From Hollow AI Draft to Rigorous Science: Our 5-Week Paper Rewrite | Karthik

Review Overview: Dissecting an AI-Generated Draft

[0:00]This is Liam in for Kumar Carik. You took a hollow AI generated literature review and turned it into something with real citations, honest gaps, and an actual argument. Walk me through what that rewrite actually looked like week by week. The first draft looked finished. It covered every method, cosmic ray removal, baseline correction, all of it. But every section only defined its topic. It never said why any of it mattered or what to actually do with it. Closing that gap between defining and explaining is what five weeks of rewriting was actually spent on. The original draft was AI generated and it showed sections read like a

The Guidelines Document: Defining the Three Forced Questions

[0:35]Glossery. Cosmic ray removal is this. Baseline correction is this with no thread connecting one idea to the next. Worse, several references either did not exist or did not say what the draft claimed they said. Before any science could be fixed, the structure itself needed a plan. So, the team built a guidelines document. not a style guide, a set of forced questions. Every section had to answer three things. What has actually been done? Why does it work or fail? And what do we recommend based on the evidence? A section that only defined a term without answering those three questions was not finished. The introduction went to the project manager first since it needed chemistry

The Two-Week Rhythm: Verifying Citations and Rewriting

[1:10]Background neither of us had. Then we layered in the machine learning framing ourselves. From there, the section split. My teammate took 2.1 through 2.3 and 3.1 through 3.2. I took 2.4. 4 through 2.6 and 3.3 through 3.5. Each section ran on the same twoe rhythm. Week one was spent hunting down every citation, checking whether a claim was actually backed by what its source said. Week two was rewriting using whatever week one turned up. That cycle ran twice, once per section across five weeks total, real meetings, real research hours logged the whole way through. 2.4 4 was written from scratch

Technical Gaps: PCA Data Leakage and Hotspot Variabilities

[1:46]With real citations for both alignment methods and an honest caveat that alignment itself can introduce error if the reference spectrum is chosen badly. 2.5 needed something stronger. The warning that PCA has to be fit only on training data, never the full data set before splitting. Skip that rule and your test set leaks backward into training. That section also previews a point the deep learning section leans on later. Automatic feature learning needs far more data than handbuilt features before it's actually worth using. Zers has its own specific headaches. Hotspots

Cleaning the Literature: Finding Mismatched Data Rows and Fake Statistics

[2:20]Vary randomly from measurement to measurement, usually treated as noise to average away. But one 20125 study trained on deliberately varied hotspot conditions instead of identical ones and cut prediction error by 84.8%. Separately, a crosslab study measuring the same samples on 35 instruments across 15 institutes found real systematic differences that had nothing to do with chemistry, just the hardware. Two very different problems, two real numbers, both added to backup claims that had none before. Assembling combining several models predictions sounds like a free win. It only actually is one when the models fail in different ways. One study combined four classifiers and beat every one of them individually, hitting 97.9%. But another combined just two models and did worse than either alone because the two classifiers kept disagreeing over noisy data instead of complenting each other. And two textbook models, decision trees and logistic regression, don't get their own sections

Here at all because in this literature, they only ever show up as pieces inside somebody else's ensemble. The original draft stated the gap as a specific number. Reported accuracy overestimates real deployment performance by 10 to 30 percentage points. Nobody could find where that number came from. It got removed, replaced with the honest version. This has never actually been measured. The claim that no study has reached the top validation tier got the same honesty treatment scoped explicitly to the actual databases and years searched instead of stated as a fact about the entire field. The summary table had its own quiet mistake. A random forest study sitting in the SERS row had never touched ramen spectroscopy at all. It was feckal source testing included only because it used the same algorithm. It came out a long open citation gap in the PLSR row finally got closed with a real SERS

Quantification study and one ensemble row that quietly mixed two very different approaches got split into two. So the distinction reads honestly instead of looking like a duplicate. So the draft started AI generated and hollow defining terms without ever explaining them. A guidelines document forced every section to answer what's been done, why it works or fails, and what to recommend. Two people split the sections, each running the same twoe cycle. Real citations replaced unsupported claims. A mismatched table row got caught and removed. And an honest admission replaced a number nobody could back up. That's what turning a draft into a real paper actually looks like. Your turn. Ask Claude to read through your own writing and flag every claim that doesn't have a real citation backing it the same way this rewrite did. Machine learning for sers from a draft to a real paper. Lay him in for Kumar Caric.