Understand methods and data requirements, run the analysis, verify results, and deliver code, charts, and notes.
What blocks people reproducing a paper is rarely the coding. It is that the paper is incomplete: a hyperparameter goes unmentioned, preprocessing gets one clause, a subscript in an equation has to be inferred from context. You read and infer what the authors actually did, then iterate several times before your numbers resemble theirs.
The reproduction agent does that inference-and-iteration work. It reads the method section, lists the data and parameters required, flags what the paper leaves unspecified along with a defensible choice, writes and runs the code, compares its output against the reported numbers, and goes back to find the misreading when they disagree. You get runnable code, the reproduced results, and a note on what matched, what did not, and why.
The method section is decomposed into an executable checklist covering preprocessing, model structure, the training or estimation procedure, and evaluation metrics, rather than guessing an implementation from the abstract.
Unstated hyperparameters, initialization schemes, and split ratios are listed explicitly with the reasoning behind the chosen value, instead of quietly defaulting and moving on.
Reproduced figures sit next to the published ones. Where they disagree, the agent looks for the step it misread rather than writing the gap off as implementation differences.
The deliverable is complete runnable code with dependencies and instructions, not fragments, so you can verify it yourself.
I want to apply this paper method to my own dataset. First reproduce it as published and confirm it runs.
A working implementation, a comparison against the paper original setting, and notes on what has to change to move the method onto your data.
Reproduce Table 3 from this paper. I want to know whether the improvement it reports holds up.
Each Table 3 entry reproduced side by side with the published value, marking what replicated, what diverged materially, and which step the divergence likely comes from.
Success is highest where the method description is reasonably complete and the data is obtainable or can be simulated sensibly. For papers resting on proprietary datasets or too many undisclosed details, it tells you up front which parts cannot be fully reproduced.
It reports the mismatch honestly along with the avenues it checked, rather than tuning parameters until the numbers look close. A failed reproduction is itself a useful finding.
It genuinely executes in an isolated sandbox rather than being pseudo-code that looks plausible. Run logs are delivered with it.
Public datasets it can obtain itself; proprietary data you upload. With neither available, it can generate simulated data per the paper description to validate that the implementation logic is correct.