pool. The gap shrinks as the bot definition tightens and is null in 2025 when a bot must fail on 75% of its transactions. It also does not survive when each contract's pools are fixed from its successful swaps (null in 2025, inconclusive in the fresh weeks), so selection through the pool set is not ruled out. The same-block part of cross-entity pool overlap is bounded above by 1.3 or 2.8 points (upper limits 1.4 and 3.4), depending on how Uniswap v4 pools are counted. On Unichain and OP Mainnet the tests are inconclusive for lack of comparable cells. I retract a figure of about 90% contention that I posted earlier; it counted one bot repeating itself. All results are associations. Version 1.2 (2026-10-25) corrects two figures and five sentences of version 1.1. The sensitivity forest and the cross-chain forest showed cell counts and estimates from before a fix to the race code (for example 31 Unichain cells instead of 169); the text and tables were already correct. Five sentences on related work and chain mechanics now say no more than their sources. No number in the text or tables changes.