Architecture

interlock route is one pure function of the rows and the request. It does not read the clock, a ledger, the watchdog, or a device. The harness is the other door. It runs a committed spec and writes receipts. Import turns those receipts into the table the route reads.

spec.yaml  ->  interlock run  ->  receipts/*.jsonl
                                      |
                                      v
                               interlock import
                                      |
                                      v
                              table/routing.jsonl
                                      |
                                      v
                               interlock route

The parts

devices is the allow-list. The names are NPU, iGPU, CPU, and 5090. GPU.1 is the 5090 as OpenVINO sees it on this rig. The resolver refuses that index even if the name or the vendor changes, refuses a bare GPU, refuses AUTO, HETERO, and MULTI, and refuses a GPU that is not vendor 0x8086 and INTEGRATED. A list with neither device property still requires the full name to contain “Intel”. A missing device is an error.

guards is the grant, the session, the heartbeat, and the empty card. A 5090 step without --grant does not start. An NPU or iGPU step without --npu-session does not start. A CPU-heavy leg waits for --cpu-window.

receipt is the schema and interlock check. importers rebuilds table/routing.jsonl from pinned receipts. interlock table prints that table. table/gaps.md is the experiment backlog.

The harness runs a committed spec. A rejected compile, a hang, and a lost device are recorded, and they are not speed. A watcher thread guards every run from outside the main thread.

router/policy.py is the choice. When the set of rows is empty, the command exits 1 and names the constraint. It does not guess.

Two adapters

The 5090 is reached through Ollama on the local machine. Cloud model tags are refused. E4b runs on its own Ollama, pinned to the install that is already listening for the 5090, and stops that process tree at the end of the block and on abort. The everyday servers are not opened by that harness.

OpenVINO runs, when a run is actually started, use NPU, CPU, or the Intel iGPU. They never select GPU.1. An embed, NLI, or rerank step on the iGPU refuses a missing static shape before compile, unless the spec sets dynamic_shape: true, and that opt-in is recorded on the row.

What a route is not

A route is not a server, a proxy, or a launch. A remote row can carry an offrig_plan object so a later tool could price a pod. interlock does not call that tool and does not spend. There is no runpod: row in the committed table. The remote shape is what the policy returns when a row of that shape is present. It is not a measured receipt.

The schema for a --json decision is src/interlock/schema/route-decision.v1.json.