← the whole review · The mechanism · The twelve shelves · The names · The labels not on your list · Retail & Commerce · Candidates from your old lists · Brand & Identity

The open calls

Six calls on the classification model that only you can make

One line per decision, in any order: ID: CONFIRM · ID: CHANGE: … · ID: LEAVE OPEN: …. Every number here was read from the database on 2026-08-25.

A-05The two questions about the product that turned out to be one · full

A beat is one moment inside a video, and every described beat answers two questions about the product — how visible it is, and what part it plays. Do we merge them into one?

When the machine watches a film it does not only say what the film is; it walks through it beat by beat — one moment at a time, the two seconds where a hand picks up a phone. On each beat it answers a short list of questions, and two of them are about the product: how visible is it, and what part does it play?

Those two were written down as if they were different questions, and only three of the visibility answers ever really were: shown on screen (965 beats), on a device (351), in hand (87). The other five — absent (1,064), implied (311), contextual (122), in active use (73) and focal (43) — are the other question's own words, and the machine has been giving them as visibility answers anyway. That is 1,613 beats answering a question they were not asked, against 1,403 answering the one they were.

An earlier version of this decision put the drift at 238 beats against 438, and it was wrong. That reading counted only three of the crossed words and left shown-on-screen out of the comparison — a word the recommendation itself keeps — so both halves were understated. Re-read on 24 August it is five of eight, 1,613 against 1,403: a stronger case for merging than the card first made.

The two Sora pieces below are the whole argument at reading size. In the holiday release demo the interface is on screen and the product is also what the beat is about — one answer covers both readings. In the scuba-diver clip there is no product on screen at all, so today the machine answers “how visible is it” with the word absent — a word that only ever belonged to the other question. Merging is not a new idea being imposed on the data; it is writing down what the data has already been doing for months.

My take: Merge them. Two questions that share five of their eight answers are one question wearing two hats, and the only thing keeping them apart is that somebody once wrote them down as two.

What this call does to everything else written on these pieces

This reaches no further than the beat. A label is the one word saying what kind of thing a piece is. What is written on the whole piece — its one label, where it ran, who appears in it, which campaign it belongs to — is not touched, and neither are the crafts credited on it: the Codex sales film above still reads Logo Design, UI Design and Motion Graphics whichever way you rule. Inside the beat, the merged question sits beside three others — what is literally on screen, what the beat does for the story, and how it was filmed — and it shares no word with any of them, so nothing else has to move to make room. The only surface a person would notice is the filter list you browse beats by: one line disappears and three answers join another. And it makes one repair possible that is impossible today, because once the words live on one question the machine can no longer write a role word where a visibility word belongs.

The alternative, and its cost: Keep the two questions and write one sentence each saying what only that one answers. The cost is that the sentence has to hold a line the machine has already crossed 1,613 times without being told to — so you would be buying a check that catches the drift after it happens, every time, instead of a shape that cannot drift.

Where this was worked out: the review agenda, item 5 — raised while reading the live system on 11 August, with the full distribution re-read on 24 August

Your own example this touches: IBM at the US Open

A-06The three things we write on every beat that no list governs · full

Three things get written on every described beat with no agreed list of words behind them — do we govern two of them and harvest the third?

Beside the big questions, every described beat carries a handful of small ones: is this beat interesting because of the craft, the data, or something the product does; is anything moving; what does the room look like. Three of those small reads have no agreed list of words at all — nothing says which words are allowed, nothing defines them, and nothing rejects a new one on the way in.

Left to itself the machine still produced something that looks like real vocabulary: 17 interest words — brand lore, category signal, craft detail, data proof, data visualisation, ecosystem, identity update, interaction detail, kinetic type, lifestyle, lifestyle detail, micro detail, new release, none, product behavior, interface detail, visual motif — and 5 motion words. That is genuinely the good news: an open list is how a technique nobody has named yet gets a name.

Here is the bad news, and it is one word. Kinetic type is written as an interest word on 8 beats and as a motion word on 2 others — the same word answering two different questions, which is exactly the drift a written list exists to prevent and exactly the kind of thing nobody catches by reading. It is small today because the numbers are small. It is not small in a library ten times this size.

The look of the scene is a different animal: 550 free-text phrases, of which the eight most common are technical/UI (461), clean UI (247), modern office (166), casual office (108), minimal clean (82), minimal UI (72), office interview (69), minimal studio (57). Read that list and the clustering is obvious — offices, interfaces, minimalism — which is the argument for not governing it yet. It goes through the harvest-then-promote loop you already ruled for beat details — let the words come in freely, then promote the ones that keep coming back: let the words come in freely, and on a schedule promote the ones that keep coming back, with definitions.

The Commonwealth Bank case-study film below is where all three land at once. It carries 31 described beats, and every single one of them got an interest word, a motion word and a phrase for the look written on it by a machine with no list to check any of them against. Nothing about that film is wrong today; it is simply unverifiable.

My take: Govern the first two and harvest the third. The word sitting on two lists at once is not a tidiness complaint — it is a defect that keeps being written, every day nobody rules on it.

What this call does to everything else written on these pieces

These three reads sit at the beat level and today nothing hangs off them: no label on a whole piece is worked out from them, no shelf is built on them (a shelf is simply a group you browse by), and no campaign, product or company fact depends on them. So governing them breaks nothing — what it changes is what a person can do with them. An interest word becomes something you can browse and filter by rather than a word that happens to be sitting on the beat, and the two readings of kinetic type can be told apart on purpose instead of by accident. The phrases for the look of the scene stay exactly where they are until a promotion round moves them, so nothing on the Sora pieces above changes on the day you confirm.

The alternative, and its cost: Stop writing the three reads altogether. It is honest and it is cheap, and it throws away the only evidence we have of what a beat is interesting for — including the 550 phrases that are the raw material for a shelf of scene types nobody has built yet.

Where this was worked out: the review agenda, item 6 — raised on 11 August, with every word re-read on 18 August and again today

Your own example this touches: IBM at the US Open

A-07Where a product lives — one thing, two homes · short

A product lives in two connected places at once — do you confirm that the smaller of the two is the detail layer beneath the other?

A product shows up twice in the live system: 394 things inside their companies have been marked as product lines — Sora and Codex among them, the two products named on the pieces below. A separate list holds 139 products, every single one of them tied back to the company that owns it, and each able to carry its own identity, its parent product and its launch date.

Nothing is lost between the two and nothing is duplicated, so the natural reading is that the smaller list is the detail layer beneath the larger one.

The one thing that reading does not settle is which of the two a person browsing should meet first.

My take: Confirm it, and lead with the product lines inside the company. This item is not a defect, so the only real choice left is which door a person walks through.

Where this was worked out: the review agenda, item 7 — raised while reading the live system on 11 August, re-read on 17 August and again today

Your own example this touches: IBM at the US Open · McDonald's and Minecraft

A-08The middle level of the services tree · short

The middle level of the services tree still has no name — do you confirm your own July word for it, Discipline?

The outer two levels are already yours: the top is a Service — what a client actually buys, like creative production or brand strategy — and the bottom is a craft, which is the only finished vocabulary this system holds. Every one of the 73 crafts in use carries a full written rulebook saying when to tag it and when never to.

The last round printed that rulebook figure as 95 rather than 73, and both readings are true: 95 is the same tree counted whole — every live entry at all three levels, each one carrying its rulebook as well. Nothing in that tree has been created or edited since 21 July, so what changed between the two rounds is the set being counted, not the data underneath it.

What has never been named is the level in between — animation, copywriting, design, photography and post-production underneath creative production, which is exactly the level the contracts film below is credited at (Cinematography, Graphic Design and UI Design). Your own July frame already calls that level Discipline, with Specialization and Execution below it. Group and family are the alternatives if you want a different one — family names are the groups the database files labels under.

My take: Confirm Discipline. It is your word, it is the only one of the three that does not already mean something else somewhere in this system, and the crafts sitting underneath it are the standard every other list here gets measured against.

Where this was worked out: the review agenda, item 8 — your July correction pass, checked against the live services tree

Your own example this touches: The vendor's case-study page

A-09Crew roles at the level of a single beat · short

Do we keep a vocabulary of crew roles at the level of one beat — who operated, who lit, who graded that moment inside the video — or park the question deliberately?

The craft ruling in August settled which parents a craft may have and left exactly one thread hanging: whether the beat level gets its own vocabulary of crew roles.

Nothing is stored at that level today, so this is a question about what we start collecting rather than about repairing anything — the Commonwealth Bank case-study video below carries 31 described beats and not one of them names who shot it.

Its neighbour one level up is unaffected either way: the 65 credit roles written on a whole piece stay as they are, and parking is not losing, because a parked question comes back when the moments work needs it.

My take: I have no lean to offer here — this is a question about how deep you want the moments work to go, and dressing a guess as a recommendation would be worse than saying so. If you park it, park it with a date on it.

Where this was worked out: the review agenda, item 9 — the craft question left open at the August walk

Your own example this touches: IBM at the US Open · The vendor's case-study page

A-10The definitions queue · full

Most of the words this system can write have never been defined in writing — do the drafted definitions stand as the start of a standing queue?

Every word this system can write on a piece or on a beat lives in one place, and that place holds 973 words in use. 379 of them say what they mean. The rest are a word and nothing else — which is perfectly workable while one person classifies from memory, and stops being workable the moment a second reader, human or machine, has to decide whether something is a Documentary or simply a film shot in a documentary style.

Reading the craft rulebooks you said “some of these are still a little off”, and the reference page built afterwards found a wider gap: the reference page cited 370 defined of 973 when it was written on 11 August, and the same set of words read again today gives 379 of 973. Nothing in it has been created since 11 August and nothing edited since 21 July, so the difference is a different way of counting, not a change in the words themselves. Take either figure and the conclusion is the same — most of the vocabulary has never been written down.

An earlier version of this card counted 21 words carrying an engineer's note where a definition belongs. Read again today it is 12. Under the loosest test, a bracket anywhere in the entry rather than only at the start, it is still 12: the six families of beat — a different list from the family names elsewhere in this review; these group moments, not labels — and the six words for how a piece speaks. Nothing in that set has been created since 11 August and nothing has been edited since 21 July, so the 21 was a wider hand-count rather than a figure the data has moved away from.

All 73 crafts in use carry a real rulebook — which is why both podcast pieces below can be credited Brand Copywriting without anybody needing to argue about it. The last round printed 95 of 95 here instead; that is the whole services tree counted at every level rather than the craft level alone, it is equally complete, and nothing in that tree has been created or edited since 21 July. Nothing else in the system has that, and everything else in this review quietly assumes it does.

The two podcast pieces below are the whole problem at reading size. They are two episodes of the same show, five days apart, with the same two guests — the researcher's surname is even spelled two different ways across them, Hata and Hara. But the machine did not read them as the same thing at all: one is credited to ChatGPT and Microsoft and is about what people are making with images, the other to DALL-E and about how the text inside those images got good. The one thing they genuinely share is the word written on them — podcast video, which 56 pieces in the library carry. So if a third person had to decide whether either of them is really an Interview instead, there is not one sentence written anywhere in this system that would help them, and knowing what each episode is about is no help either, because that is a difference of subject and the question is a difference of kind. The outside review named six of these boundary terms — Documentary, Interview, Testimonial, Product Demo, Music Video and User-Generated — and each one is a kind of piece only when that structure is the spine of the whole thing; otherwise it is a style, a job the piece does, or a single beat inside it.

The reference page already carries a batch of drafted definitions, every one of them marked as a draft.

My take: Let them stand. The undefined words are the quiet reason this review keeps re-arguing the same three boundaries, and a queue that moves a batch at a time will beat the perfect pass nobody ever starts.

What this call does to everything else written on these pieces

A definition is not a fact about a piece, so confirming this moves nothing in the library — and yet it is the item everything else in this review leans on. The instructions the machine is given are meant to be generated from this same set of words, so a word with no definition is a word the machine guesses at every time it reads a film. The shelves you browse by are built from these words, so an undefined word is a shelf nobody can explain to a new person. The six boundary terms are precisely the pairs that put a piece on the wrong shelf in the first place.

The alternative, and its cost: Write nothing and lean on the machine's own consistency. The cost is that consistency is all you ever get: it will keep calling the same thing by the same word, and nobody will be able to say whether that word was the right one.

Where this was worked out: the review agenda, item 10 — the vocabulary reference, measured on 11 August and re-read today

Answer alongside: N-11 · N-08 — Your own example this touches: The model card on a launch tweet