Mechanical Muse

The muse proposes. The engine checks.

Mechanical Muse is an agent built on Meta’s Muse Spark. It turns a written sequence of operations into running building control logic, and that logic may run only once it passes its tests. Muse Spark drafts the logic as a block graph. A deterministic control engine replays test vectors against every draft, and failures go back to Spark for repair. Only a draft with a green verification receipt runs on an emulated building: BOPTEST’s large office, or its heat pump house on spot electricity prices.

Built solo by Justin Scott for the Meta Global AI Developer Hackathon (Indie track), October 26 to November 4, 2026.

How Mechanical Muse works: Muse Spark drafts, mm-graph emits CXF, mm-bridge replays test vectors on open-control-engine, failures go back for repair, and only a green receipt reaches the runtime, BOPTEST, the web app and a read-only BACnet mirror

Why it matters

Engineers write building control sequences as prose. A controls contractor programs them by hand for each building, and they are rarely tested against the written intent before they run. Language models can draft control logic, but nobody should let unchecked model output drive a building. Mechanical Muse puts a gate in between:

Model output is untrusted. Nothing Spark writes reaches the building or the BACnet mirror without a recorded green receipt: a content id plus “passes X of Y” test vectors, replayed on a deterministic engine.
Failures are specific. A refused draft comes back with the engine’s typed diagnostic code, or the scenario and instant where an output went wrong. Spark repairs against that, within a fixed budget.
Engineers stay in charge. Bring your own sequence as text, a PDF or a photo. Spark proposes test scenarios and point bindings; the engineer accepts or corrects them before they count.

Three authoring attempts: refused by the engine as an algebraic loop, then passing 8 of 10 scenarios, then 10 of 10 with a content id — Red, then green: a recorded Muse Spark session on the chilled-water plant reset.

The pipeline

Step	What happens	Who built it
Draft	Muse Spark (Meta Model API, structured output) reads a spec and drafts a compact block-graph JSON	Agent `mm-agent`: built in the event. Model: Meta’s Muse Spark
Emit	A catalog-backed validator and a deterministic emitter turn the graph into CXF (CDL as JSON-LD)	`mm-graph`: built in the event
Verify	The CXF is loaded with typed diagnostics and the spec’s test vectors are replayed	Bridge `mm-bridge`: built in the event. Engine: open-control-engine, used unmodified
Repair	Diagnostics and failing scenarios go back to Spark until the vectors pass or the budget runs out	`mm-agent`: built in the event
Run	A verified sequence runs in a supervisory loop against an emulated large office or a heat pump house	Runtime `mm-runtime`: built in the event. Emulator: BOPTEST, used unmodified
Mirror	Live points are served read-only as one BACnet/IP device	`mm-bacnet`: built in the event. Protocol stack: rusty-bacnet, used unmodified
Show	A React web app shows authoring, receipts, live trends, KPIs and explanations	`mm-server` and `web/`: built in the event

The Provenance page lists every pre-existing component with its pinned commit and license.

Getting startedRun the recorded demo; add a key and BOPTEST for live mode.

How it worksThe authoring loop, receipts, runtime fail-safe and the rest.

SequencesOur five control sequence specs and how each was validated.

ResultsBenchmark and BOPTEST numbers, with their caveats.