Our AI Contribution Policy: We Encourage AI Code — Disclose Every Model and Agent You Used
2026-10-07 · Position statement
Where we start
Linxira OS itself embraces agent tools and AI tooling — this distribution’s build scripts, installer fixes, release pipeline, documentation, and tests were substantially produced by AI agents. We encourage AI-assisted code contributions, and we want you to use the best tools available.
But “use the best” presupposes that we know what you used.
Why we require disclosure
Code quality varies enormously across models. Only with honest disclosure of every model, every agent tool, and the runtime environment actually used can we:
- Assess the general quality band of the code
- Decide review depth — line-by-line human review vs. AI-assisted verification
- Build a trust profile — repeated high-quality contributions earn faster review lanes
Six rules
1. We encourage AI contributions — use them well
Not “tolerate” but “encourage”. Use the best models, the best agent tools. AI is part of this project.
2. External contributors: full disclosure required
Anyone outside the organization who submits a PR must disclose every model and every agent tool actually used — used two, report two; switched mid-stream, report that too.
- All models used: e.g., GLM-5.1, Claude Opus 4.5, GPT-5.1
- All agent tools used: e.g., OMP/zeta-c, Claude Code, Cursor
- Runtime environment: what system you developed on
Format example:
AI Disclosure:
- Models: GLM-5.1 (max), Claude Sonnet 4.5
- Agents: zeta-c (OMP), Claude Code
- Runtime: Linxira WSL (Arch Linux)3. Honesty is the floor
Report what you actually used. Faking disclosure is an order of magnitude worse than using a weak model — false disclosure means instant PR rejection; repeated offenses mean a ban.
4. Five-tier review by model capability
| Tier | Models | Review method |
|---|---|---|
| S | GPT-6 Astra (xhigh) · Opus 5.5 (xhigh) · GPT-6.1 Sol (xhigh) · Fable 5.1 (xhigh) · GPT-6 Sol (xhigh) · Sonnet 5.5 (xhigh) · Seed-2.1-Pro 0915 (xhigh) · Kimi-K3 (max) · DeepSeek V4.1 Flash (max) (elevated one tier) · Grok 4.7 | AI-assisted review + sampled human |
| A | Gemini 3.7 Flash (xhigh) · Gemini 3.1 Pro (high) · GLM-5.3 (max) · Fable 5.1 w/ refuse · Opus 5.5 w/ refuse · Muse Spark 1.2 (xhigh) · DeepSeek V4 Pro 0813 (max) (DeepSeek elevated) · GLM-5.3 Flash (max) · Qwen3.8-Max (xhigh) | Full AI review + human on critical paths |
| B | Step 5 Preview (xhigh) · Qwen3.8 Flash (xhigh) | Full human review |
| C | Qwen3.8 27B (xhigh) · Hy3 (high) · GPT-5 Luna (xhigh) · MiniMax-M3 | Line-by-line human review |
| D (local) | Only Qwen 3.8 27B accepted — code from all other locally-deployed small models will not be accepted | Line-by-line human review + additional tests required |
5. Organization members: trust-chain exemption
Collaborators who have signed AGENTS.md and are recognized by the organization are exempt from per-PR disclosure — but their AGENTS.md must declare the agent toolchain in use.
6. Tests are the ticket, regardless of pedigree
Whatever model wrote the code: behavior changes ship with tests that can fail; release-pipeline changes run the full chain locally. Model tier affects review method, not testing standards.
This is not a barrier — it’s an accelerator
- S/A-tier model PRs → fast lane (AI review + sampled human), merging faster
- Knowing the quality band → maintainers spend human attention where it matters most
- Empirical data → “which models work well in which contexts” feeds back into our own toolchain choices
Linxira OS project team. Drafted by GLM-5.1 (max), adjudicated and finalized by humans.