PerfectWorld

Panels 01 to 03 happen before the work, between the manager workspace and the focus contract. Panels 04 to 06 happen on the desktop. Panels 07 and 08 show where the record goes.

Six parts sit around one task contract. Context comes in from the tools people already use, the contract says what the work is, Focus Integrity checks behaviour against it, the senior asks when unsure, the answer becomes memory, and privacy limits what any of it may keep.

Each part offers the worker something, and each asks for something in return: agreement, an explanation or a little more of how they work.

The manager workspace mapped from the Week 8 build: how a vague brief becomes a project, a plan, assignments and procedures, and finally a focus contract handed to the worker. Each branch carries a need raised at Crit 2.

Highlighted pills trace the briefing loop. A brief is questioned, planned, delegated, written up as a procedure, approved and handed to a worker as a focus contract.

This side answers to the manager. It was the clearest demo at Crit 2, and the closest to Motion. The next spread maps the side that talks to the worker.

Motion. An AI project manager that schedules tasks, predicts deadlines and balances capacity. Calm and clear, but the worker appears only as capacity.

Taken forward. Calm, dense layouts and mono numbers from planning tools, and from monitoring tools a rule of what not to do: never score without saying why.

Vanta
An agent on every computer, a tick or a cross per device, reported upward. Monitoring as compliance.

7shifts
Each employee's lates, no-shows and dropped shifts, scored for the manager. The record speaks; the worker does not.

Linear
Issues, owners and status on a fast board. The task context the senior would import.

Slack
Where the work is actually discussed. Context the senior could read, and the reason connectors need limits.

After Crit 2 I stopped growing the planner and rebuilt the desktop around the worker. Each page exposes one part of the senior: Focus starts from a task, Timeline keeps what happened, Feedback shows what was learned, Privacy shows what is kept.

The Week 7 tracker hid its reasoning behind a score. Here every page is something the worker can open, read and argue with.

The desktop senior can read a screen, but design work also happens with paper, a soldering iron or a model, and to the desktop that looks like nothing. Ticker Cube gives the senior a body in the room: a Raspberry Pi 5 with a touch display, a CCTV-style lens, a microphone, a speaker and a visible privacy light.

The face is a question. A cube that smiles when you focus makes the watching friendlier, and harder to notice.

Crit 2 asked what the Cube adds if the desktop already watches. The answer is a different kind of work, not more watching, and the Cube waits until the desktop senior works. A body also makes the senior's presence harder to ignore.

Week 10 chose the hardware and wrote down what each part has to prove. Only the first two are done; sensing and interpreting studio work remain a proposal, and local processing alone would not settle consent or authority.

The tracker watched, guessed and warned. A senior has to understand the context, estimate how unsure it is, ask, accept a correction and adapt: alive in the sense of present and responsive, not conscious. It took three models to get there.

Four layers that each do what the others cannot: existing tools hold the work, the workspace turns a brief into a contract, the desktop checks the work against it, and the Cube carries that check into the room. Only summaries cross between them.

The senior is not one large prompt. It holds the current task, the session, recent corrections and what it is allowed to do, and most decisions never reach a model. Asking is an action it chooses, like warning or staying silent.

The captured case. Task match unclear, 54% uncertain, no memory to lean on: the senior chose ask_user.

Crit 2 made ethics the project rather than a closing section. In Aotearoa the Privacy Act 2020 governs information about staff, so each risk had to be answered in the build. If the system watches the worker, the worker must be able to see, question and correct it; that makes its authority visible, not fair.

I treated the Crit 2 deck in Week 9 as a test object: if people could explain the AI senior back to me, the direction was clear; if they read it as another productivity tool, the gap was still weak. Afterwards I compared ten routes, rebuilt the desktop senior in Week 10 and checked the build against its own claims.

After the crit I laid out ten routes before choosing one, and kept the parts that put the worker's answer before the system's judgement.