Bittensor 子网 11. 第二季（过渡季）上线：矿工们在聚变政策和项目上竞争 决定根据请求，工程师上哪些开权重模型能回答代理，以及以何种组合。SN11 的去向 一个是面向可验证环境的开放协议：docs/ROADMAP.md。

TrajectoryRL 在 Bittensor 上持续运行竞赛。验证者在真实沙盒中评估每条提交 面对带有隐藏测试的任务，共识权重是利害关系，最佳提交者将保留席位直到更好的 有人接受它。竞争产出的是可部署的：坐着聚变政策将作为其 自动模式（engy/auto），适用于任何客户和任何线束。

第二季的运作方式

• 你提交的内容：一个带有SKILL.md]和融合策略（policy.json，无代码，policy.py]）的资料包 小型 SDK，或原始的 OpenAI 兼容服务器）。该策略运行在场景容器旁边的侧车中，且 调用八个允许列出的模型中的任意一个，顺序或并行，重写它们看到的内容，保持会话状态， 并把答案直播回来。

TrajectoryRL

Bittensor Subnet 11. Season 2 (the transition season) is live: miners compete on fusion policies, programs that decide, per request, which open-weight models on Engy answer an agent and in what combination. Where SN11 goes next is an open protocol for verifiable environments: docs/ROADMAP.md.

TrajectoryRL runs a continuous competition on Bittensor. Validators evaluate every submission in real sandboxes against tasks with hidden tests, consensus is stake-weighted, and the best submission holds the seat until a better one takes it. What the competition produces is deployable: the seated fusion policy will be served on Engy as its auto mode (engy/auto), for any customer and any harness.

How Season 2 works

 miner                          validator (per scenario)                              Engy
 ┌────────────┐   web-submit    ┌─────────────────────────────────────────────┐
 │ SKILL.md   │ ─────────────►  │  scenario container: Hermes + the task      │
 │ policy.py  │                 │        │  OpenAI-compatible calls           │
 │ (or .json) │                 │        ▼                                    │
 └────────────┘                 │  policy sidecar: the miner's fusion policy  │  metered calls
                                │        │  only route: the meter             │ ─────────────►  8 open-weight models
                                │        ▼                                    │   receipts
                                │meter: allowlist, $0.30 cap, cost, provenance│ ◄─────────────
                                │        │                                    │
                                │  verifier container: hidden tests → score   │
                                └─────────────────────────────────────────────┘

• What you submit: a pack with your SKILL.md and a fusion policy (policy.json with no code, policy.py on a small SDK, or a raw OpenAI-compatible server). The policy runs in a sidecar next to the scenario container and can call any of the eight allowlisted models, in sequence or in parallel, rewrite what they see, keep session state, and stream the answer back.
• How it is scored: the sum over 20 scenarios of tests passed over tests total, from a fresh verifier container. Cost is metered on every model call at a frozen price table and shown per scenario and per model; it is not in the score. A $0.30 per-scenario safety cap bounds spending.
• Consensus: stake-weighted, Winsorized across validators; the seat changes hands only on a real margin.
• No GPU, no server, no uptime: the platform stores your pack; validators run everything.