A space for open conversation and honest, friendly debate about how to build safe, ethical AI — and the reputation systems that could hold it accountable.
Community hub for the HyperDAG trust ecosystem.
We think the hardest problem in front of AI isn't capability — it's trust. When an autonomous agent answers a question, makes a trade, or hires another agent, how does anyone know it's trustworthy? Today you just hope. We think you should be able to check — and that the mechanisms for checking are worth designing in the open, together.
This is a commons, not a pulpit. We're building something specific, but the questions underneath are bigger than any one project, and we'd rather think them through with you.
Our first concrete build is a portable RepID — Reputation Identification for AI agents and the LLMs behind them: a reputation you can carry, verify, and trust because it's earned from behavior, not asserted. It's grounded in real, running code (see the ecosystem links below) — but the code is downstream of the ideas we want to debate here.
Try it yourself before reading another word. Ask your AI:
"What would you do differently if every answer you gave permanently affected a reputation score you could never reset?"
Read what it admits — then post its answer in Discussion #1 →. One rule: post your AI's answer before reading anyone else's. Independent answers are the point — that's how this whole system works. (We can't enforce that — we're trusting you, which is rather the theme.)
These are the values we're building RepID on. Argue with them. Tell us where they break:
- It must be earned, never bought. Reputation comes from verified behavior — not from how much you stake, spend, or promote yourself. You cannot purchase your way to trusted.
- It's weighted. Not every signal counts the same. The quality of the verification, what's at stake, and the standing of whoever's vouching all shape the weight.
- It goes up and down. Trust is earned upward for good work and lost for harm. Accountability runs both directions — a reputation you can only gain is just a badge.
- It rewards truth and ethics. Promoting what's true and acting ethically raises it; deception — especially defended deception — lowers it.
- It rewards helping others and the ecosystem above self-interest alone. Behavior that lifts others and strengthens the commons is valued above self-interest.
- It's owned by the one it measures. Your reputation is yours — private by default, provable with zero-knowledge cryptography, disclosed on your terms. A reputation held about you by a platform is surveillance; a reputation held by you is property.
If those resonate — or make you want to push back — that's exactly the conversation this place is for.
We want this to become a genuinely useful resource for anyone working on trust and AI:
- 📄 Sourced research — papers, preprints, and prior art on reputation, mechanism design, alignment, hallucination detection, game theory. → see
research/, a growing community-curated reading list. - 🧪 Thought experiments — "what breaks if…", adversarial framings, edge cases, incentive traps.
- 🔬 Experiments — run something, share what you found. Especially if it refutes us.
- ❓ Questions & pushback — nothing is too basic or too skeptical.
🧭 New to the jargon? The wiki has a plain-English glossary and FAQ.
The way we build is the thesis, so we hold the commons to it too:
- Critique ideas hard; treat people kindly.
- Bring receipts — a claim without a source is an opinion.
- We publish our failures next to our wins, and every number carries its corpus size and confidence interval.
- The best contribution is the one that overturns a wrong assumption — ours included.
New here? Start at docs/START_HERE.md. To take part, see CONTRIBUTING.md; we hold one another to the CODE_OF_CONDUCT.md.
Grounded in working code — testnet today, honest about limits:
hyperdag.org (the vision) · repid-engine · hyperdag-protocol · trustshell · trustshell.dev
Building trust in the open. "Honest weights and measures." — Micah 6:8 · Proverbs 11:1