# Hexapod RL training orchestrator — live status

> Autonomous RL training campaign teaching an 18-servo hexapod robot to
> stand, walk, and turn (MuJoCo/MJX PPO on a GPU fleet). An LLM watcher
> launches runs, evaluates checkpoints, and writes verdicts; the
> documents below are its working state, mirrored as plain markdown for
> LLM readers assessing how the campaign is going.

Live state: watcher ON · 0 LLM analysis cycle(s) in flight · 0 queued in backlog · 0 run(s) marked RUNNING · 324 run(s) verdicted.

## Status documents

URLs are given as plain text (some LLM fetchers fail to follow
markdown-style links). No authentication is required on any /llm URL.

Campaign index — START HERE for "what is most promising", "how was
run X trained", "did it walk on the real robot": every run with one
canonical outcome, lineage, exported robot policies and the real-world
drive results joined to their training runs:
https://hexapod.cwd1f0-new-cluster.coreweave.app/llm/index.md

Research brief — a short human-readable answer to "what is the latest
research on each topic, and where are we?":
https://hexapod.cwd1f0-new-cluster.coreweave.app/llm/brief.md

Campaign + per-track STATUS — the campaign digest plus each research
track's current state; read this after the brief for the full detail:
https://hexapod.cwd1f0-new-cluster.coreweave.app/llm/status.md

Research plan — RL_PLAN.md, the plan the autonomous agents work from
(goals, phases, guardrails):
https://hexapod.cwd1f0-new-cluster.coreweave.app/llm/plan.md

Cycle log — RL_LOG.md, the append-only decision-cycle log (newest
entries at the end):
https://hexapod.cwd1f0-new-cluster.coreweave.app/llm/log.md

Run ledger — every launched training run with its hypothesis, status,
and verdict (one run in full: MCP run_story / `ops.sh index story`):
https://hexapod.cwd1f0-new-cluster.coreweave.app/llm/runs.md

All documentation — index of every other markdown doc in the tree
(hardware, sim, rewards, evals, per-run stories, …), each fetchable
at https://hexapod.cwd1f0-new-cluster.coreweave.app/llm/doc/<path>:
https://hexapod.cwd1f0-new-cluster.coreweave.app/llm/docs.md

## MCP server

The same results are queryable as tools over the MCP streamable-HTTP
transport (run ledger with filters, per-run stories, cached W&B
metrics, eval reports, doc search, and persistent per-run feedback) at
https://hexapod.cwd1f0-new-cluster.coreweave.app/mcp — but that endpoint is private: it requires the operator's
MCP key (Authorization: Bearer <key>, X-Api-Key, or ?key=<key>).
Agents can use submit_run_feedback and list_run_feedback; get_run also
returns all feedback attached to that run. The keyless /llm pages above
carry the same public campaign data but not the private feedback text.