The second-revision letter came back with a request that sounded reasonable in isolation and impossible in combination: add more validation, and stay under 11,000 words. Reviewers wanted the fracture-and-bedding model shown to hold up on wells it had never touched, and they wanted evidence that it could handle the inclined geometry of horizontal wells, not just the vertical ones the study had been built on. The journal, meanwhile, would not move the ceiling. Every sentence of new evidence had to displace a sentence of something already there, or find somewhere else to live. That constraint turned the final revision into an editing problem as much as a modelling one: how much validation can you carry per word when the word count cannot grow?
Two demands that pull the wrong way
The paper covered a transformer that reads unwrapped borehole-image logs and returns the depth, dip, and azimuth of fractures and bedding planes across 14 vertical wells from a mid-sized Middle East carbonate operator. By the second revision the core results were settled. What reviewers questioned was reach. A model that scores well on held-out slices of its own training wells has not yet proven it generalises, and a model tuned on vertical boreholes has not yet proven it survives the geometry change when a well turns horizontal and the sinusoids on the image stretch and shift.
So the honest answer to both questions was more data on the page: blind predictions on wells the models had never seen, and a separate study on horizontal wells. Both are the right thing to add. Both cost words. And the 11,000-word cap meant we could not simply append them. The revision became an exercise in maximising evidence per word, which is the discipline the whole final round turned on.
Treating words as a budget
We stopped thinking of the manuscript as prose and started thinking of it as a fixed-capacity container with line items competing for space. The existing body was mostly non-negotiable. The two new pieces of validation were the additions we had to fit. And there was a third, cheaper item: a 17-item typo list from the reviewers that had to be absorbed but barely moved the count.
The two evidence blocks were not equal in weight. The first was blind-well predictions from both 11-well models, the combined beddings-plus-fractures model and the beddings-only model, run on wells outside their training set. That is the generalisation evidence, and it belongs in the body where a reader following the argument will see it. The second was the five-horizontal-well inclination study, which is more exploratory. On those five wells the model detected and localised sinusoids reasonably, with a blind split F1 near 63% at a detection threshold of 0.55 and a depth mean-absolute-error of 2.11 cm; parameter estimation was weaker on so small a sample, which is exactly the kind of honest, qualified result that reads fine as a supplementary figure and would bloat the main text if fully written out there.
That asymmetry is the whole move. Not every piece of evidence has to sit in the body. Supplementary material spends no words against the journal's main-text ceiling, so anything that a reader can consult rather than must read can be pushed there without losing it from the record. The craft is deciding which is which.
The instrument above is the revision as a ledger. The teal column stacks the words the manuscript spends in-body against the dashed 11,000-word line; the lever chooses how many of the five horizontal wells stay in the text versus become new supplementary figures. The orange marker is the point of the whole thing: if all five horizontal wells stayed in the body, the manuscript would breach the cap. Pushing figures to supplementary is what defeats that overflow, and it does so without dropping a single number from the paper.
Where each piece of evidence landed
The blind predictions from both 11-well models stayed in the body. They carry the generalisation claim, and a reviewer who asked "does it hold up out of sample?" needs the answer in front of them, not in an appendix.
The horizontal-well study mostly moved out. We kept a compact statement of what it showed in the main text and pushed the new figures to supplementary, where the F1-at-0.55 and the 2.11 cm depth error live alongside the caveat that parameter estimation on five wells is not something we would overclaim. Nothing was cut. The evidence is all there; it is just distributed so the main text carries the argument and the supplementary carries the supporting exhibits.
The typo list was the easy part. Seventeen corrections, absorbed inline, with a net effect on the word count small enough to ignore against the two large blocks.
The result was a manuscript that answered the reviewers' harder demand, showed both the generalisation evidence and the horizontal-well evidence, and still came in under 11,000 words. It reads as a single, coherent paper rather than a main text with a bolted-on annex, because the split was made by what a reader needs in the flow of the argument, not by what happened to overflow.
Why this is a repeatable move, not a lucky fit
The reflex under a word limit is to cut evidence. That is the wrong reflex when the evidence is what the reviewers asked for. The 14-well study and the well-count question behind it are their own story, told in How Many Wells Is Enough? A Well-Count Ablation for Fracture Detection; the point here is narrower and about editing, not modelling. A hard ceiling forces a decision every technical writer eventually faces: for each result, does the reader need it in the flow, or can they consult it on demand? Answer that per line item and the main-text budget mostly takes care of itself. The blind out-of-sample results earned their place in the body. The exploratory horizontal-well figures earned a place in supplementary. Both stayed in the paper. The cap held.
Limitations
The per-item word costs in the instrument are an illustrative allocation grounded in the second-revision letter, not a transcribed word count of the manuscript; we are not publishing the exact section-by-section word tallies of a confidential paper. The 11,000-word cap, the blind predictions from both 11-well models, the five-horizontal-well inclination study, the 14 vertical wells, the 17-item typo list, and the blind horizontal metrics (F1 near 63% at threshold 0.55 and depth MAE 2.11 cm) are the sourced facts. The horizontal-well result is deliberately reported as reasonable detection and localisation with weaker parameter estimation on a five-well sample, and should be read as exploratory validation rather than a headline claim. All well identifiers, formation names, and the operator identity are withheld under the engagement's confidentiality terms.
