Swap Tests September 2026

Abstract

What the swap tests of BlazePhoenix-Dex measured, in the numbers they produced, on 2026-09-05 with release settings. A quote executed later, on mocks: 240 samples under zero to three adverse trades and zero to ten seconds of delay — under up to 1 % of drift every quote settled within 2 % of its prediction; under 1–3 % the floor refused one settlement in four rather than fill it below the attested output, and no sample produced a third outcome; after the deadline, 20 of 20 refused with the deadline’s own code. On live Base, a 1,000 USDC quote executed 0, 3, 6 and 10 seconds later delivered exactly its prediction all four times, and with 10,000 to 5,000,000 USDC traded ahead of it through the same route all five settled inside the floor at 9,974–10,000 bps of the prediction. The protocol fee, measured from outside the Router on ten route shapes with one and two legs per hop over 2,000 fuzz runs, with the Router holding nothing after every settlement; the detection rate of every fee mutant under twenty seeds with a Wilson 95 % interval; what a quote costs through the ABI on mocks and on live Base; ten hostile venues and two regime covering arrays with no third outcome in any row; the sandwich curve from the attacker’s side; and the 40 stateful invariants in 14 campaigns.

The figures, with their denominators

64 / 64

drift-free quotes settled, delivering exactly the prediction at every delay

43 / 43

settled under up to 100 bps of adverse drift — mean 9,952 bps of the prediction

102 / 133

settled under 100–300 bps of drift; the other 31 refused by the floor, none filled below it

10,000 bps

live Base, 1,000 USDC → WETH executed 0–10 s later: delivered exactly the prediction, four of four

9,974 bps

live Base, 1,000,000 USDC traded ahead through the same route, the stale calldata run 10 s later

2,000 runs

fee seals over ten route shapes, one and two legs per hop, the Router holding nothing after every settlement

0 third outcomes

hostile-venue matrix (20 rows) and covering arrays of strength 2 (63) and 3 (168): settle or refuse with a selector of ours

~2.7 %

of a 1 %-of-depth trade: what the floor caps a sandwich at before it refuses

Reading the tables

Time does not move a quote; drift does. A settlement never delivers below the floor the preview attested, a drift-free quote settles at every delay and delivers exactly its prediction, and there is no third outcome — each of these is asserted on every sample, not printed as an average. Detection rates are per test at 256 fuzz runs per seed; a test that misses at that budget may catch at a larger one. No probability of correctness follows from any table.

Companion

Every instrument of the apparatus — the mutation guard, the covering arrays, the oracles, the N-version lane, the metamorphic relations — is the Verification Report. The floor the tables keep returning to is derived in the aggregator whitepaper.

Cite this document

Mitra. (2026). BlazePhoenix-Dex — Swap Tests: Results (September 2026). BlazePhoenix. https://blazephoenix.xyz/swap-tests-report-2026-09.md

@techreport{blazephoenix_swap_tests_report_2026,
  author       = {Mitra},
  title        = {BlazePhoenix-Dex — Swap Tests: Results, September 2026},
  institution  = {BlazePhoenix},
  year         = {2026},
  month        = {09},
  note         = {Text CC BY 4.0. Machine-readable text: https://blazephoenix.xyz/swap-tests-report-2026-09.md}
  url          = {https://blazephoenix.xyz/research/swap-tests-report/}
}

DOI: assigned on Zenodo registration — until then cite the URL above; the version and date identify the text. · Machine-readable text: /swap-tests-report-2026-09.md