Code Review With Spam Filter
Trusted public repos, thin studies, and spam patterns with a strict check before reuse.
How verification works
Public RSI code ranges from strong systems to thin wrappers. Many repos reuse the same keywords with little code, few tests, no benchmarks, and no history. The atlas applies a strict pass before trust.
A repo earns trust when it shows runnable code, a clear loop, an evaluator that is separate from the worker, logged runs or benchmarks, recent maintenance, and a license. A repo is marked thin when it shows only prompts, only slides, only a fork with no changes, or only keyword text with no execution path.
Trusted references
Darwin Godel Machine is the reference for open ended agent evolution with archive search and coding benchmarks. Godel Agent is the reference for self referential runtime with readable modules and task scripts.
Recursive improve by Kayba is the reference for trace driven product loops with capture, skill, benchmark, dashboard, and ratchet discipline. Karpathy autoresearch is the reference for minimal research loops with one editable file and a fixed budget.
The self improving research kernel is the reference for gated promotion with immutable guards and capability fixtures. Reflexion remains the classic baseline for reflection memory. Each trusted repo maps to one atlas role so readers can borrow mechanics without copying hype.
Thin studies and spam signs
Some repos hold a useful idea with a small community or early code. Program genetic RSI, three phase recursive agents, NFH separation loops, recursive selves, SEED evolution, curious creation, and small Godel inspired frameworks sit in this band. They merit study with tests before reuse.
Spam signs repeat with steady rhythm. Watch for copied abstracts with no code. Watch for chat wrappers with RSI in the title. Watch for forks with no diff beyond the readme. Watch for benchmark claims with no logs. Watch for self judges with no separate check. Watch for update bursts that only touch badges. The atlas keeps these out of the trusted path.
Repository verdicts at time of writing
| Repository | Signals | Verdict |
|---|---|---|
| jennyzzt dgm | Paper plus benchmarks plus archive search | Trusted reference |
| Arvid pku Godel Agent | Paper plus runnable modules | Trusted reference |
| kayba ai recursive improve | Traces plus benchmark plus dashboard | Trusted reference |
| karpathy autoresearch | Minimal loop plus wide forks | Trusted reference |
| self improving research kernel | Immutable guard plus full test gate | Trusted reference |
| reflexion baseline | Classic reflection memory | Trusted reference |
| program genetic RSI study | Pareto selection with docs | Useful study |
| recursive agents study | Three phase scaffold | Useful study |
| nfh loop study | Hard shell guards | Useful study |
| keyword only repos | No code or no runs | Spam pattern |