The reusable idea in this paper is scoring the three calibration outputs against each other instead of one at a time. At matched 1.02 ms latency, that change takes leverage RMSE from 0.127 to 0.098, the independent flux residual from 0.203 to 0.154, and repricing from 136 to 115 bps, against a leverage-only network of the same speed. No architecture story is needed to explain it. Forcing the heads to agree is worth 15 to 24% depending on which of those three errors you look at.
A note on what we could and could not do with it. The paper is generic and synthetic rather than tied to any exchange-traded market, so we took the vanilla-surface calibration idea across to US listed equity and ETF options. The forward-start and cliquet tests could not transfer, because those instruments are not in our data. Our data is end-of-day, so intraday calibration latency is untestable here. And we ran a residual-coupled IV/Dupire substitute rather than the paper's DeepONet/FNO solver. Our run is not a replication and not a test of the paper's claims.
What the operator replaces
A desk calibrating local-stochastic volatility runs three engines in sequence. Fit an arbitrage-admissible implied-volatility surface to the quote book. Differentiate it to get Dupire local variance. Then solve for the leverage function. That step is where the pain lives. Gyöngy's projection says leverage squared times the conditional moment of stochastic variance equals local variance. The conditional moment is computed under dynamics that already contain the unknown leverage, so the calibration is a fixed point in laws. Particle methods solve it with Monte Carlo and kernel error, sequentially in time, and the whole chain reruns for every surface, backbone and scenario.
Wang, Després, Dureau and Buet-Golfouse learn one conditional map from a quote book and a stochastic-volatility backbone descriptor to all four fields at once: log implied variance q = log I², Dupire local variance, the conditional moment, and leverage squared. A permutation-invariant Deep-Sets encoder reads the unordered quotes. A DeepONet branch-trunk decoder (or an FNO) writes the fields, and softplus and exponential heads keep them positive. Training minimizes a normalized residual system rather than a label loss: bid-ask distance, static-arbitrage margins, a division-free Dupire residual derived in log-implied-variance coordinates, and the projection residual. Autodiff supplies the derivatives, so the Dupire term binds between quote nodes.
Everything is synthetic. Forty-eight training and twelve held-out rough-Bergomi market states, a Heston-type application backbone, quote grids of six maturities by eleven log-moneyness nodes, with main-suite metrics restricted to |k| ≤ 0.25 and T ≥ 0.2. The authors verified the numerical plumbing before fitting anything: Black inversion error 3.5e-16, rough-Bergomi Volterra covariance error 1e-16, price-coordinate versus IV-coordinate Dupire agreement tightening from 2.02% to 0.043%, Heston Lewis-versus-Monte-Carlo agreement of 7.6 bps.
The 164x shrinks to 2.53x
Calibration latency falls from 98.5 ms per surface for the particle solve (N = 6,000, 50 steps, including its own 0.53 ms Dupire finite-difference input) to 0.60 ms for the operator, paired on the same hardware. The paper is direct about what that does and does not buy. Adding the shared 512 ms two-dimensional ADI exotic solve on a 96 x 48 x 128 grid, the complete-workflow speedup is 1.15 to 2.53x. The learned operator removes the calibration fixed point while the payoff solve stays, and the authors write it that way.
An implementer should also read the full timing tables rather than the headline. The 0.60 ms figure is a paired exotic-configuration timing. A separate 984-node checkpoint puts the operator at 2.37 ms per surface at batch 1, dropping to 0.19 at batch 16 and 0.074 at batch 64. The same checkpoint puts particle costs at 77.93, 385.43 and 783.93 ms for 2^10, 2^12 and 2^13 particles over 64 steps. The paper flags these as distinct checkpoints. Batching is where the economics sit: 0.074 ms per surface at batch 64 against 2.37 ms at batch 1, so a scenario grid is cheap and a single surface on demand is thirty times dearer per unit.
One cost is simply not on the table. The authors state that matched offline costs were not recorded, so this is an online frontier and not a total-cost claim. They give the crossing formula for when amortization pays, and it needs the offline training time, which nobody has.
How much of the gain belongs to the baseline?
The two headline field results are the ones I trust least, and the authors tell you why. On the rough-Bergomi trusted interior the IV head gives 41.5 bps RMSE versus 227.6 bps for SSVI, and local-volatility RMSE falls from 0.056 to 0.0356, a 36% reduction. But the operator is scored against the exact Dupire identity while the 0.056 comes from a noisy one-surface SSVI-Dupire estimator. Same shape on leverage: at 500 states, two FNO seeds give 0.0615 and 0.0687 against a 24-run particle mean, 7 to 16% below the 0.0735 deviation of a single particle run. The paper treats it as noise-averaging evidence only.
On prices, the honest reading is parity. Held-out vanilla repricing is 58.2 ± 3.3 bps, and that spread is two seeds rather than a confidence interval. The comparators are 60.3 for local volatility, 61.5 for Heston, and a 59 bps particle self-repricing floor. The authors say the 1.8 bps gap to the floor is not statistically meaningful. Against high-fidelity LSV truth with 60,000 particles and 150 steps, the operator's forward-start error is 2.0% versus 1.9% for the reference particle implementation, and its cliquet error 0.6% versus 0.4%. The operator is slightly worse on both, no Monte-Carlo intervals are reported, and the paper explicitly declines to call it equivalence.
The ablations are where I would spend replication effort, because they compare the method against itself. Predicting q = log I² rather than I cuts IV error from 136.4 to 102.7 bps, roughly 25%, for a coordinate change. Residual training produced zero observed calendar or butterfly violations across five seeds on the evaluated grids, versus 2.9% butterfly violations without the residual terms. Scaling from 8 to 48 training market states walks IV error 451, 186, 147, 124, 116, 103 bps and repricing 238 to 113. That curve has not flattened.
Judgment calls
The cleaning map. Theorem 1's identification runs through a fixed cleaning convention, supplied either as a cleaner input or as collocation supervision. Proposition 2 proves no exact teacher exists from finite quotes: for any interior point off the quote grid there is a smooth perturbation vanishing on every quote that moves local variance. The measured SSVI-versus-market local-volatility commutator is 10.5%. So the implementer picks what counts as truth, and that choice is not neutral.
Teacher annealing. Supervision flips sign in usefulness. With scarce labels, teachers prevent collapse and deliver local-volatility and leverage RMSE of 0.0477 and 0.1501. With full labels, teacher-heavy variants inherit about 233 bps of SSVI bias. The paper says annealing begins after validation residuals stop improving. Our end-of-day surfaces gave no clean-label target, so the teacher weight had nowhere to anneal to.
The gate thresholds. Each inference carries a seven-component diagnostic vector routing to Accept, particle Polish or Reject. The paper writes down the two quantities a deployment must report, selective risk and coverage. It then says the experiments do not estimate them, so the vector is not a calibrated confidence certificate. The 0.59 correlation between the projection residual and the particle residual in the vol-of-vol sweep is described as preliminary ranking evidence. With coverage unmeasured, the 0.60 ms figure is an upper bound on what a desk would actually realise.
The domain. Accuracy outside the trusted interior degrades to 175 bps at the boundary and short maturities against 41.5 inside, and repricing runs 107, 139 and 182 bps on interpolation, out-of-distribution and stress boxes.
On training itself, we did not find network widths, optimizer settings or a training budget in the paper's text. And the residual weights, curriculum stages, margin floors and out-of-distribution ridge parameter are chosen on validation data drawn from the same synthetic generator as the test states.
The authors do not hide any of this, and they answer it. "The present results therefore support the mechanism and measured ordering, not universal dominance," they write, and the abstract claims only that "these results support amortizing the LSV fixed point: the expensive solve moves offline." That is a fair description of what 12 test surfaces can carry. My view is that a mechanism-only claim is still worth a desk's afternoon. The mechanism is cheap to test, because residual coupling across heads costs one loss term. Either the 15 to 24% joint-versus-leverage-only gain at matched latency reproduces on your own surfaces in a week, or it does not. What it does not buy is a decision to retire a particle engine, and the unmeasured Accept coverage is exactly why.
We traded a substitute
We could not reproduce this paper. Our data is end-of-day only, so intraday calibration latency and every intraday use case are out of reach. We used end-of-day option prices and implied volatilities rather than the paper's bid-ask-weighted quote residuals. The paper's evidence is synthetic LSV field accuracy and exotic pricing error, not trading performance, and forward-start and cliquet instruments are not in our data. Building the full DeepONet/FNO stack would have required an offline synthetic generator and a particle teacher. So we ran a practical substitute. It fits the arbitrage-constrained q = log IV² surface, applies the butterfly and calendar margin gates plus a Dupire residual gate, and trades the residual.
That adaptation buys options whose market IV sits at least two MAD-scaled standard deviations below the fitted surface and sells those at least two above, preferring same-underlying, same-expiry, closest-delta pairs. Top 50 US stock underlyings, 73 to 365 days to expiry, log-moneyness within ±0.25 (the paper's trusted interior, imported), delta 0.10 to 0.90. Exit on convergence at |z| ≤ 0.5, adverse widening at |z| ≥ 3.5, or gate failure. Delta hedged in the underlying at every close. We charged 50 bps of premium slippage on option legs, 5 bps on hedges, $0.65 per contract and $0.004 per share.
Over 2020-01-01 to 2025-10-08 that book returned 5.59% in total, with a Sharpe of 0.53, a Sortino of 0.78, a Calmar of 0.48, volatility of 1.83% and a win rate of 47.53%. A 0.53 Sharpe on a 1.83% volatility book is a thin result. The paper reports no strategy, no P&L, no Sharpe and no drawdown anywhere. No like-for-like comparison exists in either direction: their metric is calibration RMSE and latency on synthetic market states, ours is a listed-option relative-value book. Read our figures as evidence about the substitute.
Two of our choices matter most for the shape of that result. Net vega is capped at 2% of initial capital and delta is flattened daily, so this is a small-exposure book by construction, and both the returns and the risk figures are scaled accordingly. And the diagnostics I most wanted to exercise, the out-of-distribution distance, the independent flux metric, the projection residual and any particle LSV check, are all disabled in our version. The accept/polish/reject routing is the part of the paper a desk would actually depend on, and it is precisely the part we did not test.
First place a desk should put this
As a warm start and a scenario engine for an existing LSV stack. In the 500-state study, two FNO seeds produced leverage RMSE 7 to 16% below the 0.0735 single-run particle deviation, measured against a 24-run mean. In the separate paired exotic-configuration timing, one operator evaluation takes 0.60 ms against 98.5 ms for the particle solve. The paper already frames the Polish mode as initializing a short particle solve from the predicted leverage. That use needs no certification: a bad warm start costs iterations, not marks. Replacing particle calibration outright needs Accept coverage and selective risk on rolling historical quote books. Rolling historical books are on the authors' own production-validation list, alongside more seeds and states, exotic confidence intervals, hardware latency percentiles and targeted high-vol-of-vol failures. Selective risk and coverage they write down in equation (25) and do not estimate.
What would move me: a run on real quote books where the residual gate's 0.59 correlation with particle difficulty turns into a measured coverage number, and the boundary error stops being four times the interior error.