Skip to content

blueprint.json

The single state file, at docs/airgap/blueprint.json in the target project. An abbreviated example:

{
"schema_version": 1,
"project": "example",
"created": "2026-11-20",
"language": "tr",
"blueprint_path": "docs/superpowers/specs/2026-11-20-example-airgap-blueprint-design.md",
"answers": { "org.users_total": 180, "org.departments": 4 },
"assumptions": [
{ "id": "sizing.chat.active", "value": 0.2, "evidence": "estimate", "note": "default" }
],
"tier": {
"base": "T2",
"floors": { "llm-gateway": "T2", "identity": "T2" },
"reasons": ["org.audit_required raises llm-gateway to T2"]
},
"sizing": { "concurrency": { "low": 9, "high": 18 }, "total_gpus": 8, "servers": 1, "power_kw": 10.1, "fits": true },
"decisions": [
{
"layer": "llm-gateway",
"components": ["litellm"],
"existing": [],
"custom": [],
"rationale": "Four departments need separate keys and budgets; audit logging is mandatory.",
"alternatives_considered": ["kong"],
"evidence": "docs",
"override": null
}
]
}
  • components holds catalog IDs to deploy; it may be empty.
  • existing names infrastructure that is reused and not carried in.
  • custom names components the user insists on that are not in the catalog. They are labeled estimate and appear in the transfer manifest as “enumerate artifacts manually”.
  • benchmark is optional: catalog IDs that are carried in only to be measured against the decision’s choice and are never deployed, so they are not in components. references/layers/llm-serving.md says when an engine goes there. Lint reports a list that is not a list of strings, repeats an ID or repeats one of components as E1, an unknown ID as E2, and an ID that the verification section does not name as E9.
  • override is null or { "reason": "..." }.