03 — In Plain English
One start, then a simple loop.
There's no black box here. Every recommendation Wake gives comes from the same short, transparent path — one you can follow end to end, and trust.
First, you set Wake up — once:
Set the goal
You give Wake your app's goal — the outcome every decision should drive toward — and it breaks that into clear, ordered steps. If your app lets each customer set a goal of their own, Wake works toward that too, on top of your app's goal; where a customer sets none, your app's goal is enough on its own.
Every swimmer starts with one goal: move through the water more efficiently. She can add her own, like break 1:05 by September. The app sends it to Wake's goal compiler, shows her a preview of what it will weigh, and passes back any weight she changes.
In our robot-arm demonstration (a demonstration, not a result, and not hosted anywhere yet), an operator types a task, such as putting the mug in the cabinet, and it goes to the goal door for the arm's domain. Wake does not yet read the task's words for this domain, so the plan is not yet shaped by the goal.
Connect your data
You tell Wake which real-world signals to watch — the feeds and sensors it reads from your world. New readings stream in every turn, but you wire them up just once.
A swimmer uploads a stroke clip. V-JEPA, MoveNet, SAM 3D and Gemini read it, and each clip is matched to its closest swim-technique concepts, such as the elbow-up position at the catch. Today she sees Gemini's written read and coaching in the app's Language Model Intelligence tab. The World Model Intelligence tab, where the video model's own correction will appear, gets it from a backend change that is being merged now; until then that tab stays empty.
The demonstration's simulated camera looks at a household scene: a table, a cabinet with a door, a drawer, a mug, a bowl, a can, two blocks, a sponge and a cloth. No video reader trained for the arm is deployed yet.
Set the moves it can make
You tell Wake the moves it's allowed to make — the levers it can pull and the actions it can recommend. Wake plans within the moves you set — every move is declared up front. Strict enforcement, which rejects a planned move outside those lines, is on in staging and not yet in production.
Nine kinds of moves: a swim set, a form drill (such as the elbow-up catch), a strength block, a run, rest, a nutrition task, a mindfulness task, a goal review, and a form video review.
Twelve kinds of moves, from the 1,348 robot videos its world model trained on: pick, put or place, remove, open, close, move, push or pull, turn, pour, fold, stack and wipe.
Then this loop runs, again and again:
Take it in
Your app sends what's happening right now, and Wake turns the raw data into a tidy read of the moment.
A video of the swimmer's stroke comes in, and the video pipeline reads it into a picture of what her stroke is doing right now.
The demonstration sends one rendered camera frame to the perception door. That door answers 503 today.
Imagine what's next
Wake pictures what could happen if it went one way or another — several futures, not just the next word.
The app can ask the world model for her day-plan outlook, but that outlook has not yet returned a plan to a swimmer. What she gets today is a month of workouts built around the drills she adds.
The demonstration asks the world model for a plan at the level of moves, and the world model plans.
Weigh the options
It scores each option against the goal you set and what you care about — and shows the breakdown of why each option costs what it does.
Her compiled goal sets the weights Wake scores each option against; she can see and change them in the goal preview.
The demonstration sends no arm-specific weights yet, so nothing arm-specific shapes the score.
Act
Wake delivers a recommendation to your app from its planner. A quick-reflex mode for split-second calls is designed but not built yet.
She taps “Add to my month”, and the drill, such as the high-elbow catch, lands in her month of workouts.
The plan comes back at the level of moves, but no motor command exists yet, so the arm is not driven by it.
Learn
You report what really happened, and Wake learns from it for next time.
When she changes a weight in her goal preview, the app sends it back to Wake. A drill she accepts from the World Model tab will count as feedback once that tab is live; drills from the Language Model tab are scheduled but not counted.
The demonstration reports each cycle as not executed, because the arm did not move. It does not invent an outcome.