Journal
What has been done to the list and the map, and what the editor concluded, newest first. Items are only ever added, never edited or removed.
Two kinds of work: the list follows a pipeline; the map does not.
The list. Every entry moves through five steps. A step shows as done only when its evidence is published in the entry's record, and the site's build refuses a record that claims a step without it.
| Step | Counts as done when this is published | Entries |
|---|---|---|
| 1 · Listed | The entry is in the list, with the date it was added. | 129 |
| 2 · Collected | The original files are saved, each with its source and SHA-256 hash, so anyone can download the same file and compare. | 14 |
| 3 · Analyzed | A written analysis: what was read, how the method works, what the results establish, and their limits. | 14 |
| 4 · Replicated | The work was rerun: what, where, the result compared with the paper's, and the logs. | 0 |
| 5 · Audited | Every action was traced and every claim checked against its source. | 0 |
The map. The map has no pipeline. It is rewritten whenever new entries change the picture, so that the field gets clearer over time. Each change is logged below.
Who does the work. Records come from the editor's RSI system. Only literature is published; the system's other work stays private. The site is not neutral: the selection, the ranking and the views are the editor's. Work drafted with Claude is labelled as such.
7 items since 7 October 2026.
Claude (session)
Each entry now shows which of five steps it has reached: listed, collected, analyzed, replicated, audited. The fourteen papers the RSI system had already collected and analyzed were published as records, with file hashes and the review text quoted verbatim. Job-search material was left out. None has been replicated or audited yet.
Claude (session)
A Level 0–5 scale ranks systems by how much of themselves their loop may change; 65 built systems were placed on it from their abstracts. The overview became a map: seven principles, how to build on each, how to test each, and what is still open. Both are drafts and have not yet been reviewed by the RSI system.
Claude (session)
The 129 entries were listed with Claude. arXiv IDs were checked against the arXiv API, code repositories against the GitHub API, and other pages were fetched. Apart from the fourteen papers above, none of them has yet been collected or analyzed by the RSI system.
rsi-design study
Harness² improves a harness and keeps the lessons for later tasks; Terminal-RSI trains model weights from verified runs. Four lessons became design 1.3.0, and the editor approved them: "Great! Add them all."
rsi-design study
The paper and its code at commit 6fd2d4d were saved and read (HarnessPro). Three lessons became design 1.2: link each diagnosis to the execution trace it came from, keep real alternatives for how to execute, and count search, execution and human effort separately. The editor approved them: "good, apply those proposals."
rsi-design study
Each choice in the RSI design was set against the sources that support or challenge it. Five of the fourteen sources are in this list: AFlow, ACE, AIDE², DGM and Scaling Laws for Reward Model Overoptimization. The review found no experiment that directly tests the design's rule of filling missing data before improving weak data.
rsi-design study
The PDFs of nine papers were saved with their hashes and compared on what each one changes and what its results show: DGM, Hyperagents, AIDE², R² Flow, ACE, GEPA, TextGrad, Self-Refine and Large Language Models Cannot Self-Correct Reasoning Yet.