Skill

AI Research Reproduction

lllllllama

Research

Faithfully re-runs a published deep-learning repository to verify its documented results - reading the README first, picking the smallest honest inference or evaluation target, executing it under strict no-silent-edits patch rules, and logging evidence, deviations, and human sign-off points into a standardized outputs bundle. It is about REPRODUCING someone else's already-published numbers, not summarizing a paper, gathering fresh sources, or drawing business conclusions.

Situations it fits

You downloaded the freely available implementation of a famous image-classifier and need to independently check whether its headline accuracy claim was reached by running the lightest possible benchmark, documenting every deviation from the original setup, and signing off with your own numbers in a neatly packaged file before your committee reviews it.

AI Research Reproduction. This skill is specifically for faithfully re-running a published deep-learning repository to verify its documented results, including reading the README, selecting the smallest honest inference or evaluation target, executing under strict no-silent-edits patch rules, and logging evidence, deviations, and human sign-off into a standardized outputs bundle. It wins because the scenario demands an independent verification of a headline claims using the lightest possible benchmark while faithfully recording every deviation and sign-off, not organizing scattered files, compressing qualitative anecdotes, or sizing a commercial opportunity.

Your team suspects a 2022 generative-art paper inflated its numbers, so you fork the authors' public code, erase your own commit history to dodge accidental patches, and rerun the exact benchmark script to see if you can match their claims before the press release drops.

AI Research Reproduction. This skill is about re-running another team's published code under strict patch-free conditions to verify that their published numbers are honest. The distractor synthesize-research-anthropics is tempting because it also handles research, but it actually compresses scattered qualitative notes into themes and counts - it cannot execute a benchmark or verify numerical claims from a public code repo - so it is the wrong fit here.

3 scenarios in the bank answer to AI Research Reproduction. The rest are in the game.

Skills it gets confused with

These share a family with AI Research Reproduction, which is another way of saying they are the ones you might reach for by mistake.

Knowing what AI Research Reproduction does is the easy half. Telling it apart from the others under time pressure is the game.

Today's session

AI Research Reproduction is part of lllllllama/rigorpilot-skills. Licence: MIT. The description above was written for this game, not taken from the skill.