GetEYT feature adoption matrix

52 scored dimensions across 29 products, from 5,817 lines of teardown research and 434 screenshots. Machine readable version: FEATURE-MATRIX.json.

Executive summary, the ten numbers that change a decision

0 of 4 booking platforms print a capacity denominator. The best any member can learn anywhere in the category is 3 SPOTS LEFT, which in a Pilates room cannot distinguish a six machine studio from a twelve machine one.
0 of 4 ever show a waitlist position. Not one product tells a member whether they are first or eleventh.
0 of 4 achieve all three of: see fullness free, see price free, book without an account. Punchpass has the right order and the wrong hardness, Arketa the right hardness and the wrong order, Momence the softest wall and the weakest information.
0 of 4 model equipment. The specialist vendor in the category sells spot level booking for bikes and explicitly does not offer it for reformers, in its own marketing.
0 of 4 treat level as a field. It is proven three times on three platforms: studios smuggle level, price and location into the class name because there is nowhere else to put them.
0 of 6 admin products open a record in a drawer. All six navigate away to a full page, and Cloudscape publishes the rule saying they should not.
1 of 12 metric rows does anything when you click it. Plausible's row is the chart's tab bar and fits six metrics in 70px, where shadcn spends 130px on four inert ones.
1 of 9 products that show a percentage delta print what the delta is a delta of. Everyone else makes the reader do the arithmetic.
5 of 14 admin products open with a row of equal KPI cards, and all five are ranked among the least distinctive screens in the set. Rejecting that row puts us in the majority, not on a limb.
3 of 3 white label platforms stamp their own name on the studio's member facing surface, one of them directly under the payment form.

How to read the counts

A product is yes or no only from an explicit statement in a findings file, or from a screenshot opened and read during this pass. Everything else is unknown. Two claims were resolved this way: Kimai's landing screen and Cloudscape's in product topbar.
Denominators exclude products a dimension cannot apply to, and every exclusion is listed in the JSON as not_scored. A small honest denominator was preferred over a padded one.
Two products (Directus, Saleor) are in the product list and in no denominator, because their captures yielded no product at all.
Where the research contradicts itself, the conflict is named and a side is picked with a reason. There are nine of them, listed at the end.

Denominator bases. Admin products evaluable: 14. Plus two component galleries where metric dimensions apply: 16. Products with an observed table: 10. Products with an observed record detail: 6. Booking platforms walked live: 5, of which class booking products: 4. Published design systems surveyed: 8, of which two could not be read.

1. Layout and shell

Feature	Count	Who has it	So what
Opens with a row of equal KPI cards	5 of 14	Ant Design Pro, Tabler, shadcn admin, Umami, Refine	All five are ranked among the least distinctive screens in the set. The correlation is stated outright in the research: a KPI row is what you reach for when you have not decided what the user's first action is
Sidebar is collapsible	6 of 10	Grafana, Cloudscape, Refine, Ant Pro, shadcn, Umami	Table stakes. Its absence would read as an omission, not as restraint. Ant Pro's handle straddling the sidebar edge is the one worth copying
Sidebar uses text group headings	2 of 10	Umami, shadcn admin	The field is against them. Refine states the threshold: eight items do not need headings. Our studio app crosses it, the portal does not
Global search in the topbar	3 of 14	Grafana, shadcn admin, MUI template	A differentiator, not a catch up. Member lookup by name is the most repeated front desk action in our product and 11 of 14 admin products do not offer a global search at all
Command palette surfaced	2 of 14	Grafana, shadcn admin	Low value as built. Plausible instead prints a one letter key cap beside all fourteen options in its date menu, which is the only keyboard affordance surfaced in the UI anywhere in the teardown
Breadcrumb	4 of 14	Grafana, Cloudscape, MUI template, Evidence	Three of the four products that skip it are the three ranked most distinctive. Atomic CRM, Ant Pro and Plausible all independently ship "no page title, no breadcrumb, no greeting"
Page or record level tabs	5 of 14	Cloudscape, Ant Pro, Refine, shadcn, Plausible	Low confidence, 7 unknowns. Cloudscape's variant is the one to take: a summary card of four identity facts pinned above the tab strip so you never lose which record you are in
Density or compact toggle	3 of 14	Cloudscape, Kimai, Ant Pro	Rare. Follow Cloudscape's published rule rather than the count: comfortable by default, compact in the HQ console only, never in the studio app where it would shrink touch targets
Ships both light and dark	8 of 16	Cloudscape, Atomic CRM, Tabler, shadcn, Mantine, Tremor, Umami, MUI	Exactly split. Grafana ships dark only and Plausible light only, and both are ranked at the top of the distinctiveness list, so one committed theme per surface is defensible

So what for the shell. The distinctive products delete chrome rather than arrange it. Plausible has no navigation at all, Atomic CRM has four top tabs and no sidebar, Evidence has a two item text rail. shadcn admin stacks three navigation levels before any data and ships two visibly disabled tabs, and it is the one the client called average.

2. Data and tables

Feature	Count	Who has it	So what
Whole row click opens the detail	3 of 10	react-admin, Atomic CRM, Refine	The majority put navigation in the first cell or in an overflow menu. Cloudscape's published rule says first cell link plus whole row hover, and only its own product follows it. This contradicts our current spec, which says the row is the hit target
Detail opens as a drawer or side sheet	0 of 6	nobody	The single most decision changing count in the admin set. Every one navigates to a full page, and Cloudscape's own decision table uses our exact case ("who is booked into a class, seen from the schedule grid") as the worked example for a side sheet
Checkbox selection with batch actions	6 of 10	Cloudscape, Kimai, react-admin, Refine, Ant Pro, shadcn	Expected where batch actions exist. Cloudscape's disabled but visible header buttons beat a bar that appears on selection, because you can see what selection would enable before selecting
Filter state persists in URL or session	3 of 14	Grafana, Kimai, Plausible	Nearly free and nearly nobody does it. 0 of 14 ship a named saved views feature; the closest anything comes is Plausible's shareable filtered URL
Column chooser	3 of 10	Cloudscape, react-admin, shadcn admin	HQ console only. Every column on a roster is load bearing, so a chooser there is a way to break the screen
Multi column sort with visible precedence	1 of 10	Cloudscape	Cloudscape's numbered `1▲ 2▲` indicators appear nowhere else in the teardown. Its own rule says single column by default and multi only for cross dimensional analysis, which is exactly the HQ console
Detail is the edit form, no read mode	1 of 6	react-admin	Removes an entire screen, an entire mode and an Edit button from the build. Keep its maximum separation of Save and Delete, drop its disabled until dirty Save per Cloudscape's rule
Designed empty or null state observed	5 of 14	Grafana, Cloudscape, Kimai, Atomic CRM, Refine	Grafana's is the best in the set: `No data` rendered as the stat's value in the same 40px type a real number would use, so the layout never reflows between loading, empty and populated. Tabler ships a route called "empty" that is literally blank

So what for tables. Two rules from the same source resolve the recurring conflict: hover the whole row so the eye can scan it, navigate from the first cell so checkboxes, links and inline buttons can coexist in the row. And the drawer count is the opening: nobody in either product set keeps the list visible behind the record.

3. Charts and summary numbers

Feature	Count	Who has it	So what
The metric row is also the chart's control	1 of 12	Plausible	Six metrics in 70px that double as the chart's tab bar, against shadcn's 130px for four that do nothing. This is the direct answer to the client's rejection of the KPI row and it is essentially unique
Comparison baseline printed as a number	1 of 9	Metabase	`up 0.9% vs. previous month: 1,217`. Everyone else ships `+20.1% from last month` and makes you work out what it is 20.1% of. One extra field, and it is the difference between a number and a fact
Heat strip or segment small multiple	1 of 16	Tremor	Ninety segments, 7px wide, 24px tall, one per day, with `90 days ago` and `Today` as the only axis. Ninety data points, one line of type, zero chart chrome
Metrics live in the chart's legend slot	1 of 16	Tabler	Removes the metric row instead of redesigning it. The numbers move into the legend slot the chart needed anyway, colour bound to the lines
Metrics set into the page header band	1 of 16	Ant Design Pro	Zero vertical space spent on a metric row. Includes a rank rendered as a fraction, `8 / 24`
Annotations on a time series	2 of 12	Grafana, Plausible	Only Grafana pairs the marked chart with a bound event log beneath it, so you read the shape above and the reason below without a tooltip
Sparkline inside a summary metric	5 of 16	Refine, Ant Pro, Tabler, Tremor, MUI	Three separate good ideas across five products and no product combines them: Refine's shared baseline, MUI's direction coloured fill, Tabler's dashed previous period ghost line
Sparkline or bar inside a table cell	2 of 10	MUI template, Tabler	A whole trend per row without leaving the table. Plausible's related trick is cheaper still: the proportional bar is a tinted background on the row itself, so the encoding costs zero vertical space and the row stays clickable
Metric written as a sentence	8 of 16	Grafana, Cloudscape, Ant Pro, Tabler, shadcn, Mantine, Plausible, Evidence	Exactly half. Proves the pattern is normal, not eccentric, which matters because it is one of our two signature candidates
Gauge, dial, half donut or ring	4 of 16	Refine, Ant Pro, Tabler, Mantine	The research is emphatic against all four for this product. Refine's speedometer spends a 380 by 360 panel on one percentage and its needle implies precision the data does not have
Chart tooltips	0 yes, 0 no, 16 unknown	not answerable	A screenshot harness cannot capture hover. This is a method limitation, not a finding, and it is the one thing on this list still to verify with a live pass

So what for metrics. The three best answers to the four card row all cost less space than the row they replace: Plausible's control band (70px for six), Tabler's legend metrics (zero added), Ant Pro's header band metrics (zero added). None of them is a better looking card.

4. Craft

Feature	Count	Who has it	So what
Skeleton loading rather than a spinner	1 yes, 1 no, 14 unknown of 16	Grafana skeletons, Plausible spinners	Mostly unanswerable from static captures. Material is the only design system publishing numeric thresholds and they are testable, so use those instead: nothing under 200ms, skeleton to 5s, determinate progress beyond
Any keyboard shortcut surfaced in the UI	3 of 14	Grafana, shadcn admin, Plausible	Plausible's date menu prints a key cap beside all fourteen options, teaching the shortcut at the moment of use rather than hiding it in a help modal. The other two just show a chip in a search box

So what for craft. Motion, focus states, route transitions and toast versus inline confirmation are not observable from a screenshot harness and are not scored here. They are covered instead by findings/03-motion-craft.md, which reads real source constants (Sonner, Vaul, Motion) rather than inferring from images, and by the prescriptions below.

5. Booking, the category we are entering

Feature	Count	Who has it	So what
Capacity denominator shown	0 of 4	nobody	The single biggest opening in the category. Two studios were so short of it that they typed "Limited spots available per class" into their own hero band by hand, where it is unverifiable and never updates
Waitlist position or depth shown	0 of 4	nobody	`Join waitlist` and `Waitlist` are the whole story anywhere in the set. It is the difference between joining and giving up
All three free before commitment	0 of 4	nobody	See fullness, see price, book, without an account. The compound test nobody passes
Apparatus or equipment modelled	0 of 4	nobody	Walla sells "bike spots" for cycling on its own modality card and offers no equivalent on its Pilates card. The specialist has said out loud that this does not exist for reformers
Spot or machine pickable	0 of 4	nobody	Same source. Members have a preferred machine and no platform lets them say so
Level is a real field	0 of 4	nobody	Proven three times on three platforms: `CultReformed: Reformer Level 1.5`, `Group Apparatus 1`, `37th St -New Client Orientation Class $20`. Level, price and location all get smuggled into a text field
Teacher is browsable	0 of 4	nobody	Every product renders the teacher as a 20px circle and a first name. The consumer app members arrive from, Alo, puts INSTRUCTORS in its primary navigation
Studio controls typography inside the widget	0 of 3	nobody	Measured on Arketa: the studio owns everything down to a 1px border it drew itself and nothing one pixel inside it. The only studio specific token in the whole widget is a `utm_source` parameter
Cancellation policy is its own field	0 of 4 (3 unknown)	nobody	Punchpass concatenates the class description, the parking directions and the refund policy into one free text blob
Fullness visible before login	1 of 4	Punchpass	And it paints it backwards: `CLASS FULL` is white on solid teal, `1 SPOT LEFT` is a quiet grey chip. The urgent state is loud and the scarce state is quiet
Price on the schedule row	1 of 4	Momence	And it puts the price in the top right corner and the button in the bottom right of the same panel, roughly 480px apart with a rule between them
Guest booking, no account demanded	1 of 4	Momence	The only genuine guest checkout in the set. Punchpass walls before availability, which is the worst possible position; Arketa walls after a $473.02 total
Default view is a seven column week grid	1 of 5	Arketa	Right shape, wrong content. The other four are two agendas, one single day view and one month grid. Nothing in the category treats a weekly slot as a relationship a member has
Studio timezone stated on the row	1 of 4	Arketa, mobile only	Its desktop grid is the degraded one. Momence's schedule and its own checkout, one click apart, disagree by five hours for the same class
Substitution shown inline on the row	1 of 4 (3 unknown)	Punchpass	Called the best single detail in the category, and it costs nothing. It exists because members care who is teaching
Reviews or ratings anywhere	1 of 4	ClassPass	12,500 ratings, a histogram, and every review tied to a named class. No studio software has any review concept at all
Empty state names a recovery action	1 of 4	ClassPass	A live Arketa studio returns `No Classes` for all seven days with no explanation and no next available date
Paginates inside a single week	1 of 3	Momence, as a defect	Caused by a 583px phone layout pinned into a 1280px viewport
Price visible without an account	3 of 4 (1 unknown)	Punchpass, Momence, Arketa	The category does disclose price. It is availability it hides
Waitlist offered at all	3 of 4 (1 unknown)	Punchpass, Momence, Arketa	Table stakes. The gap is not the waitlist, it is everything the waitlist refuses to say
Platform stamps itself on the member screen	3 of 3	Punchpass, Momence, Arketa	In the page title, in the footer, and in Arketa's case as a badge directly under the payment form
Named layout defect at 390px	5 of 5	all of them	Every member arrives on a phone and every product breaks something there. Momence spends roughly 470px of an 844px screen on filters before a single class appears

So what for booking. The category hides the one number that decides the booking and shows the one number that decides nothing. Fullness is communicated by the verb on a button in three of four products, while the studio side of the same platform is sold a "Capacity filled 20.4%" tile. The number exists, the studio sees it, the member never does.

6. What published design systems prescribe

Counts across the eight systems surveyed. Two of the ten intended sources could not be read (Atlassian is client rendered and returned only navigation chrome; legacy Polaris usage guidance no longer resolves), which is why some denominators are five rather than eight.

Prescription	Count	So what
Publish a KPI card or stat card component	0 of 8	The component the client rejected does not exist in any mature system. What they offer instead is Cloudscape's key value pairs, a one to four column label over value grid inside one container with no per metric chrome
Publish numeric loading thresholds	1 of 8	Only Material: nothing under 200ms, loading indicator 200ms to 5s, progress indicator beyond 5s. They are the only testable numbers available, so they become our acceptance criteria
State one primary button per page or section	3 of 8	Nord's per section scope is the version that survives a real booking page, which has several independent commit points
Agree on empty state anatomy	4 of 4 that document it	Visual, heading, description, primary action, secondary action. They split on alignment (Primer centres, Carbon left aligns and publishes why) and on button weight (Cloudscape says always secondary)
Call the family "toast"	0 of 5	The client was right. Carbon calls it Notification with toast as one of seven types; Cloudscape calls it Flashbar and has no transient variant at all
Auto dismiss their transient notification	2 of 3 that publish a timing	Material publishes the argument against: auto dismissing messages are inaccessible to people who need more time. Our front desk is the exact at risk case, since they are looking at a member, not the screen
Scale icons from one master rather than redrawing	1 of 5	Nord scales from a 20px grid, which is the domain match: its icons exist to sit inline with small text, exactly like ours

So what. Two of these override the observed counts outright. Nobody ships a KPI card component, and nobody in the systems literature thinks a toast should be the record of what happened.

7. Our own system, scored the same way

From findings/04-distinctiveness.md, which audits design/style/ against the same evidence base.

Measure	Value	So what
Tells found in our own system	14 of 25 catalogued	We avoid every tell on our own anti pattern list and are still anonymous, because avoiding tells is not the same as making decisions
Component rules using the display serif	2, across roughly 10,000 lines, both at or below our own 20px floor	The brand's voice is declared in the tokens and absent from the product
Type scale range in the product band	2.08x	Against Linear at 4.4x in one block and ClassPass at 5x inside one sentence. No size in our system has the job of being looked at first
Occurrences of `--accent`	102, plus 25 of `--accent-soft`	A colour used 102 times is a background. No colour in our system appears exactly once, which is the single mechanism that makes Linear's amber work
Prescriptions in section 7 of the design system	0, against 13 prohibitions	A system defined only by what it refuses converges where every careful system converges, which is "good but average"
Symmetric `auto-fit` grids	6	Every collection in the product is an evenly divided row that cannot have a focal point by construction
Motion tokens in genuine use	61 micro, 17 component, 9 overlay, 4 springs	A real hierarchy, and it beats shadcn admin and MUI outright. Do not touch it

So what. The fix is additive, not corrective. Nine things in the system are genuinely designed and should not be touched. What is missing is a signature, and the recommendation is the arc meter: Jessica's concentric arc logo geometry used as the product's capacity primitive, on the single most repeated data point in the product.

8. Where the research disagrees with itself

Nine conflicts. Each is resolved with a reason rather than averaged.

Screen count. File 06 claims 113, its own ledger sums to 102, the filesystem holds 119 PNGs of which 14 belong to the two unavailable products. Picked: 105 from products that were entered, because the filesystem is the only checkable number. Plausible is also understated at 5 when 8 files exist.
Cloudscape keyboard support. File 01 records a Command+K chip, file 06 records none. Picked: none. File 01 captured the cloudscape.design docs index; file 06 walked the demo applications. Verified in deep/jpg/p02-cloudscape-02-dashboard.jpg: the in product topbar has no search field at all.
Punchpass capacity. File 02 says it shows nothing, file 07 says it is the only product that publishes exact spot counts. Picked: file 07. File 02 saw one studio with the setting off; file 07 walked four and quotes the literal strings from three.
Punchpass price. File 02 says price never appears, file 07 says the full list is public one click away. Picked: file 07, which names the route.
Arketa dual timezone. File 02 credits the checkout, file 07 shows the mobile row carries both zones and the desktop grid carries one. Picked: file 07, which probed both viewports.
"You made 265 sales this month." File 06's matrix places it in Atomic CRM's column; two prose passages attribute it to shadcn admin. Picked: shadcn admin. Atomic CRM is scored unknown rather than credited with something it may not have.
Grafana sidebar structure. File 01 describes ten chevron parents, file 06 calls them eleven groups. Unresolved, scored unknown. It changes no build decision.
Plausible loading. File 06 says none needed, file 01 says a bare spinner in a 380px empty container. Both true at different scopes: the fixed width metric row needs no loader, the breakdown panels use a spinner. Both facts carried forward.
The dash in file 06's feature matrix. The file states a dash means the product does not have it, but that cannot hold for rows like "loading state" where every product necessarily has one. Split the reading by row type, declared per dimension: feature presence rows score no on a dash, always present behaviour rows score unknown.

What this changes about GetEYT

Ranked by decision weight. Each is backed by a count above and specified tightly enough to build from.

Build these because nobody has them (the moat)

1. The spot meter, with a denominator. Backed by 0 of 4. A segmented bar of N cells where N is the studio's real reformer count, plus the literal string to its right. Open: 4 of 12 open in teal. Two or fewer: same meter, string in plum, because colour marks scarcity and not fullness. Full: meter filled at 40% ink, string Full. A reformer out of service drops the denominator and the row reads 4 of 11 open. The same component renders at the same weight on the schedule row, in the class sheet, on the staff roster and in the HQ console, so a member and an owner looking at Thursday 9:30 read the identical fact. This is the product's signature data component and it is drawn as the arc meter wherever there is one value and the eye lands, as a tabular number wherever many are scanned down a column.

2. Waitlist position, everywhere the waitlist appears. Backed by 0 of 4. The action label carries it: Join waitlist, 3 ahead. The portal carries it: You are 3rd of 5, in a section never mixed with confirmed bookings, plus an Auto book me if a spot opens toggle defaulting on.

3. Apparatus and level as real fields with their own chips. Backed by 0 of 4, proven on three platforms. Level chip and apparatus chip side by side on line 3 of the row, both filterable, Inter 600 at 11px, teal on 12% teal. The studio never types a level into a class name again. The same chips become the filter row, which is a single horizontal row of the studio's real class types, not five stacked rows of dropdowns.

4. Teacher pages, plus a Sub badge. Backed by 0 of 4 browsable, 1 of 4 substitution. The teacher's name on every row links to a real page with a credential line and their upcoming classes. Substitution is a first class Sub badge on the row, not a parenthetical inside the name. Follow a teacher is a portal feature: their classes sort first and a new slot notifies.

5. Guest booking, with the wall in the right place. Backed by 1 of 4 and 0 of 4 on the compound test. Order is availability, price, class selection, account, payment. Availability at zero clicks, drop in price at zero clicks, the full option set at one tap, then email only with a magic link and no password field. Collect email plus first name, create the account silently, send the link with the confirmation.

Build these because the best in class do them and nobody else does

6. The metric band is the chart's control. Backed by 1 of 12. Studio app home. One cream container, plum hairline dividers, six cells: CLASSES TODAY, BOOKED, WAITLIST, NO SHOW RATE, NEW CLIENTS, AVG FILL. Labels Inter 11px uppercase, values 22px, delta inline on the value's baseline. The selected cell gets a cream fill and a 2px plum top accent and drives the one chart beneath it. Cells are fixed width so a range change never shifts layout. This replaces the rejected four card row, adds two metrics and costs half the height.

7. Class detail opens as a side sheet, never a page. Backed by 0 of 6 and 0 of 4. 480px from the right on desktop, bottom sheet on mobile, schedule visible behind it. The same container carries the roster and the quick edit. The full class definition, and only that, gets its own page, per Cloudscape's rule that a split panel must not replace the details page.

8. Print the baseline with every delta. Backed by 1 of 9. A hard system rule: no percentage delta ships without the number it is a delta of. Retention 83 percent, down from 87 percent in June, never 83% (-4%).

9. Class fill as a tinted row background. Backed by 2 of 10. The next seven days as one row per class: day, time and teacher left, 9/12 right aligned, and the row background filled to 75% at roughly 8% teal. Under filled classes read as visibly emptier rows with no chart and no added height, and the row is already the click target. Highest value single steal from the admin set.

10. The term heat strip on class and instructor detail. Backed by 1 of 16. One segment per session across a term: teal ran full, amber ran under filled, muted cancelled. 12 weeks ago left, Today right. It answers "is this slot working", which a line chart cannot.

11. Metrics in the legend slot for the HQ console. Backed by 1 of 16. Studio names as the legend, each with its current figure at 22px under its name, hairline separated, right of the plot. Fourteen studios' current numbers at zero added vertical cost.

12. Key caps in the date range menu. Backed by 3 of 14 surfacing any shortcut. Today D, This week W, This month M, This term T, Compare X, right aligned in a key cap on every option. A studio manager opens that control twenty times a day. Per the motion research, keyboard initiated actions get no entrance animation at all.

Build these because the category's defaults are wrong

13. Filter and scope state lives in the URL, from day one. Backed by 3 of 14, and 0 of 14 saved views. Every filter, date range and studio scope. It gives us shareable views, a correct back button and deep links for free.

14. Never render a past day as No Classes, never paginate a week. Backed by 1 of 3 paginating and a live studio showing seven blank cells. Past days collapse to a 40px date stub, elapsed classes collapse to a 24px line reading 9:30 Reformer L1 · done, and the grid opens on today. An empty week names the next real class and renders it inline and bookable.

15. Studio local time is primary, labelled, and unchanged through every screen. Backed by 1 of 4. Viewer's zone appears as a secondary line only when it differs, on the row and identically at checkout. The five hour contradiction between one product's schedule and its own checkout is the failure mode to test against.

16. Price and action are one stack at the bottom of the cell. Backed by 1 of 4 printing a price at all, and separating it from its button by roughly 480px. Drop in price, then the saving prompt only when true, then the primary button. For a signed in member the price line becomes Free with your 8-class pack, 3 left.

17. Never pre select a subscription, never print a sentinel. From the checkout that ranks a $459 recurring charge above a $15 intro class and ships expires after 999999 days to paying customers. At most four options after Book, cheapest entry first: what you already own, the intro offer, the single class, then See all 9 options as a disclosure. If a pass does not expire the string is No expiry. Compute the per class rate, never store it, because the platform that stores it ships titles and bodies that contradict each other.

18. No GetEYT mark anywhere a member can see it. Backed by 3 of 3. It belongs in the HQ console. Per studio theming goes three levels deeper than the category ceiling: display typeface, accent, photography and copy, all the way through checkout, with no visual seam at the money moment.

19. Design 390px first. Backed by 5 of 5 having a named defect there. Every failure in the set is a desktop layout surviving badly on a phone. A class row must be visible in the first viewport. Filters are one chip row.

20. Four named empty states, not one component with a prop. Backed by 5 of 14 having anything designed at all. Nothing yet (positive imperative heading, secondary button), No matches (action text exactly Clear filters), No access (explanation plus a route), and errors routed to the notification system instead. Left aligned per Carbon. Plus starter content for a brand new studio: two example class types and one schedule template, labelled as examples, one click to delete.

Fix in our own system before building more

21. Give the display serif a real job. It appears in two component rules across roughly 10,000 lines, both below its own stated floor. One display step at 32px or above, spent on the class name, the day heading, and the single primary number per screen. Never in a table cell, never in an element repeating more than five times on a screen.

22. Break the accent's monopoly and add one colour used exactly once. 102 occurrences is a background. Reduce --accent to interaction only, then spend a single warm ochre on one semantic in the entire product: the class in progress right now. Written as an absolute rule, because a second use kills it.

23. Write prescriptions, not prohibitions. Section 7 is 13 bans and 0 rules that say do this distinctive thing. Every decision in this document that survives review becomes a prescription in that file, with its count attached as the reason.

Sources: findings/01 through findings/07, plus direct reads of deep/jpg/p03-kimai-01-home.jpg and deep/jpg/p02-cloudscape-02-dashboard.jpg to resolve two ambiguities. Every count in this document is reproducible from FEATURE-MATRIX.json, where each dimension carries its per product scoring and its file and line citations.