Governance
How the registry is run, and the policies that govern submissions, moderation, and lifecycle. All policies are versioned in the public repository; changes are announced 14 days in advance.
How Agentarium is governed
Agentarium is a registry of scientific agents hosted by the University of Alabama in Huntsville (UAH) as community infrastructure. This page is the short, plain-language explanation. The linked policies are the long form.
What the registry verifies — and what it doesn't
The registry verifies format (does this submission have all the structured fields it needs?) and topic (is this actually a scientific agent in a recognized domain?). The registry does not verify scientific correctness, the accuracy of the author's claims, or the reliability of the tool endpoints the agent depends on.
Every author-stated field — intended use, known failures, validation metrics, guardrails — is shown to readers labeled as author-stated. You should judge it the way you'd judge a preprint: structured enough to read, but not peer-reviewed.
Who decides what
| Decision | Who decides |
|---|---|
| Does this submission conform to the schema? | Automated conformance gate |
| Is this submission on-topic? | Automated check; humans decide borderline cases |
| Can a first-time author publish? | One moderator endorses them (arXiv-style) |
| Is a tool's endpoint safe to register? | Human review for local-action tools; automated reachability for remote-query |
| Is a flag valid? | Moderator review within 7 days |
| Should an agent be withdrawn? | Author can always withdraw their own; moderators can force-withdraw with cause |
| Who can be a moderator? | Nominated by existing moderators, two-mod approval, conflicts disclosed |
Trust tiers
| Tier | Who | What they can do |
|---|---|---|
| 0 | Anyone | Browse, search, cite |
| 1 | ORCID-verified | Draft submissions |
| 2 | Endorsed author | Publish agents and tools; auto-approved when the gate passes |
| 3 | Moderator | Approve endorsements, review flags, process withdrawals |
| 4 | UAH admin | Full override, manages the moderator pool |
Service-level commitments
| Action | Target turnaround |
|---|---|
| First-time author endorsement | 72 hours |
| New tool registration review | 72 hours |
| Flag review | 7 days |
| Withdrawal processing | 24 hours |
How decay is handled
Tools break. Models go out of date. Authors move on. Agentarium tries to be honest about this rather than pretending it doesn't happen:
- A registered tool is pinged every 6 hours. After ~48 hours of failures, the tool is marked unreachable, and agents that depend on it flip to degraded with a public banner.
- Agents whose tested-on model is 4+ generations old are marked stale.
- Agents broken for 30+ days move from degraded to stale; at 90 days, a moderator reviews for forced withdrawal.
Withdrawn agents stay citable. The page resolves with a clear withdrawal notice and the reason. This matches academic norms for withdrawn preprints.
How to push back
- A submission was wrongly rejected: appeal in-app. The appeal auto-routes to a different moderator.
- A flag against your agent is wrong: respond in the moderation thread.
- A moderator misbehaved: contact UAH admins at the contact address in
/about. - Something in the schema is broken: open an issue on the registry repository.
The policies in detail
- Code of Conduct — community and moderator behavior.
- Endorsement Policy — what endorsers are vouching for, how to be endorsed.
- Moderation Policy — auto-approval matrix, SLAs, appeals.
- Withdrawal Policy — voluntary and forced withdrawal.
- Lifecycle Policy — broken-tool grace periods, model-staleness rules.
- Conflict of Interest — recusal rules for moderators.
- Moderator Handbook — the internal playbook, public for transparency.
Changes to these policies
Policy changes are proposed via pull request to the registry repository, reviewed by the moderator pool, and accepted by simple majority. Material changes are announced 14 days in advance.