Escalane
Not generate-and-pray. Decide, then act.
The model chooses; the harness is the only thing that runs.
Composition
Industry Agent = LLM + tools
Escalane Escalane = Decide + Act
Same loop energy as an agent — closed action set. Expanded: System One + catalog harness. Agentspropose tools; Escalane selects among authored playbooks.
Two jobs. One boundary.
Language models are good at judgment. They are bad authors of production argv. Escalane keeps those roles apart — decide among a catalog, then act.
Decide
System One answers typed questions — playbook choice, safe-to-actnoul, blast score — over a frozen catalog plus escalate.
Act
The harness binds params from inventory, runs gates, renders argv from templates only, dry-runs by default, and writes an episode audit. The model never invents shell.
What you can prove
- Catalog — every mutate is a configured playbook (SysOps, Databricks CLI, homelab Docker/systemd).
- Gates — A0–A3 ordered vetoes; choice-only automation is not the product.
- Audit — catalog hash, state digest, answers, gate steps, argv in episode JSONL.
Research prototype (escalane). Seed eval numbers are claim-fenced — seeClaims.