docs/standalone_layout.md

# Standalone Layout

Clean product one-liner: this repo should ship one standalone control stack where the kernel, policy, benchmarks, API, and operator UX all live locally.

## Shipped layout

| Slice | Local path | Upstream reference | Why in plain English | Real-world analog |
| --- | --- | --- | --- | --- |
| graph kernel | `vendor/meta3-graph-core/` | `meta3-graph-core` | durable graph state and receipts | engine block |
| control language | `policy/` | `nix-star` | bits, vectors, tensors, gates, confidence | instrument cluster |
| product runtime | `runtime/` | `causal-workbench` | route work, compile manifests, emit receipts | drivetrain |
| conversational API | `api/` | `NIX.codecli` | one front door over graph state | front desk |
| orchestrator UX | `ux/` | `NIX.codecli` | one operator surface over the API | cockpit |
| control benchmarks | `benchmarks/` | local eval doctrine + `NIX.codecli` patterns | prove the layers work | test track |

## What belongs where

| Concern | Home |
| --- | --- |
| bits / vectors / tensors definitions | `policy/` |
| control scorecards and eval runners | `benchmarks/` |
| memory, tensor, safety, economics reports | `benchmarks/` |
| request and response contracts | `api/` |
| operator flows and display modes | `ux/` |
| routing, manifest lifecycle, guarded execution | `runtime/` |
| receipt-first graph operations | `vendor/meta3-graph-core/` |

## Why this cut wins

Pareto read:

| Choice | Outcome |
| --- | --- |
| benchmarks as side files | loses coherence |
| API and UX left in another repo | loses standalone boundary |
| policy without a local kernel | loses execution truth |
| one local slice per concern | wins |

Layman version: a standalone machine needs its gauges, dashboard, and test track inside the garage, not parked down the street.