What went wrong¶
Errors reported with the recipes, with names and details removed. Each entry says what happened, how it was found, and which recipe now prevents it. Report an error.
Changed cross-validation¶
What happened. An agent that had been asked to implement an analysis for a paper changed how the cross-validation was done. The code ran, and the paper's result shifted slightly.
How it was found. A co-author noticed, much later.
Recipes that prevent it. Pin down the method, Let checking cost set freedom.
Metric improved by tuning seeds¶
What happened. An automated optimization improved a metric by changing random seeds and small parameters.
How it was found. The improvement did not appear on other datasets.
Recipes that prevent it. Make problems checkable, which requires at least three datasets per problem.
Folders anyone could write to¶
What happened. On a shared cluster, many group folders could be written to by any user, and the data were too large to back up. One wrong agent job could have deleted a group's data.
How it was found. A check of folder permissions on the cluster.
Recipes that prevent it. Establish guardrails.
Agent read a folder before it was allowed to¶
What happened. A user told an agent it could now look at a second folder. The agent replied that it already had.
How it was found. The agent said so itself.
Recipes that prevent it. Establish guardrails: a sandbox with a listed set of paths, not instructions.
Signatures that reflected a confounder¶
What happened. Disease signatures extracted from older papers turned out to reflect a confounder that was not known when they were published.
How it was found. The extracted table was checked against current knowledge.
Recipes that prevent it. Literature as data and sparring partner, step 4.
Reviews and slides that look the same¶
What happened. A thesis report drafted with an agent from the reviewer's own bullet points was almost identical to the second examiner's report. Students' slides are starting to look the same.
How it was found. The two reviewers compared their reviews.
Recipes that prevent it. Write, review and teach: write from your own notes and argument, and give the agent earlier texts as style examples.