> ## Content Index
> Fetch the complete content index at: https://aligrithm.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# 5.64 Lead-Lag Arbitrage That Survives Latency: DAX Cross-Listings
- URL: https://aligrithm.com/lead-lag-arbitrage-that-survives-latency-dax-cross-listings/
- Published: 2026-11-07T06:00:00.000Z
- Updated: 2026-11-07T05:59:59.000Z
- Description: Lead-lag arbitrage on DAX cross-listings: a multinomial logit calls the lagging book's next quote cluster, and a one-tick limit collects it after fees and a 5 ms wire.
- Author: ali askar
- Tags: 5. Microstructure Alpha

Deutsche Bank's best bid on Chi-X moves, and the same bid on Xetra follows about 5 milliseconds later. Poutré, Dionne, and Yergeau measure that lead with an overlap correlation on the raw quote updates, then fit a multinomial logit to the sign of the lagging book's next run of updates. The trade is a marketable limit on that book, a resting limit one tick past the signal, and a stop two ticks the wrong way. The fill that pays is the resting limit. A market order on the same forecast pays the spread, and on these tapes the spread is larger than the move inside one run of quotes.

## The lag that matches the wire is the one you can trade

The old article ["Lead-Lag Relationships in Global Markets"](https://aligrithm.com/lead-lag-relationships-in-global-markets/) finds a lead by correlating one market's return today with another market's return k periods later, on a clock both series share. Chi-X and Xetra do not share a clock. Quote updates arrive at each book's own times, and forcing both onto a fast grid drives the correlation toward zero even when the two prices are the same stock. The estimator here multiplies price changes whose time intervals overlap, and it never resamples.

The lag ell, in milliseconds, is how far the second book's intervals are slid before the overlap test. Rho-hat of ell is the lagged overlap correlation. A positive ell in the Pearson definition they open with, correlation of X now with Y ell later, means X leads Y.

$$ \\hat\\rho\_{X,Y}(\\ell) = \\frac{\\sum\_i \\sum\_j \\Delta X(I\_i)\\, \\Delta Y(J\_j)\\, \\mathbf{1}\\{I\_i \\cap (J\_j+\\ell)\\neq \\emptyset\\}} {\\sqrt{\\sum\_i \[\\Delta X(I\_i)\]^2 \\; \\sum\_j \[\\Delta Y(J\_j)\]^2}} $$

Delta X on interval i is the price change of the first book between two of its own updates, in price points. The indicator is 1 when that interval shares positive duration with interval j of the second book after both of j's endpoints have been moved forward by ell. The denominator is each book's summed squared changes. Sliding the timestamps leaves those changes alone, so the ratio is unitless. The lag that maximizes the absolute value is ell-star.

Worked toy, in price points and seconds, using the shift as printed. X rises 4 over (0, 10\] and 3 over (20, 30\]. Y rises 4 over (−8, −3\] and 3 over (12, 17\]. At ell of 0 nothing overlaps, so the numerator is 0 and the correlation is 0\. At ell of +8, Y's intervals land on (0, 5\] and (20, 25\]. The products are 4 times 4 and 3 times 3, the numerator is 25, each summed square is 25, and the correlation is 25/25, which is 1\. At ell of −8 the only overlap is X's 4 with Y's shifted 3, the numerator is 12, and the correlation is 12/25, which is 0.48\. The peak is at +8, and what lined up were Y's earlier moves. Under the printed shift, a positive ell means the second book leads.

The old article ["Lead-Lag Is Multidimensional: Trade Price and OBI Predict Others' Midpoints"](https://aligrithm.com/lead-lag-is-multidimensional-trade-price-and-obi-predict-others-midpoints/) reads this same shift that way. This paper's opening sentence says the opposite: a positive ell-star means the first book leads. The table follows the sentence. Deutsche Bank, with Chi-X as the first book and Xetra as the second, has ell-star of +5 ms, correlation 0.100, and they name Chi-X the leader. That reading also matches the wire. Light and fiber between Slough and Frankfurt are a few milliseconds, and they treat any estimated lag shorter than that travel time as suspect. I keep their table. The printed shift and the sentence that defines a positive lag point in opposite directions, so one of those two lines is wrong.

The lead-lag ratio compresses the whole curve into one number. It is the sum of squared overlap correlations on the positive lags, divided by the same sum on the negative lags. Above 1, on their definition, the first book leads, and a larger ratio means a more one-sided curve.

$$ LLR\_{X,Y} = \\frac{\\sum\_{g \\in G} \\hat\\rho\_{X,Y}(\\ell\_g)^2}{\\sum\_{g \\in G} \\hat\\rho\_{X,Y}(-\\ell\_g)^2} $$

G is their grid of positive lags: every millisecond from 0 to 50, then steps of 5 out to 100, steps of 100 out to 1,000, and steps of 1,000 out to 15,000\. Whether the zero lag sits in both sums is not settled by the prose, which says positive lags, against a search grid that starts at 0\. Squared correlation is unitless, and so is the ratio.

Worked toy. Correlations 0.08 and 0.06 on the positive side, and 0.05 and 0.05 on the negative side. Squared sums are 0.0064 + 0.0036 = 0.0100 against 0.0025 + 0.0025 = 0.0050, and the ratio is 2\. Deutsche Bank's measured ratio of Chi-X against Xetra is 1.73, with the single biggest absolute correlation at 5 ms and a value of 0.100\. The ratio and the peak correlation are different objects. HeidelbergCement against BATS has the largest correlation in the table, 0.306, and a ratio of only 1.07\. Lufthansa against BATS has a ratio of 7.76 and a correlation of 0.138\. A lopsided curve can be a small curve.

Chi-X is the leader on almost every DAX 30 name against both Xetra and BATS, from February through July 2013, 125 sessions. Henkel has no Chi-X versus Xetra lead. RWE has no Chi-X versus BATS lead. The reason they give is activity and the fee, and Deutsche Bank's own counts support the activity half: Chi-X prints 211,214 quote updates a day against 119,713 on Xetra and 105,426 on BATS. Chi-X also pays a 0.15 basis point rebate to add liquidity, while Xetra charges 0.36 basis points to add it and 0.36 to remove it. The busy, cheap book leads the home market. You watch Chi-X. You trade the other two.

## A cluster is one book talking while the other is quiet

A cluster is a run of updates on one book with no update from the other book inside it. Alsayed and McGroarty used the idea on futures mids. Here a cluster can be the best bid, or the best ask, of either venue. You score the sign of the sum of price changes inside the lagging book's next cluster, once the leader's latest run has finished.

![Alternating runs of X updates and O updates, bracketed into clusters C1 of X, C1 of Y, and C2 of X](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/10/article_570-cluster_timeline.png)

$$ r\_{C\_i^Y} = \\sum\_{j=1}^{n\_i^Y} C\_{i,j}^Y, \\qquad R\_{C\_i^Y} = \\operatorname{sign}(r\_{C\_i^Y}) $$

The cluster return r is the sum of the raw price changes inside cluster i of the lagging book, in euros. The sign is +1 when that sum is positive, −1 when it is negative, and 0 when the updates inside the run cancel. You are forecasting that one sign, at the moment the leader's latest run ends and the lagger's next run is about to start.

Worked on a €30 name, whose tick in their Xetra schedule is €0.005 because the price sits in the €10 to €50 band. Two bid updates of +€0.005 and +€0.005 sum to +€0.010, and the sign is +1\. Two updates of +€0.005 and −€0.005 sum to 0, and the sign is 0\. A model that is obligated to call up or down will invent a trade on that second run. Theirs is allowed to call flat.

## Three probabilities, and a threshold that can refuse

The logit is a three-class model of that sign. The reference class is up. Two linear scores, theta-down and theta-flat, are log-odds against up. Each score has its own intercept, its own slopes on the leader's cluster returns, and its own slopes on the lagger's past cluster returns. D, the lag length, is fixed at 10 for every name and every window: on the first two weeks, lags past 12 were insignificant, lags under 10 lost accuracy, and moving D by 2 changed accuracy by less than about 2 points. That gives 4 times 10 plus 2, which is 42 coefficients, fit by maximum likelihood with BFGS, separately for the bid and the ask, and refit every five days on the days before the five you trade.

$$ \\theta\_{i,-1} = \\alpha\_{-1} + \\sum\_{j=0}^{D-1} \\beta\_{j,-1}\\, r\_{C\_{i-j}^X} + \\sum\_{j=1}^{D} \\gamma\_{j,-1}\\, r\_{C\_{i-j}^Y} $$ $$ \\theta\_{i,0} = \\alpha\_{0} + \\sum\_{j=0}^{D-1} \\beta\_{j,0}\\, r\_{C\_{i-j}^X} + \\sum\_{j=1}^{D} \\gamma\_{j,0}\\, r\_{C\_{i-j}^Y} $$

The leader sum starts at the cluster that finished, j of 0, and runs back through D−1\. The lagger sum starts at the previous cluster, j of 1, and runs back through D. The cluster being scored stays out of both sums. Cluster returns are in euros. The betas and gammas convert them into log-odds, which are unitless. The source's multinomial display writes the lag as a second subscript on cluster i. The least-squares cousin they say is the same model writes the lag as i minus j. The card follows i minus j.

$$ p\_{i,-1} = \\frac{e^{\\theta\_{i,-1}}}{1 + e^{\\theta\_{i,-1}} + e^{\\theta\_{i,0}}} $$ $$ p\_{i,0} = \\frac{e^{\\theta\_{i,0}}}{1 + e^{\\theta\_{i,-1}} + e^{\\theta\_{i,0}}} $$ $$ p\_{i,+1} = \\frac{1}{1 + e^{\\theta\_{i,-1}} + e^{\\theta\_{i,0}}} $$

The denominator is one plus the two exponentials, and the one is the reference class, up. The three probabilities sum to 1\. The source prints the down and up lines and the adding-up constraint, so the flat line is the same denominator with the flat exponential on top. The call is the class with the largest probability, and only if that probability clears a threshold K. Otherwise the call is flat, and a flat call sends no order.

$$ \\hat R\_{C\_i^Y} = \\begin{cases} +1 & \\text{if } \\max(p\_{i,-1}, p\_{i,0}, p\_{i,+1}) = p\_{i,+1} \\text{ and } p\_{i,+1} \\ge K \\\\ -1 & \\text{if } \\max(p\_{i,-1}, p\_{i,0}, p\_{i,+1}) = p\_{i,-1} \\text{ and } p\_{i,-1} \\ge K \\\\ 0 & \\text{otherwise} \\end{cases} $$

K lives in the unit interval. They search it on each training block over 0.35, 0.375, and so on up to 1, and they keep the value with the best training accuracy. That value is the peak. The traded results then walk the peak down in steps of 0.025.

Worked probabilities, after the 42 coefficients have already been collapsed into two scores. Take theta-down and theta-flat both equal to −3\. e to the −3 is 0.04979\. The denominator is 1 + 0.04979 + 0.04979 = 1.09958\. The up probability is 1/1.09958 = 0.9094\. Down and flat are each 0.04979/1.09958 = 0.0453\. Up is the maximum. At a K of 0.90 you call +1\. At a K of 0.95 you call 0 and you send nothing.

On the first training window of Deutsche Bank's bid, Chi-X against Xetra, that tradeoff looks like this. The blue accuracy line peaks near 87% short of a threshold of 0.90, and by then the red opportunity count has already fallen from about 27,000 to a couple of thousand.

![Training accuracy and the number of opportunities against the probability threshold, for Deutsche Bank's bid in the first window](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/10/article_570-threshold_accuracy.png)

Accuracy, the way they score it, ignores flat calls. A potential opportunity is a non-flat prediction. Accuracy is the fraction of those that match the realized sign.

$$ PO = \\sum\_{i=1}^{N} \\mathbf{1}\\{\\hat R\_i \\neq 0\\}, \\qquad \\text{Accuracy} = \\frac{1}{PO} \\sum\_{i=1}^{N} \\mathbf{1}\\{\\hat R\_i = R\_i \\text{ and } \\hat R\_i \\neq 0\\} $$

Both are counts over the N lagging clusters in the window. Flats never enter the denominator, so a timid model can look accurate by refusing the hard runs.

Worked on their peak-threshold row, out of sample, bid and ask together. Xetra contributes 915,666 opportunities at 84.2%. BATS contributes 708,580 at 78.3%. The counts add to 1,624,246\. Hits are 0.842 times 915,666, which is 770,990.8, plus 0.783 times 708,580, which is 554,818.1, and 1,325,808.9 divided by 1,624,246 is 0.8163\. They print 81.6%. The component percentages were already rounded, and they still rebuild the total.

Strip the leader slopes, every beta set to 0, and the same peak rule falls to 43.2% on 2,705,221 calls. Their coin-flip of the up and down classes, flats aside, is 40.10% on Xetra and 40.48% on BATS. The lagger's own history is a few points over that coin-flip. The lead is the model. They also say that dropping the lagger's own lags, and keeping the leader, trims accuracy only a little, and they do not print that figure. The least-squares cousin on the same ten lags reaches 79.6% on 1,412,282 calls. Loosen the logit to peak minus 0.125 and accuracy is 79.7%, a tenth of a point from the least-squares 79.6%, on 3,386,495 calls, 2.40 times as many.

## Sell the bid you expect to fall, and buy one tick lower

The logit is fit twice, once on best bids and once on best asks. You only act on the direction that lets a resting limit collect one tick.

On the bid, a call of −1 sends three orders at once. A marketable sell limit of size V at the current bid, so you hit the bid and you are short. A buy limit for the same size at the bid minus one tick, which is the profit exit. A stop-buy, stop and limit both at the bid plus two ticks, which is the loss exit. A call of 0 or +1 on the bid sends nothing. The source writes that idle case as the open interval from 0 to 1, which contains none of the three labels. The action list is the part that trades, and it fires only on −1\. The ask side is the mirror, and its idle set is written cleanly as down or flat: a call of +1 buys the current ask with a marketable limit, rests a sell one tick higher, and parks a stop-sell two ticks lower.

$$ \\text{sell at } b, \\qquad \\text{buy limit at } b - \\delta, \\qquad \\text{stop-buy at } b + 2\\delta $$

b is the lagging book's best bid in euros when the call fires. Delta is the tick, in euros, from the Xetra schedule that Chi-X and BATS copied: €0.001 under €10, €0.005 from €10 up to €50, €0.01 from €50 up to €100, and €0.05 from €100 up. The cash if the exit fills is one tick per share. The cash if the stop fills is two ticks the wrong way per share, and only when the stop trades. Keep one open position per side. The next signal on that side waits until it closes. If it is still open after M minutes, a market order flattens it. M is 15 minutes on Xetra and 20 minutes on BATS.

Worked at b of €30.000, so delta is €0.005\. You sell at 30.000, you rest a buy at 29.995, and the stop sits at 30.010\. On 500 shares, next to the Xetra mean size of 503.64 in their trade table, the exit is worth 500 times 0.005, which is €2.50 before fees. The stop, if it fills, costs 500 times 0.010, which is €5.00 before fees. Both tickets sit on the bid. A one-tick spread on this name is 30.000 by 30.005, and a market-order short that covers on the ask donates that spread on the way out. The old article ["Why Transaction Costs Should Be Added Before You Fall in Love"](https://aligrithm.com/why-transaction-costs-should-be-added-before-you-fall-in-love/) is this arithmetic done before anyone likes the hit rate: the gross signal has to clear the round trip, and here the round trip of a taker is the spread itself.

They make that concrete with a forecaster that is perfect on the next mid-quote cluster and still uses market orders. At a one-tick threshold the book does 11,383,116 round trips, 0.50% of them net profitable. Gross wins are €69,459\. The spread losses are €44,766,290\. Fees are €2,677,747\. The sum is 69,459 minus 44,766,290 minus 2,677,747, which is −€47,374,578\. The model made no directional error in that table. The spread did. Alsayed and McGroarty's own mid-quote rule, run on these tapes with full latency, loses €37.23 million combined. No threshold they tried on that rule finishes net positive.

![](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/10/article_570-visual-0.png)

## BATS carries the profit

Look at the orange line. That is BATS, where a resting order pays no fee, and it does most of the climbing. The blue line is Xetra. The green line is the sum.

![Cumulative net profit in euros for Xetra, BATS, and the total across the out-of-sample days](https://storage.ghost.io/c/27/cb/27cb0fc8-2c77-4434-af9e-d5d32a916994/content/images/2026/10/article_570-cumulative_net_profit.png)

Full latency, February to July 2013, threshold at the training peak minus 0.175\. BATS nets €718.41 thousand, Xetra €244.85 thousand, and the printed total is €963.26 thousand. At the euro the venue columns are €244,850 and €718,406, one short of the printed €963,257, and every other offset from the accuracy peak down through peak minus 0.200 is net positive, with this row the high of the menu.

## Charge the wire twice, and the €1.9 million is a doubled half-year

The latency they charge is two waits, and both use the same number. You sit at Chi-X. An Xetra message is held for 5 ms before the strategy is allowed to see it, and an order you send back to Xetra is held another 5 ms before it may rest. BATS is 1 ms each way. The 5 ms is their rounding of 4.150 ms of fiber plus 1.100 ms of exchange time, which sums to 5.250\. The 1 ms is a rounding of 0.165 ms up to the nearest non-zero integer, so the BATS delay is harsher than the measurement. A microwave path London to Frankfurt was already about 2.3 ms one way in that period, which means the 5 ms book is the slow wire, and the euro figures are the slow trader's.

Adding the outbound wait, the second half, cuts Xetra's net on the highlighted row from €290,017 to €244,850, a drop of 15.6%. Across the menu the Xetra cut runs from about 16% to 21%. BATS moves the other way: €681,707 becomes €718,406, a rise of 5.4%, and BATS is higher with the extra delay in every row of the table. The paper calls that "about a 5% change" and leaves the sign out. Extra delay raising the nearer venue's profit is the opposite of a pure latency cost, and the source never isolates a fill rule that would produce it. I keep both columns. The two printed totals, €971,725 without the outbound wait and €963,257 with it, differ by €8,468, and that small gap hides the Xetra hit because BATS rose. Each of those printed totals sits one euro above the sum of its venue columns.

Fees are the larger wedge, and they have to be reconciled with the size column before the euro net means anything. Xetra charges 0.36 basis points to remove liquidity and 0.36 to add it, so a winner who takes to enter and makes to exit pays 0.72 basis points of notional. Trading costs on the highlighted row are €1,528,736 across 1,381,501 Xetra trades.

$$ \\frac{1{,}528{,}736}{1{,}381{,}501} = €1.1066, \\qquad \\frac{1.1066}{0.000072} = €15{,}369.44 $$

The 0.000072 is 0.72 basis points in decimal form, a take plus a make at Xetra's 0.36\. Divide €15,369.44 by the mean size 503.64 and the implied price is €30.52 a share. Treat that size as shares and the fee schedule agrees with the trade count. Table 14 puts a euro sign on the same column and prints 338.45, which would make this fee about 22 basis points, several times the schedule. I keep 503.64 as shares. The euro label in Table 14 looks wrong. BATS lands in the same neighborhood: €192,261 across 1,226,857 trades is €0.1567 a trade. Divide by 0.000015, one 0.15 basis point take with the maker side free, and the notional is €10,447\. Divide by 352.29 shares and the price is €29.65\. Losers who get stopped pay a second take, so the true average price sits a bit under that.

On that reading the Xetra gross of €3,365,103 splits into €1,591,517 of losses, €1,528,736 of fees, and €244,850 of net. Fees take 45.4% of the gross and the losers take 47.3%. The net margin is 244,850 / 3,365,103 = 7.3%, which they round to 7%. BATS keeps 718,406 / 2,108,945 = 34.1% of its gross, and the fee there is 9.1% of gross. Same signal, two schedules. The venue with a free resting order pays you. The venue that taxes both sides keeps a sliver.

The average loser is larger than the two-tick stop. At Xetra the mean net on a winning trade is €1.79 and the mean net on a losing trade is −€8.20.

$$ \\frac{8.20}{1.79} = 4.58 $$

One loser costs 4.58 winners. The two-tick stop on 500 shares was €5 before fees, and €8.20 is past that, which is the 15-minute market flatten and the stops that miss. The win count on Xetra is 1,158,049 out of 1,381,501 fills, 83.83%, so the loser count is 223,452\. Weight the rounded means by those counts: 1,158,049 times 1.79 is €2,072,907.71, and 223,452 times 8.20 is €1,832,306.40\. The difference is €240,601, against the reported net of €244,850\. The gap is the rounding in the two per-trade means. BATS is 5.51 / 1.95 = 2.83, which they print as 2.82\. Filled trades win 83.83% of the time at Xetra, against the model's 80.6% accuracy at this same threshold.

Turn the stops, the max level-1 variation rule, and the no-microstructure-change rule off, and leave only the time flatten. Xetra net falls from €244,850 to €177,580, a cut of 27.5% (they print 27%). BATS falls from €718,406 to €441,575, a cut of 38.5% (they print 39%). The loss-to-win ratio goes from 4.58 to 14.33 / 2.14 = 6.70 on Xetra, and from 2.82 to 35.52 / 1.92 = 18.50 on BATS. Mean holding time goes from 27.82 seconds to 75.22 seconds on Xetra, against a median near 1 second when the exit fills quickly. The stop is how you get the position out of the way so the next cluster can be traded. Without it the book is still net positive.

I am against quoting €963,257 as a pre-committed result. The per-window peak of K is chosen on training data, and that row, full latency, nets €446,131\. The offset of 0.175 is the row of the test table where profit stops rising. Every row makes money. The highlighted row is still the best of nine offsets, picked after the six months were already on the page. Trade the training peak if you want the number you could have had without peeking. Hit rate at the traded offset is 78.2%. The abstract's "above 80%" is the accuracy-peak row.

I am also against the annualization. Six months at €963,257 doubled is €1,926,514, which is the "over €1.9 million" and the introduction's "all of 2013." The tape is February through July. On the 58 name-venue pairs that had a leader, 29 against Xetra and 29 against BATS, that doubled total is €33,216 a pair, which is where "more than €33,000 per stock per exchange" comes from. Spread the same euros over all 60 slots, including Henkel and RWE, and the average is €32,109, under the sentence they wrote. The daily means divide back to 113 days, 244,850 / 2,167 and 718,406 / 6,358, two short of the 115 sessions left after the first ten-day training block. Small, and unexplained.

BATS's worst printed day on this row is still a gain, €2,812\. Xetra's worst day is −€2,290\. One six-month window in 2013, with a free maker rebate, can look like that. It is a property of this tape.

What the backtest leaves out is the part you would pay in real life. Deutsche Börse's co-location menu starts near €5,000 a month in their footnote, and that price is the Xetra menu. Data fees are unstated, and Chi-X and BATS co-location are separate bills. The fill rules live in a supplement and in QuantOffice's price-time priority, so a suspicious fill has to be taken on trust. The public replication set is one month of transformed data. The fee schedule is the one in force on 2 January 2013\. Before you believe the orange line on a live book, re-estimate who leads, re-read both fee columns, and charge yourself the wire you have. These limits can reach the part of the lag slower than that wire, and only when the resting order gets to the front of the queue.

## KEY POINTS

- Measure the lead with the overlap correlation before you forecast. Deutsche Bank on Chi-X leads the Xetra bid by 5 milliseconds.
- Score the next quote cluster up, down, or flat with the ten-lag logit, and skip calls under the threshold. Out-of-sample accuracy at the training peak is 81.6%.
- Collect the call with a one-tick resting limit on the lagging bid or ask. A perfect mid-quote forecast traded with market orders loses €47.37 million.
- Judge the book on the full-latency cumulative profit. Six months net €963.26 thousand.
- Put the fee on the trade before you double the half-year. At Xetra the average loser costs 4.58 times the average winner.

## References

- [Lead-Lag Is Multidimensional: Trade Price and OBI Predict Others' Midpoints](https://aligrithm.com/lead-lag-is-multidimensional-trade-price-and-obi-predict-others-midpoints/)
- [Why Transaction Costs Should Be Added Before You Fall in Love](https://aligrithm.com/why-transaction-costs-should-be-added-before-you-fall-in-love/)
- [Lead-Lag Relationships in Global Markets](https://aligrithm.com/lead-lag-relationships-in-global-markets/)
- [The profitability of lead-lag arbitrage at high frequency (Poutré, Dionne, and Yergeau, International Journal of Forecasting, 2024)](https://doi.org/10.1016/j.ijforecast.2023.09.001?ref=aligrithm.com)

---

**A note on AI.* The ideas, research, analysis, and conclusions in this article are my own. I use AI tools to help with editing and wordsmithing, because English is not my first language, and I am not shy about that. AI-generated ideas and AI-assisted writing are not the same thing: the first is empty slop from a generic prompt, the second is a tool for communicating years of real research more clearly. Judge the work by its substance, not by whether software helped polish the prose.*