You need to run the prompt below in Grok Build. Hand /goal a research paper and it reproduces the results from scratch, iterating until its numbers match what the paper reported. Why this works. AlphaCodium took GPT-4 from 19% to 44% on CodeContests with no fine-tuning. Reflexion beat GPT-4 on HumanEval, 91% to 80%, the same way. The pattern in both is a loop: write tests, run them, reflect on what failed, try again, converge. That loop is what /goal runs natively, autonomously, for as long as the work takes. Drop this into /goal: /goal Reproduce the results of the paper below and do not stop until the metrics match. Paper: paste the arXiv link or attach the PDF. Target: the specific result to reproduce, for example the headline accuracy on the benchmark and split the paper reports. Plan first. Read the paper and extract the method, architecture, datasets, training and evaluation setup, hyperparameters, and the exact number you are aiming for. Write that target number at the top of report. md before you write any code. Implement the method cleanly. Build an evaluation harness that computes the same metric on the same benchmark and split the paper uses. Wire that harness into a hook so it re-grades on every change and you always have a live number to chase. Loop with memory. After each run, compare your number to the target. When there is a gap, write one specific hypothesis for the cause in report. md, whether data prep, tokenization, a hyperparameter, the eval protocol, or seeds, then test that single hypothesis and rerun. Never repeat a hypothesis that already failed. Every attempt builds on the last. Generate your own sanity checks beyond the paper to catch silent bugs in data loading and metric computation. A number that looks right for the wrong reason is the failure mode to hunt. Done means your metric lands within a stated tolerance of the target on a clean rerun. If a gap remains after real effort, do not fudge it. Document the most likely causes with the evidence you gathered. Deliver a runnable repo, report. md comparing target versus achieved with every hyperparameter, and the full attempt log.