Take the per-proof budget apart before deciding whether it transfers. Proving is $0.003 and verifying is about $0.02 [8][9]. Those sum to $0.023, inside the guide's own ceiling of under $0.025 per proof [1][10]. Verification is therefore about 87 percent of the money [2]. Halving the proving cost saves $0.0015 [6]; the line that moves your invoice is Arbitrum gas.
The $0.02 is where the transferability question lives. 280,000 gas at $0.02 works out to roughly $0.071 per million gas [4]. The results table gives the gas and the dollar figure without stating the ETH price or L2 gas price in effect at measurement [21]. For your deployment to land under 2.5 cents, your gas has to sit near that level at settlement, and your verifier has to be the same 247 lines doing the same pairing check [3].
The proof-size row is listed as 47 KB of compressed calldata [7], sitting next to a 280,000 gas verification cost [9], and the table does not say whether the gas figure covers posting that calldata [21]. On an L2 those are two different things to pay for. If the gas number is the pairing check alone, the last row is a floor rather than a price.
The 100x claim over RISC Zero rests on 10M iterations of recursive Fibonacci, picked because it matches the oracle's cycle count [11]. The oracle's 10.2M cycles are HTTP, JSON and float math [5], and the author's stated reason for needing SP1's continuations is that a coprocessor doing HTTPS and JSON parsing is heavy on syscalls [13]. Fibonacci matches on cycle count and on nothing else. For the ratio to carry into your workload, execution has to split into segments that prove in parallel [12], and the syscall-heavy stretches must not serialise them. The figure I would quote for this oracle is the end-to-end one: 10.2M cycles in 2.3 seconds is about 4.4M cycles per second [5].
What the contract learns is narrow, and the guide states it precisely: verify this Groth16 proof, and if it is valid then price_cex > price_dex is true [18]. It is true of the bytes the guest was handed, and the guest gets its bytes by calling the endpoints named in its input struct, a Binance ticker and a 0x quote [19][2]. Proving you parsed a response correctly is a different thing from knowing the response was honest, so the promise of doing this without trusting an external oracle network [1] resolves to trading a signed feed for an unsigned fetch whose parsing is checkable.
Two parts of this are good craft. The Groth16 wrapper reuses the perpetual powers of tau, so a new circuit needs no new ceremony [14], which removes a coordination problem rather than a compute one. And full std support means reqwest, serde_json, alloy and tracing compile into a guest that is just a Rust binary targeting RISC-V [15][16]. The manifest shows where the real boundary sits: the fixed crate is pinned because Solidity has no floats [17]. Floats are fine inside the zkVM; fixed point is what crosses to the contract.
