Lab
Stretch Lab
One place for the Stretch 4 book, the lab dashboards, and the task ledger.
- Stretch 4The book: how the robot works, its hardware, firmware, Body API and code, in one table of contents.manual, wiki and codebase, in reading order14 chapters14 chapters
- DashboardsWhat is running right now, on which host, and the door into each dashboard.12 with a web dashboard, moving to sbl715 of 28 services up15 of 28 services up
- TasksThe evidence-backed ledger, the projects behind it, and every open task.16 active projects, ledger updated 2026-09-0317 done, 6 in progress, 4 blocked17 done, 6 in progress, 4 blocked
- Stretch 4.5The mid-cycle revision: one board order that removes this year's failure classes, and the firmware that rides with it, in two months.
- Stretch 5The next platform: vision and open questions.
How the work is done
Built and kept current by one engineer with local AI agents doing the reading, searching, drafting and bookkeeping. Every number on this page is read from the same files the site renders.
- All data stays local
LLM-POLICY.md - Claude Code and Codex run on the lab machines; nothing is pasted into a cloud chat window. Every project declares a privacy tier in its manifest.
- Schema before prose
validate:data - Program state lives in JSON files next to a JSON Schema. Pages are generated from manifests, and the validator refuses a project without a privacy field or two services on one port.
- An evidence-backed ledger
work-inventory.json - A done entry names the commit, file or measurement behind it; a pipeline entry names what blocks it. A claim without evidence does not get a row.
- Guardrails where the model is not trusted
AGENTS.md - Model-produced reference data is labelled and checked against the schematic before it drives hardware. Bench rigs park their relays on any host failure.
More on the workspace itself in About Stretch Lab.