Aletheia is an open-source agent loop for uncertain investigations
Aletheia is an for questions where the answer cannot be checked in a simple way. Many work best when there is a clear test, such as code compiling, tests passing, or a task being visibly finished. Aletheia is built for cases where evidence is partial, messy, or conflicting, such as checking whether a vendor’s growth claim is believable, whether a company is financially healthy, or whether a science headline matches the actual study.
Its cycle is belief, act, observe, and update. It keeps track of what may be true, chooses searches that could change that view, and lowers when opposing evidence appears. Its first working use is an investigator for company and vendor diligence.
It returns a verdict with evidence, conflicting signals, stated , and unresolved unknowns. It can also stop without forcing a conclusion when the evidence is too weak. It currently runs with Claude Code and .
Key points
- Aletheia targets investigation tasks where the answer is hard to verify directly.
- Example uses include vendor claims, company health checks, and science headline checks.
- The follows belief, act, observe, and update.
- It can lower when evidence conflicts and can stop without a forced answer.
- It currently runs with Claude Code and .