The vug paper came back from its second revision with the kind of review that separates a defensible method from a fragile one. Not a demand to open the code, which is the fight two of our other manuscripts had already fought. This time the reviewers went at the method itself, one technical claim at a time, and the useful thing to record is which comments we pushed back on, which we conceded, and which we answered off the page. The paper is a classical computer-vision pipeline that counts and measures individual vugs on borehole-image logs from a mid-sized Middle East carbonate operator we partnered with. What follows is the concession map: four challenges, three verdicts, no contract broken.
The generalisation charge, and why the defaults were frozen
The editor's sharpest line was that the method looked over-tuned. The phrase, near enough, was that there was too much hyperparameter tuning for the result to generalise. It is a fair suspicion to bring to any pipeline with a dozen knobs, and a classical vug detector has them: a mode count, an intensity-separation step, a block size for adaptive thresholding, a circularity band, a centroid-matching tolerance, an overlap threshold. If each of those is retuned per well to make the numbers look good, the method is a curve fit dressed as an algorithm.
Our answer was to show the knobs were not retuned. The detector runs on frozen defaults across wells: k = 5 modes, an intensity separation of delta-m = 5, a 31 x 31 thresholding block stepped one pixel at a time across a roughly 380 x 360 image patch. Of the pipeline's parameters, six are universal and held fixed for every well - the mode count, the separation, the centroid-difference tolerance, the overlap threshold, the bounding-box expansion, and the mean-based radius - and only two are set per well, the thresholding block size and the local constant adaptive thresholding needs. A method that generalises is one where the parameters you would be tempted to tune are the ones you refused to move. Stating that plainly, with the count, was the defence.
The concession we did not fight
The comment that mattered most did not ask us to change anything. A reviewer asked whether the method distinguishes an interconnected vug system from a set of isolated vugs. It does not, and it cannot, and the honest thing was to say so without hedging.
The reason is physical, not a shortfall we could engineer away. A borehole-image log is a two-dimensional reading of the well wall - the electrical response of the rock at the borehole face, unrolled flat. Two vugs that touch behind the wall, in the rock the tool never sees, look exactly like two vugs that merely sit near each other on the imaged surface. Connectivity lives in the third dimension, and the sensor does not measure the third dimension. No area threshold, no shape filter, no circularity band recovers information the physics never captured. So we granted it: the pipeline measures vug morphology on the wall, and separating interconnected from isolated pore systems is out of its reach, flagged as future scope for a graph-based approach that would need volumetric input the image log cannot provide.
That single concession is the load-bearing one, and this instrument is built around it.
Naming a real limitation is not a weakness in a paper; it is the thing that makes a reader trust the claims you do make. A method that answers every question with a defence is a method a good reviewer stops believing. The interconnected-versus-isolated grant is what earns the frozen-defaults defence its credibility two paragraphs up.
The shape-distance suggestion, and keeping circularity
A different reviewer pushed on the false-positive filter. Our pipeline rejects image-log artefacts that are not real pores partly on shape: a real dissolution vug is roughly round, a drilling-induced streak is not, so a circularity measure separates them. The reviewer suggested we were reasoning in a circle by using a circle-based ratio and proposed a Frechet or Hausdorff distance between contours instead - established curve-similarity metrics that compare two shapes point by point.
We kept the area-ratio circularity, and the reason is narrow and honest. The area ratio of a contour to its enclosing circle falls on a bounded scale from zero to one, where one is a perfect disc, and that boundedness is what lets the same circularity band travel unchanged across wells. A Frechet or Hausdorff distance is unbounded in the general case and needs a reference contour to compare against; adopting it would have reintroduced exactly the per-well tuning the editor had just accused us of. We were not defending the prettier metric. We were defending the one that keeps the parameter count honest.
This is the connection between two comments that only shows up if you read the whole exchange at once: the shape-distance suggestion and the generalisation charge pull in opposite directions, and the area ratio is the choice that satisfies both.
The physics we disclosed, but not in print
The last request was the one the confidentiality agreement touched directly. To sanity-check the resistivity contrast the vug detector keys on, a reviewer asked for the connate-water resistivity and the mud salinity - the fluid properties that set how conductive the rock and the borehole mud are, and therefore how a vug appears against its matrix on the image.
These are real numbers from a producing field, and the field data belongs to the operator. Printing them in a journal would put a national oil company's fluid measurements on the public record, which the agreement forbids. So we did the thing the contract does allow: we disclosed them privately, to the reviewers, so the physics question got a complete answer while the manuscript carried none of the field values. The lever in the exhibit above reveals the payload we handed over - a formation-water salinity of 220,000 ppm, a mud-filtrate resistivity of 0.045 ohm-m, a mud-cake resistivity of 0.128 ohm-m - the same numbers that stayed out of print. A reviewer's need to check the physics and a contract's demand to keep field data confidential are not in conflict if you answer in the review correspondence rather than the manuscript.
The map, not the outcome
The vote went our way, and the fracture-and-bedding sibling paper from the same programme was accepted at a petroleum-engineering journal and is in copyediting. But the transferable thing is the map, not the result. Four challenges arrived; each got a different verdict. We defended the defaults because they were genuinely frozen and we could prove it. We granted the interconnected-versus-isolated limitation because the physics makes it unfixable, and granting it bought credibility for everything we defended. We kept the circularity ratio because the alternative would have broken the generalisation case. And we disclosed the salinity numbers to the people who needed them without putting the operator's field data in print.
Reviewing a confidential-data method well means deciding, comment by comment, which of those four moves each objection deserves. Fight the ones you can win on the merits. Concede the ones the physics has already decided. Disclose the rest where the contract lets you.
Limitations
This is one paper's second-revision exchange, not a general theory of reviewer response. The four challenges here had clean verdicts because each mapped to a real property of the method or the data; a comment that turned on the field identity, or one demanding the raw well data itself rather than a physics parameter, would have had no analogous move inside the same confidentiality boundary. The interconnected-versus-isolated concession is a genuine ceiling, not a temporary one: the graph-based future scope we named would require volumetric input a two-dimensional wall image cannot supply, so it is a direction, not a promise. And the account is anonymised throughout - the parameter counts, the salinity and resistivity values, and the method are reported, while the operator, the field, and the wells are held back under the agreement that shaped the whole exchange.
