Arbitrage research — measured, and mostly negative

The question. Does a $100 retail account have any tradeable arbitrage edge in crypto?

What was measured. 10+ arbitrage classes — triangular, stablecoin depeg/cross-rate, cross-venue spot, single-venue and cross-venue funding carry, maker spread capture, CEX↔DEX (native and bridged), Solana and multi-chain meme cross-DEX, and a public "$0.4–0.7/min" earnings claim — on real public order books, against the project's own fee/slippage cost model, at a $100 notional basis.

The verdict. No class is net-positive for a $100 retail taker. The blocker is the retail fee/spread floor, not the absence of signal: where a gross dislocation exists it is smaller than the round-trip cost, or it is not an arbitrage at all but a structural basis (bridge risk), or it is statistically real but economically trivial and capacity-bound below the trade size.

The selling point is the methodology. Along the way the work caught three of its own measurement artifacts before they became false claims (see Methodology).

Paper only. Measurement only. No orders, no keys, no wallets, no transfers, no account state. Public unauthenticated endpoints only. Every number is traceable to a file in evidence/ and reproducible from scripts/.

---

Verdict table

#	Class	Measured edge	Cost floor	Net	Verdict
1	Triangular (spot)	best mean net −28.5 bps; 0 of 100 sampled triangles (of 354 discovered) ever positive	30 bps RT taker (6 bps maker what-if)	negative	NO
2	Stablecoin cross-rate / depeg	round trips −20.1 to −30.0 bps; widest genuine spread Gate TUSD ~10 bps; no depeg in window	20 bps RT taker	negative	NO (rare-event bet only)
3	Cross-venue spot (same pair, 2–3 venues)	max gross dislocation +11.0 bps; 0 of 30 directions net-positive	≥20 bps RT (2× taker)	negative	NO
4	Single-venue funding carry	best hedgeable ZROUSDT $0.0221/day; decisive survivors $0.040–$0.070/day	30 bps taker / 24 bps maker hedged RT	~$0.02–$0.07/day per $100	marginal / illiquid-corner only
5	Cross-venue funding spread	325 of 1,674 pairs pass the statistical gates; best BTW $0.0815/day gross, $0.0775/day net; median candidate ~$0.0073/day net	4 fee events (2 perp RTs) + measured half-spreads	tiny; top-ranked capacity-bound	statistically real, not collectable at $100 for top names
6	Maker spread capture	best VIP0 case −9.64 bps/fill; median spread < fee on 22/24 pairs	VIP0 maker == taker == 10 bps/leg	negative everywhere	NO at VIP0 (0/24); 2 bps tier 1/24 = NOT ACHIEVABLE
7	CEX-DEX native/deep	1–3 bps gap (SOL, USDT/USDC)	150.5 bps @$100 (42.05 @$1,000)	negative	NO
7b	CEX-DEX bridged	41–100 bps "gap"	150.5 bps @$100	structural basis, not mispricing	NOT an arbitrage — bridge-risk basis, captured by MEV searchers
8	Public "$0.4–0.7/min" claim	$576–$1,008/day required	—	576–1,008%/day on $100	impossible for a $100 account; unverifiable as stated
9	Solana meme cross-DEX (90 venue-pairs)	0 of 90 cross-venue venue-pairs ever net-positive; widest gross 4.78 bps; best net −5.43 bps; $0.0000/day	pool fee 8–60 bps RT (mean 34) + 1–2 bps on-chain + impact <5 bps = ≈35–60 bps required	negative	NO
10	Multi-chain meme cross-pool (22 chains measured, 8 with candidates)	positives only in pools ≤$70k; capacity $76–$577; 1.3–14.6% of samples	per-chain gas + pool fee + impact	capacity-bound, transient	NOT a strategy
11	CEX↔DEX meme (282 verified samples)	0.0% net-positive; best −7.2 bps	10 bps CEX taker + 10 bps buffer + pool fee + gas	negative	NO
—	WARP expanded venues (9 venues)	best T $0.354/day — FALSIFIED (artifact #2)	4 fee events + half-spreads	non-executable (different assets)	artifact, not an edge

One-line summary: the retail fee floor, not the absence of signal, defeats every class; the only statistically rich class is a funding spread that is real, tiny, and capacity-bound below the trade size.

---

Methodology — three artifacts we caught on ourselves

This is the part worth reading. Each of these would have been a headline "edge"; each was caught and killed by an explicit identity/partition guard before it became a claim.

1. DEX symbol collision — three different TRUMP contracts

Unguarded symbol grouping produced phantom edges. On ethereum alone, search returned three distinct TRUMP contracts priced 0.03 / 2.34 / 5.05 — a symbol-grouped "spread" of ~100,000 bps that is really three different assets.

guard	robinhood headline	ethereum CEX↔DEX
grouped by symbol (first draft)	2,932 bps	1,287 bps
grouped by (chain, contract address)	493 bps	—
+ CEX↔DEX canonical-contract allowlist	—	45 bps

A collision is a fabricated order-of-magnitude edge, not a small bias. Guards: cross-pool rows keyed by (chain, contract address); CEX↔DEX rows carry identity_verified, true only for a known canonical contract. Evidence: reports/10-meme-arbitrage.md §6, reports/11-meme-multichain-cex-dex.md.

2. T ticker collision — Threshold vs a USDT/USD stablecoin perp

A WARP scan's best result (T, $0.354/day) was built by string-stripping the Kraken symbol PF_USDTUSD into base T, colliding a USDT/USD stablecoin perp ($0.9998) with Binance's Threshold token (TUSDT, $0.005414) — a 184.67× price ratio between two different assets. Verdict: FALSIFIED; corrected net $/day 0.0 (non-executable: the two legs are different assets). Evidence: evidence/verify_t/verify_t.json.

3. Venue-partition artifact — the Kraken "survivors"

A spread_series robustness pass reported 8 surviving pairs (4× BAT, 4× QNT) whose OOS bootstrap CI excluded zero in 21/21 seed×block combinations. The split was a perfect partition by venue: every pair with a kraken_futures leg scored 21/21, every pair without one scored 0/21. The falsification test showed the "spread" is a venue-mechanism level offset — Kraken's hourly relative funding rate is structurally different from the other venues' cadence — and every one of those pairs fails the programme's own same-sign gate (same_sign_share 0.39–0.47 vs the 0.60 bar). Not a token-specific, tradeable edge. Evidence: evidence/partition/partition_report.json, evidence/spread_series/robustness.json.

---

Precision rules (binding on any citation)

These phrasings are carried verbatim from the source reports and are mandatory:

• Triangular: write "0 of 100 sampled triangles (of 354 discovered) ever showed positive net edge" — never "0/354" (only 100 of the 354 discovered triangles were sampled).
• Cross-venue funding: write "325 of 1,674 pairs pass the statistical gates", then immediately state that statistical persistence ≠ collectability: the #6-ranked 龙虾 has Gate-leg capacity $75.66 < $100, and 5 candidates have capacity below $100 (BOME×2, CHIP, SKY, 龙虾).
• 龙虾_USDT 208-day re-test: survives (positive both halves, OOS CI [+5.87, +15.48]) but is regime-dependent ($0.110/day long-run vs $0.035/day most-recent-30d) and thin/illiquid — present as "thin, illiquid, regime-sensitive", never validated.
• Maker: the verdict is VIP0 = 0/24 positive; the 2 bps column (1/24) is "NOT ACHIEVABLE at $100 (high-VIP/rebate tier)"; adverse selection is a "COARSE PROXY, LOWER BOUND".
• Meme: "0 of 90 cross-venue venue-pairs ever net-positive"; "widest gross 4.78 bps vs a 35–60 bps required bar"; multi-chain positives are capacity-bound ($76–$577) and transient (1.3–14.6% of samples), NOT a strategy; the meme simulator's $/day is SYNTHETIC-fixture arithmetic only (sim_meme_SYNTHETIC.json: mode SYNTHETIC, $629.52/day — never a market claim).
• Reachability is intermittent: the same hosts answered 5/5 OK and HTTP 000 minutes apart. A two-venue strategy cannot assume simultaneous, continuous access to both legs.

---

Reproduction index — which script produces which report

Script (scripts/)	Produces
arb_common.py	shared fetch / backoff / depth-walk helpers
arb_triangular_scan.py	reports/00 §2.1, evidence/triangular/
arb_stable_crossrate.py	reports/08 (class A), evidence/stable/results.json
arb_crossvenue_spot.py	reports/08 (class B), evidence/crossvenue/results.json
arb_funding_scan.py	reports/01, evidence/funding_ranking.json
arb_funding_multivenue.py	reports/02, evidence/multivenue_ranking.json
arb_funding_decisive.py	reports/03, evidence/decisive_carry_results.json
arb_funding_spread.py	reports/04, evidence/crossvenue_funding/spread_ranking.json
arb_longxia_longhistory.py	reports/05, evidence/crossvenue_funding/longxia_longhistory.json
arb_maker_scan.py	reports/06, evidence/maker/maker_results.json
arb_dex_scan.py	reports/07, evidence/dex/
arb_warp_scan.py	reports/09, evidence/warp/
arb_meme_solana.py	reports/10, reports/12, evidence/meme/
arb_meme_multichain.py	reports/11, evidence/meme/multichain_reanalysis.json
arb_meme_exec.py	reports/12, evidence/meme/raw_pool_fees.json
arb_spread_series.py + instrument_identity.py	reports/13, evidence/spread_series/
arb_partition_test.py	artifact #3, evidence/partition/partition_report.json
arb_verify_t.py	artifact #2, evidence/verify_t/verify_t.json
arb_paper_sim.py / arb_meme_sim.py	paper simulators (sim_funding.json, sim_meme.json)
arb_claim_math.py	reports/00 §3 (the "$0.4–0.7/min" reverse-engineering)
arb_small_capital.py	evidence/small_capital/small_capital.json (venue minimums at $25/$50)

See METHODOLOGY.md for the cost model, the IS/OOS + block-bootstrap approach, and the identity guards. See scripts/README.md for how to run the scanners (they import the project's cost model; a vendored subset is included).

---

What this does NOT claim

• No profit promise. Nothing here establishes a positive expected return for any account size or fee tier.
• No validated strategy. The only statistically surviving carry name (龙虾) is presented as thin, illiquid, regime-sensitive — not validated.
• No claim that these classes are impossible in principle — only that they are not net-positive for a $100 retail taker on the fee schedules and venues measured here.
• Paper-only / measurement-only. No orders were placed, no keys or wallets were used, no account or book state is included, no strategy code is included.
• $100 retail basis. Fees are the project's base tier (spot taker 10 bps, perp taker 5 bps, maker 2 bps; slippage k=0.5, exit ×1.5). A negative-maker rebate tier, colocation or $100k+ capital is a different account, not a tweak to this one.
• Point-in-time. Windows are short (25–40 min for microstructure classes, ~30 days for funding, 208 days only for the single 龙虾 re-test). These are one regime on one host on 2026-10-03, not a permanent property of the market.
• Venue fees are partly unverified where the fee pages were unreachable; where verified, the project model under-charges (Gate perp taker is API-verified 7.5 bps vs the 5.0 bps modelled). Execution, latency, market impact and partial fills are unmodelled or proxied; adverse selection in the maker study is a lower bound.

Prior work cited (not copied)

• ~/back/FUNDING_OR_ALT.md — funding-data obtainability + a 21-symbol carry test (negative in both IS/OOS halves).
• ~/back/COST_FLOOR.md — VIP0 taker fee verified at 5.0 bps/leg; corrected RT 12 bps; no per-trade edge in the project is net-positive at corrected VIP0.
• ~/polymarket/PROFIT_REALITY.md — the same structural finding on Polymarket (YES+NO ask sum minimum 1.0010 vs the <1.000 needed for a riskless pair).

License

MIT — see LICENSE.