Model cards
Agent model cards
A model card says, in plain words, what one of Runink TIDE’s agents does, what it reads, what it produces, what a person still decides, which model it runs on, and where it stops. There is one card for each agent you meet in the console, in a pull request or in an issue. Agents that work as a pair share a card.
Every card has the same nine parts: what it does, what it reads, what it produces, human oversight, model, where it runs and data handling, guardrails, limitations, and evaluation.
Four things true of every card
- Agents propose; people decide. No TIDE agent merges a pull request, deploys a change or sends anything outside your organisation by itself.
- Your own inference. Every agent calls the model plane of your own TIDE deployment, on your Server or in your cloud. No third-party AI service is called. See Models & inference.
- Base models today. Agents run on open base models. Runink domain adaptation is planned for each agent; this release uses the base models.
- No published scores yet. None of the cards carries an evaluation score, because none has been published.
The base models
| Tier | Model | Maker | Licence | Used by |
|---|---|---|---|---|
| coder | Qwen3-Coder-30B-A3B | Qwen | Apache-2.0 | Code Reviewer, Fix Drafter, CI Doctor (failed-build diagnosis), the FORGE coding step |
| general | Qwen3.6-35B-A3B | Qwen | Apache-2.0 | every other agent that calls a model |
The Independent Assessor also uses a small ranking model, Qwen3-Reranker-0.6B (Qwen, Apache-2.0), to narrow evidence. The Data Governance Checker uses no model at all.
The cards
| Card | Agent | Model | What a person decides |
|---|---|---|---|
| Code Reviewer | reviewer | coder | which findings matter and what merges |
| Fix Drafter | fixer, resolver | coder | whether to merge the draft correction |
| CI Doctor | self-heal, opsdoctor | coder (self-heal), general (opsdoctor) | whether to apply the proposed fix |
| Issue Triager | triage | general | the final labels and who works on the issue |
| Backlog Hunter | hunter | general | what to do about each reported item |
| Dependency Risk | risk | general | which upgrades to schedule |
| Compliance Evidence | compliance | general | how to close a gap; any claim of compliance |
| Release Curator | curator | general | whether to merge the curator’s pull request |
| Independent Assessor | judge | general, with a ranking model | what to do with each verdict |
| Rules Reconciler | recon | general, for one question | whether to fix the rule or the engine |
| Data Governance Checker | datagov | no model | owners, credentials and masking |
| Task Deployer | deployer | general | whether to merge the release-notes pull request |
| Ask TIDE | console Ask and chat | general | what to do with an answer; every provisioning action |
| FORGE (preview) | FORGE planner, forger | general (planner), coder (code) | every planned step, and every merge |
The roster names in the second column are the ones the Agent Fleet page uses.
Not given a card
evalis a test harness for the agents themselves, not an agent you meet. Its trace-scoring step uses the general model.usersapplies access grants from anadd-usersissue. It uses no model.sessionis thecore sessioncoding agent, which you drive yourself. See Coding Sessions. The Fix Drafter’s resolver and FORGE’s forger use it.