I am working on an ML-based problem for my undergrad thesis, based on an experimental dataset and prediction results published in a paper two years ago by an associate prof. in a very high ranking journal - the paper has 200+ citations. For months, I'd been trying to replicate the paper's results using the methods provided as well as those of my own but all in vain. The paper states that the code used would be made available on request, so out of curiosity I mailed the corresponding author for it.
I received a quick response, and for another month I was trying to decode how the author obtained their prediction results. One day when I was discussing implementation issues in my code with my supervisor, I spotted a grave error (or manipulation trick, as it seems to me) that completely invalidated all results in both their main paper and supplementary information. I was dumbstruck.
I found a very similar question on Academia SE - but here's a slight difference. I asked my supervisor to permit me to write an independent comment on the article as I tried to reach out to the authors regarding the mistake (or intended manipulation, as the case might have been), which they did not respond to. The authors deliberately avoided producing graphs that could demonstrate the error (as most papers in the field do contain those graphs) and did not define the exact performance evaluation metrics.
The problem is that I'm unable to convince my supervisor as he thinks that (1) the paper might be retracted or significantly altered as it provides both the code and the experimental dataset and (2) it would be an malpractice on ethical/humane grounds to request the authors' code and then use it to bring their published, widely-cited article down that also helped them gain grants.
I believe I should at least publish a comment or request the authors to collaborate to work on this together as answered in the linked question above, but I think the latter is unlikely given that a likely retraction would affect their grants.
What would be the best course of action in this case? How do I convince my supervisor to help preserve academic integrity and not let my months of efforts go to waste (as unfortunately, academia and thesis committees still undervalue reporting poor/negative results in a thesis)?
EDIT: I found it all out too late as I have less than a week for my thesis defence.
EDIT 2: I think I must clear a few more things. The paper was the first to present the dataset and results of its kind. The authors provide the data preprocessing code publicly but the model development code on request. Also, I think I can analogize the exact error in the paper with a simple example. Imagine you want to compute estimates of a quantity and require scaling it down in proportion to a reference. Instead of scaling the quantity down, you scale down the root mean square of the estimation errors between the true unscaled quantity and your unscaled estimate and exclaim you get spectacular estimation results!