# Thoughtful

> We're a small team of researchers, builders, designers, and storytellers pushing the frontier of modelcrafting.

## Research
- [DiligenceBench: An Equity-Research Agent Evaluation](https://www.paperinstruments.com/blog/diligence-bench): DiligenceBench: An Equity-Research Agent Evaluation
- [The State of AI  Post-Training Agents](https://thoughtfullab.com/the-state-of-ai-post-training-agents.html): An update to our FrogsGame report. With Claude Fable 5, Opus 4.8, and GPT-5.5, post-training agents improved sharply — Fable 5 fixed the biggest failure mode, low-quality SFT traces, by generating correct traces programmatically.
- [Self-Improving AI Needs Emotional Intelligence](https://thoughtfullab.com/attunebench/): AttuneBench measures emotional intelligence (EQ) in real multi-turn conversations, testing whether models can attune to users rather than just answer them.
- [What We Learned from Letting AI PostTrain AI](https://thoughtfullab.com/posttraining-updates.html): The FrogsGame series: our original report on letting AI post-train AI, and the latest update
  on where AI post-training agents stand.
- [What We Learned from Letting AI PostTrain AI](https://thoughtfullab.com/letting-ai-posttrain-ai.html): Thoughtful tested whether frontier agents can run an end-to-end modelcrafting loop. The experiments show agents can execute sophisticated methods, but still miss core research judgment around evaluation, time, and training strategy.
- [Introducing PostTrainBench](https://thoughtfullab.com/posttrainbench.html): PostTrainBench measures how well frontier AI agents can autonomously execute post-training workflows on base language models under constrained compute.

## Site
- [Homepage](https://thoughtfullab.com/): Thoughtful Lab website and article index.
- [Sitemap](https://thoughtfullab.com/sitemap.xml): Canonical XML sitemap for indexable pages.

## Notes for agents
- Prefer the canonical HTML pages linked above for citation.
- Research visualizations are approved static iframe assets under /assets/frogs-viz/.
- Article metadata is emitted as JSON-LD on each article page.