Aletheia is an open-source uncertainty loop agent for Claude Code that uses belief-update over guess-and-summarize, delivering verdicts with explicit confidence and residual unknowns.
What Changed — The Specific Update
Aletheia is a new open-source agent built for investigations where the truth is hidden and the evidence is noisy. Instead of the typical "think → act → repeat" loop that guesses and summarizes, Aletheia runs a belief → act → observe → update loop — the shape of a POMDP (Partially Observable Markov Decision Process).
It's designed for Claude Code and OpenAI Codex. The core idea: treat every answer as a hidden truth, every search result as a noisy clue, and let contradictory evidence lower confidence rather than ignore it.
What It Means For You
Most AI "research" assistants run a few searches, then summarize whatever came back loudest. They sound most confident exactly when they're most wrong. Aletheia flips that.
Ask it something like "Is this vendor really at $10M ARR?" and it:
- Holds an explicit belief about what's likely true
- Spends each search where it will reduce its own uncertainty the most
- Lets contradicting evidence lower its confidence
- Stops only when the evidence has earned an answer — or says INCONCLUSIVE when it hasn't
You get back a Verdict: a bottom-line call, plain-English confidence for each claim, the evidence with sources, and the residual unknowns it couldn't resolve.
Try It Now
To install and run Aletheia with Claude Code:
git clone https://github.com/nsankar/Aletheia.git
cd Aletheia
pip install -r requirements.txt
# Configure your API keys
# Run with Claude Code:
claude code "use Aletheia to investigate whether Acme Corp is really at $10M ARR"
Key prompt pattern:
Use Aletheia's uncertainty loop to investigate [claim].
Return a verdict with confidence levels, evidence, and residual unknowns.
Three engineering choices make it work:
- Value of information search — Each next look is the one most likely to move the answer, at the least cost. Fewer searches, not more.
- Dual stopping conditions — A single lucky strong result clears the confidence bar but not the uncertainty bar, so the loop keeps looking rather than committing early.
- Honest INCONCLUSIVE — When evidence isn't there, it says so instead of hallucinating an answer.
When To Use It
Aletheia shines in investigations where:
- You need calibrated confidence (not just a summary)
- The truth is hidden and evidence is noisy
- You want to know what you don't know
Examples: vendor due diligence, competitive analysis, verifying claims, research synthesis.
The Bigger Picture
This is part of a broader trend in Claude Code ecosystem: moving from "guess and summarize" to structured reasoning with uncertainty. As Claude Code's terminal-native agent matures (Opus 4.8 scores 78.9% on Terminal-Bench 2.1), tools like Aletheia add a layer of epistemic rigor that's missing from default agent loops.
Compare with the recent "Build a Bulletproof Claude Code JSONL Parser" (Jul 5, 2026) — both focus on deterministic, verifiable outputs over probabilistic guesses. Aletheia extends that philosophy to research tasks.
Source: github.com
Originally published on gentic.news
Top comments (0)