One contract, six surfaces, and a product that had to feel like a person made it.

The starting point was not a feature list. It was lived context: what two people had watched together, what they had watched apart, and which of those things carried something worth keeping. A generic catalogue answers none of that, because the interesting information is not the film. It is whose memory it is.

So the product model came first and the software came second. Shared memories, and distinct lenses for each person and for both together, kept separate from who is signed in. A seasonal calendar. Dates that matter. A companion character with a written voice, sarcastic and fond, rather than a general assistant wearing a name.

The artwork was chosen by hand and approved as final. No automated process was permitted to regenerate or replace an accepted master. That rule exists because a system that can improve your artwork can also quietly lose it, and the pieces in this product are not interchangeable.

The delivery history holds 202 commits and 990 tracked files at the Build 11 release candidate, commit 18bfcb9, and the first and last of those commits are 21 hours and 50 minutes apart. Those figures are pinned to that commit rather than to a date, because the repository is still moving and a count taken on a Tuesday is wrong by Wednesday.

That figure is the repository, and it is not the whole job. Concept, product model, artwork selection and the seasonal rules were settled before the first commit existed, and none of that appears in a git log. The total elapsed effort was longer and I have not measured it, so it is not stated here as though I had.

I was the product, creative, governance and acceptance authority. The concept, the relationship context, the artwork decisions, the seasonal calendar, the companion’s voice, the privacy boundaries and every correction came from me. An AI build lane did the implementation and integration under that direction, and a second lane did bounded review, audits and focused repairs.

The thing worth reporting is not that agents wrote code. It is what the arrangement had to look like before it produced something shippable. One lane owned the authoritative branch and the right to accept, reject or supersede anything before it entered the product. Everything else ran as bounded work with explicit files and acceptance criteria.

The finding, and it cost something to learn: the fastest arrangement is not the one with the most agents writing at once. It is one authoritative integration lane, strong human direction, and parallel agents given work narrow enough to verify independently. Open-ended competing builds create merge, context and worktree overhead that erases the parallelism that motivated them.

An Apple TV build reached TestFlight and was shown as waiting for review. That is Apple’s beta distribution system, and it is the furthest this product has travelled toward anyone outside it.

The current release candidate is synchronised across all five Apple targets in source, and is deliberately unsigned and not uploaded. Signing, upload, App Store platform records, processing, tester assignment and physical-device acceptance are separate authorised gates, and none of them is claimed here. Feature completeness runs ahead of release readiness for exactly that reason.

The backend follows the same rule. Local source completion is kept apart from production deployment authority, and only migrations with separate authorisation and a readback plan are activated.

SCOPE is the same direct-and-review arrangement on a single build, with the notarisation record attached.

Family Care Hub is the other private application on this site, built for my own family, and published as a synthetic case study for the same reason this one names nobody.