Question 1

What is a fusion policy and how does it work?

Accepted Answer

A fusion policy is a program that intercepts every request between the agent and the LLM catalog and decides which models to call and in what order. It can pin a single model, escalate from a cheap model to a stronger one on certain signals, ask multiple models in parallel and pick the best answer, or any custom logic. The policy sees the full conversation stream (everything the agent has done so far) and can keep state across turns. It runs in a sidecar container next to the scenario, and every model call it makes goes through a metered endpoint that enforces cost caps and logs tokens.

Question 2

How is a policy evaluated and scored?

Accepted Answer

Each policy runs against 26 scenarios in isolated Docker containers. For each scenario, the agent produces a deliverable file (code, config, response, etc.), then a fresh verifier container runs the test suite and reports tests passed over total tests. The policy's score is the sum of tests passed across all scenarios, ranging from 0 to 26. Validators run independent evaluations and Bittensor consensus (Winsorized mean from SPEC 16 onward, weighted by stake) picks a winner; the policy must beat the incumbent by a noise-aware margin (3% in the normal range) to take the seat.

Question 3

What models are available and what do they cost?

Accepted Answer

Eight models are allowlisted on Engy: deepseek-v4-flash-0731, deepseek-v4.1-flash, qwen3.6-35b-a3b, qwen3.8-27b, glm-5.3-flash, glm-5.2, glm-5.3, and kimi-k3. Prices range from $0.04–$1.95 per million input tokens, depending on model. A $1 per-scenario cap is enforced—the meter stops accepting calls once the episode budget is exhausted. Cost is reported but not scored; quality and cost are separate, so you can optimize for either or trade them off.

Question 4

Can I test my policy before submitting?

Accepted Answer

Yes. You can run the exact same harness, meter, and verifier locally using the `eval_pack.py` script provided in the repo. Point it at your policy and a subset of scenarios, and you get per-scenario scores, cost breakdowns, and full transcripts. Local runs use your own Engy API key, so you pay for the evaluation yourself; validators run it on their keys.

Question 5

What prevents miners from hardcoding solutions or gaming the verifier?

Accepted Answer

Every submission passes through a server-side LLM-as-judge integrity check (Phase-1 pre-eval) that looks for hardcoded responses, instruction overrides, scenario-specific routing, and other red flags. Packs that fail are rejected before the challenger queue and never reach validators. Additionally, the policy runs in a sandboxed network with no internet and no filesystem access to the scenario's task files or tests. The verifier runs fresh after the agent completes, in a separate container, so the policy cannot observe or modify test execution. And every model call the policy makes is logged and cryptographically signed—provenance verification ensures the assistant message the agent received came from a model, not the policy itself.

Question 6

How long does evaluation take and how often can I submit?

Accepted Answer

A full session (all 26 scenarios) takes 10–30 minutes on a typical validator's hardware, depending on agent complexity and LLM latency. Submissions are cooldown-gated at 20 minutes per hotkey, so you can submit a new pack every 20 minutes. Each submission must be backed by an on-chain `recycle_alpha` burn (fee), which deters spam. Packs landing in the challenger queue are evaluated in the next available epoch (typically 20–30 min blocks).

Question 7

What is the seat and how does it change hands?

Accepted Answer

The seat is the policy currently earning 100% of miner alpha. A new challenger takes the seat only if it scores significantly better than the incumbent—from SPEC 16 onward, a noise-aware multiplicative margin δ(s) that is 3% in the normal range and decays to zero near the score ceiling. This margin resists evaluation variance and copy-paste attacks. Additionally, a seat's protection decays with age: after ~5 epochs of holding the seat, the margin softens and eventually becomes a −3% handicap, so a stale champion can be displaced by a slightly weaker challenger (anti-camping). The winner takes 100%; all other miners get 0%. When no seat exists (cold start or deregistration), emissions are burned.

TrajectoryRL

TrajectoryRL

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup