The Risks Atomicity Hid

On a single chain, atomicity was your bodyguard. Your flash-loan arb either executed both legs in one transaction or reverted as if nothing happened — buy and sell were welded together, settled in the same block, hedged for exactly zero seconds. You could quote profit on a snapshot of the AMM curves and that snapshot was the truth, because nothing could change between leg one and leg two. There was no “between.”

Cross-chain settlement deletes that bodyguard. As the earlier lessons established, a cross-chain arb is two independent legs on two independent ledgers, and nothing forces them to settle together. Leg one commits on chain A; leg two settles on chain B some minutes later. In that gap you are exposed — naked, directional, at the mercy of the tape. This lesson is about decomposing exactly what lives in that gap.

The unifying idea is the exposure window: the stretch of wall-clock time between when your first leg is committed and your second leg settles, during which you are not hedged. Atomicity hid three risks inside a window of length zero. Stretch the window to minutes and all three wake up. Let’s name them.

The exposure window

Analogy. Atomic arb is a simultaneous swap: two people slap their cards on the table at the exact same instant, and a referee guarantees both hands flip or neither does. Cross-chain arb is a relay race in slow motion: you hand off your baton on chain A, then sprint to chain B, and only when you arrive does the trade complete. The whole time you’re running, the price is allowed to move. That run is the exposure window.

Definition. The exposure window $\Delta t$ is the elapsed time from commitment of leg one to settlement of leg two. For a cross-chain arb it decomposes cleanly into two waits:

$\Delta t = t_{\text{finality}} + t_{\text{relay}}$

where $t_{\text{finality}}$ is the time you wait for your source-chain leg to be deep enough that the bridge will trust it (confirmations), and $t_{\text{relay}}$ is the time the bridge takes to carry the message and let leg two execute on the destination.

Render the hero below and drag the sliders — the confirmations slider grows the finality wait, the bridge latency slider grows the relay, and you’ll watch both the exposure window and the illustrative price drift swell. The “atomic baseline” chip is your reminder of where this number used to live: ~0.

The math the window reintroduces

Price doesn’t drift linearly with time — it diffuses. Under the standard random-walk model, the standard deviation of the return over a window scales with the square root of time:

$\sigma_{\text{window}} = \sigma_{\text{ann}}\sqrt{\Delta t / T}$

where $\sigma_{\text{ann}}$ is annualized volatility, $T$ is one year in seconds ($31{,}536{,}000$), and $\Delta t$ is the window in seconds. Let’s plug in the canonical scenario: ETH priced in USDC, chain A at $2000, chain B at $2020 — a 1% gross gap of $20/ETH. Block time ≈ 12s, so 12 confirmations ≈ 144s of finality wait; bridge relay ≈ 90s. That gives an exposure window of:

$\Delta t = 144 + 90 = 234 \text{ s} \approx 4 \text{ min}$

Crypto annualized vol runs hot — take $\sigma_{\text{ann}} = 0.80$ (80%). Then:

$\sigma_{\text{window}} = 0.80 \times \sqrt{234 / 31{,}536{,}000} = 0.80 \times 0.00272 \approx 0.0022$

So a 1-sigma move over your 4-minute window is about ±0.22% of price — roughly $0.0022 \times 2000$, i.e. ±$4.4 on a $2000 ETH. Set that against your 1% gap ($20): one sigma is about a fifth of your edge. That sounds comfortable until you remember that a 2-sigma move (±$8.8) is routine and a 3-sigma move (~±$13.2) is a Tuesday in crypto — and an adverse 3-sigma move eats two-thirds of the gap. The edge that looked like a sure thing is actually a thin cushion against a fat-tailed coin flip.

The deep point: atomic arb’s exposure window is ~0, so $\sigma_{\text{window}} \approx 0$ — risk-free in the model. Cross-chain arb’s window is minutes, and because risk scales with $\sqrt{\Delta t}$, you can’t make it vanish by being clever; you can only trade it down by waiting less, which (we’ll see) conflicts with waiting enough for safety.

When it matters

The exposure window dominates your risk whenever the gap is thin relative to volatility and the window is long. A fat 3% gap on a low-vol pair with a 30-second window barely cares about drift; a 0.4% gap on volatile ETH with a 6-minute window is mostly a volatility bet wearing an arbitrage costume. Always size the window against the gap before you size the trade.

Inventory risk

Analogy. You agree to buy a house for $2000k and flip it to a waiting buyer for $2020k — $20k locked in, supposedly. But you have to close on the purchase before the resale paperwork clears, and the housing market moves while the lawyers shuffle documents. For those weeks you simply own a house, fully exposed to the market, regardless of your tidy flip plan. Inventory risk is owning the thing in the gap.

Definition. Inventory risk is the price risk of the asset you are holding between legs. After leg one, you have a real, settled position — long the asset you bought (or short the asset you sold and can’t yet offset). Until the offsetting leg settles, that position is naked directional exposure. Its P&L per unit of adverse move is the full position size, not some hedged fraction.

Worked example

Buy 100 ETH on A at $2000 (cost $200,000), intending to sell on B at $2020 for $202,000 — a gross edge of $2,000. Now suppose ETH drifts down 0.5% during the window. The price you’ll actually sell at on B tracks the market, so it slides too:

A 0.5% adverse move erased half your edge. Push the adverse move to 1% and the loss on inventory is $-0.01 \times 100 \times 2000 =$ −$2,000, which exactly cancels the gap — you did all that work for zero. Any drift past 1% and the “arbitrage” is a loss. Note the asymmetry: the gap is fixed at $2,000, but the inventory loss scales without limit as the move grows.

When it matters

Inventory risk scales with position size × adverse move, so it bites hardest when you’re pushing size to make a thin gap worth the gas. Doubling your ETH to double the $2,000 edge also doubles the dollar loss from any drift — the risk grows in lockstep with the reward, which is precisely why blindly scaling up a thin cross-chain gap is a trap.

Latency risk

Analogy. You see a great fare on a flight and start a slow checkout that takes ten minutes to process. Two things can go wrong: the airline reprices the seat (the gap reverts), or someone else grabs the last seat first (a competitor closes it). Either way, the deal you saw isn’t the deal you get. Latency risk is the price of being slow in a race where the prize disappears.

Definition. Latency risk is the risk that the time to settle leg two is long enough that the opportunity changes character before you capture it — the gap reverts, or a competing arbitrageur (racing the same public gap) closes it first. Unlike inventory risk, which is about price moving against your held position, latency risk is about the opportunity structure itself decaying during the wait. Bridges are the slow link: a fast liquidity-network bridge clears in seconds-to-minutes, while canonical or optimistic bridges can take minutes-to-hours.

Bridge latencies and capture probability

The intuition in numbers: suppose a given 1% gap historically persists for an average of ~3 minutes before competition closes it. A 30-second bridge captures it almost always; a 4-minute bridge captures it maybe half the time; an optimistic bridge with an hours-long window captures it essentially never. Your expected profit is the gap times the probability the gap is still alive when leg two lands — and that probability falls steeply with latency.

When it matters

Latency risk dominates for fast-mean-reverting gaps and crowded routes. A structural, slow-moving mispricing tolerates a slow bridge; a fleeting micro-gap on a heavily-watched pair is a foot race you’ll lose the moment your bridge is slower than the field’s. Match the bridge speed to the gap’s half-life, or don’t take the trade.

Probabilistic finality & reorg risk

Analogy. Probabilistic finality is ink that’s still wet. On chains where recent blocks can be reorged (reorganized — replaced by a competing chain segment), the last few blocks aren’t truly settled; they can be peeled off and rewritten. Acting on the destination before the source is dry is like shipping a signed contract before the other party’s signature has actually set — if their ink smears away, you’ve shipped against nothing.

Definition. Reorg risk is the risk that the source-chain block containing leg one gets reorged away after you’ve already acted on the destination, leaving you with a one-legged (naked) position. Confirmations are your defense: each additional block buried on top of yours lowers the probability it can be reorged. Ethereum reaches full economic finality only after 2 epochs (12.8 minutes); acting on a single confirmation is acting on something that is statistically likely but absolutely not final.

The cruel trade-off: confirmations buy safety against reorgs but cost time, and time is the exposure window. More confirmations → lower reorg probability → but a longer window → more price drift. You cannot minimize both at once. This is exactly the tension the timeline island makes physical — drag the confirmations slider up and watch reorg-safety improve while the exposure window (and the drift readout) gets worse.

Worked example: confirmations vs. drift

Block ≈ 12s, 80% annual vol, 90s bridge relay held fixed. Compare confirmation depths:

Going from 1 to 64 confirmations cuts reorg risk dramatically but triples the expected drift cost (0.14% → 0.42%) — which against a 1% gap is the difference between a comfortable cushion and a coin-flip. The “right” depth is wherever the marginal reorg risk you remove stops being worth the marginal drift you add.

When it matters

Reorg risk matters most on chains with weak or slow finality and when the source leg is the one you commit first. On a fast-finality destination you can be aggressive about acting; on a probabilistic source you must wait out confirmations — and that wait, not the bridge, is often the bigger slice of your exposure window for volatile assets.

Putting it together: the risk-adjusted gap

Analogy. This is the cost-stack mindset from the on-chain arb course, with new floors added to the building. There, your gross gap had to clear gas + slippage + fees. Here the window stacks two more floors on top: expected adverse drift and a tail allowance for the rare reorg/hack catastrophe. You only profit if the gap is taller than the whole stack.

The inequality. Let $g$ be the gross gap (as a fraction of price), $f$ the bridge fee, $d$ the expected adverse drift cost, and $\tau$ a tail allowance. The trade clears only if:

$g > f + d + \tau$

Worked decision

Take the canonical 1% gap and stack the costs:

Thin but positive — $0.43 of edge per $100, before you’ve even priced the gas on both chains. Now run a 0.3% gap through the same stack:

$0.30\% - 0.05\% - 0.22\% - 0.30\% = -0.27\%$

Negative. The 0.3% gap looked like free money on the screen and is a guaranteed loser once the window’s costs are honest. The lesson: cross-chain arb has a much higher minimum viable gap than atomic arb, because atomicity used to zero out $d$ and shrink $\tau$ — and the bigger your bridge latency and confirmation depth, the higher that minimum climbs.

When it matters

Always — this inequality is the gatekeeper for every cross-chain arb. The terms shift with the route (fee), the asset’s volatility and your window (drift), and the chains’ finality and bridge security (tail), but the discipline is constant: never take a trade whose gap doesn’t clear the full stack with room to spare. The thinner the gap, the more ruthlessly you must price $d$ and $\tau$.

Recap