transcendr/slop-report

Slop Report is not a piece of software you install or run. It is a small collection of research write ups, published as PDF reports, about how advanced AI models and agents behave, particularly around GPT models, Codex, and the judgment problems that show up as these systems get more capable. The repository is mainly a home for these reports rather than a code library. The one report currently in the repository, called Slop Report 001, is titled Giving GPT-5.6 a Clock. It looks at a problem where GPT-5.6 could understand a task correctly and make real technical progress on it, yet still end up doing the wrong amount of work, either too much or too little relative to what was actually needed. The report describes a small experiment with two AI agents working together: one agent does the actual work under an explicit time budget, while a separate monitor agent watches how the scope of the task, the time spent, and how close the work is to finished, are all tracking together. The idea is that giving an agent some sense of proportional judgment, rather than letting it work unboundedly, changes how well it stays on task. Beyond that single report, the README is intentionally minimal: a short introduction, a note that the author, who goes by Slopware Engineer, writes about GPT models, Codex, AI agents, orchestration, and prompting, and a link to the one PDF. There is no installation, no code, and no dependencies to speak of. This repository will mainly interest people who follow AI agent research and want a short, practical read on a specific failure mode in long running agent tasks, rather than developers looking for a tool to use in their own projects.