A page is a contract, not a canvas

• h1
• subtitle
• primaryCta

Framed hero with the pricing banner (.hero-offer): « NEW PRICING », the entry price, no date

• eyebrow
• title
• subtitle
• cta.label
• methodologyLabel
• cardLabel
• measured
• disclosure
• measures.<key>
• close

Since 2026-09-24 (Antoine): the accuracy ranking of the latest edition, BlitzAPI in the brand blue. It replaced DataQuality. The only comparison allowed outside /benchmarks: the recipe lists this page in COMPARISON_PAGES, and tolerates competitor names only inside the block, which carries its dated scope and the disclosure

The /sculpt design since 2026-10-14: section heading, framed 16:9 facade with the HTML poster, nothing loads before the click

100 % driven by JSON: pillars, chips, proof

Dark matrix, one cell per workflow

Hidden in production while QUOTES_ARE_PLACEHOLDERS is true

BreadcrumbList JSON-LD

• h1
• subtitle
• primaryCta
• secondaryCta

secondaryCta = the deep link to the API reference page

Product variant: text cell + client marquee

• items[] with title, text, asset

Opening B, only when features is the FIRST section. Later features render as GrayCards

• items[] with title, text, href

« Related APIs », 3 cards. relatedProducts: [] removes the band

Native <details>, FAQPage JSON-LD emitted automatically

variant="strip", links to /pricing

• h1
• subtitle
• primaryCta
• a spec in motion/specs/use-cases/

Without a scene the hero is a centered text block

Use case variant: centered hook + marquee

• asset
• items[] text bullets

Text on the left rail, proof frame on the right

• items[] title + text

• href on EVERY item

The current page cell becomes « You are here »; every other href is typed by lib/destination.ts, or the build throws

• asset
• exactly one item {title, href}

Proof on the left, text + button on the right. Without the href item it silently falls back to « points »

• asset
• items[] with href; bullets separated by \n

The P1 frame shows UC_META.endpoint

proof is cut when the page already shows that signature

Native <details>, FAQPage JSON-LD emitted automatically

variant="strip", links to /pricing

• h1 (« … for <persona> » or **accent**)
• primaryCta

• pains split: title, subtitle, items[] (lead + bullets)
• pains split: ONE item {title, href} without text, the column button (label and link); no fallback
• features WITHOUT title: exactly 3 cells with title and text (throws otherwise)
• scene in motion/specs/teams/ + SCENE_LABELS

closesOnBand: no bottom padding, the gray band's top rule closes its grid

• title (**accent**)
• exactly 6 items with title and text (the outcome, not the call)
• href optional: without it, a plain CAPABILITY cell

Gray band; destination type from lib/destination.ts (shared with PlaybookGrid and NextStep), an href it cannot type throws; stretched title link, no aria-label

• title, subtitle
• items[]: {product: "/products/<slug>", job}, 4 to 8, no duplicate; job = what goes in, what comes back

Name, endpoint, family, link and plans read from the product JSON (sections/product-ref.ts), the call drawn by EndpointCall; optional note at the foot (agencies: client work; developers: Blitz Embedded); stretched name link

• title, subtitle
• steps[]: {product, text, output}, 3 or 4; text = what the step does with the previous output
• every step product listed in toolkit.items

Dark panel on the gutter; the eyebrow counts the calls; the call drawn by EndpointCall (tone dark)

• founder split: ONE item with title « Name, Role » and text
• stack split: title starting with « Plug BlitzAPI », one item with text

Four cells; no pipeline diagram, the chain above carries it

• docsUrl
• TEAM_BANDS entry

Native <details>, FAQPage JSON-LD emitted automatically

variant="strip", links to /pricing

• a spec in motion/specs/hubs/

• 3 items

• href only on what is really documented

• asset
• proof.head
• proof.foot

• 6 cells (full grid)

ONE « Where to go next » matrix, never two consecutive card bands

Native <details>, FAQPage JSON-LD emitted automatically

variant="strip", links to /pricing

align="center" compact

• three plans (Start, Scale, Build Your Infra), same decision at the same height
• specs: speed, fair use policy as a LEVEL (the number lives in the tooltip and the table), support

Quality reassurance without an accuracy figure

Four groups (what changes, included everywhere, custom terms, service), a decision row in <tfoot>

Primary CTA = cal.com booking, the contact form as a secondary link

• html
• name
• role

variant="about", authorized KPIs only

Principles, all unfolded

align="left": the left variant carries the section envelope geometry

• three states: sending, sent, error

• title
• description
• heading
• path

Never indent the markdown: 16 leading spaces once published most of the terms as code blocks

Opt-out: [hidden]{display:none!important} scoped rule, because state frames set display:flex

HomeHero geometry; the right column is the page visual in HeroVisualFrame (slot "visual": GuestPass on /sculpt, PerkDeck on /stateofgtme, decorative), the left column carries PromoCode (functional)

/stateofgtme only (Antoine, 2026-10-27): the home ranking, without links to the head to heads; the page is listed in COMPARISON_PAGES, names tolerated inside the block only

/stateofgtme only: five use cases in plain words, each row a link

• quote = verbatim with a dated source, or bio = prose, never a bio in quotes

Eric Nowoslawski (verbatim) and Jordan Crawford (bio) on both pages

/stateofgtme only: five flagship products + the stack cell

Inline facade, nothing loads before the click; YouTube, Loom, Vimeo or a file

The discounted plan (prop `plan`, Scale on both pages) large, the same code on the other self-serve plan in one line; discounted prices are computed

/sculpt only: 3 × 2 matrix, five teams + the stack cell

buttonAttrs carries data-promo-cta so the button flips when a dated code expires

• title
• description
• category
• publishedAt
• takeaways (3 to 5)

Also emitted as structured data

components/content/* (Definition, Callout, Steps, Compare, Stats…)

Visual mounted on the Visual Generator, caption required

« Keep reading »

Generative 16:9 cover, platform tag, level badge, lessons and minutes

Course landing

Reading bar, remaining time, keyboard arrows, resume

• Objective first
• Recap last

Objective → prose + shared blocks → Payload → TryIt → Troubleshoot → Quiz · Checkpoint → Recap

BreadcrumbList JSON-LD

• kicker
• h1
• one paragraph

• eyebrow with the Measured on line
• h2 linked
• one solid action

A BORDERED card (session 47, Antoine: make it read as a block of the page that gives access to the actions): 1px --color-border, white on the page background, square corners like the report recap, padding --pad-8xl. It wraps the eyebrow, the h2 and the action row, nothing else; the actions sit on a rule inside it. The h2 carries its underline AT REST: painted like the h1, nothing marked it as a link before hover, and there is no hover on a finger. The solid pill sits directly under it, in the first screen

• claim={false}: no claim, by decision

Right under the card, outside it. Its claim (« No provider here is ahead on every measure ») was removed in session 47 at Antoine's request: ChartFrame is called with claim={false} and renders nothing in its place; the readAs note stays. Columns are measures, rows are ranks. The order is DERIVED in the component, not written in the page: measures whose separation was tested first, then the one where the publisher places lowest, whichever rank that is, then the tightest race. On the September run that opens the figure on coverage. Rank 1 is marked by weight alone. The four-sentence rule paragraph under the figure was removed in session 46: what it said is IN the figure, in the column headers (separation not tested) and on the rows (too close to separate). Work history is deliberately absent: its 80 point gaps are endpoint scope, not quality

BELOW the figure: before it, it reads as a barrier to a result

Three, not six: more repeats /benchmarks/methodology. Step 01 and the hero lede DATE the queries from run.queried_on of the latest edition (dayRange() of viz.ts): « same day » is written only when every provider was queried on one day

ALWAYS rendered, even with one edition: a shape never seen on screen has never been measured. Year grouping keys off the SPREAD of years, not a count; label falls back to the long date for the whole list as soon as two editions share a month; data-benchmark on EVERY row. The fold over twelve editions is specimened on /brand-lab/sections with visible=0

cta id benchmarks.cta_section.primary

• kicker
• h1
• standfirst
• disclosure
• measured on line
• one action

The standfirst is report register, 25 to 40 words (session 47): a figure-free human framing of what the series measures, then a GENERATED sentence naming the edition and its providers in ALPHABETICAL order, never rank order. It carries no measure and does not repeat the Measured on line right under it. The disclosure names what the reference is FOR (« the reference every answer is scored against »): the shortened standfirst no longer introduces it, and the PDF reader meets it on page 1. The measured on line keeps the REFERENCE date (testDate); when the queries do not all fall on it, a « Queried on » item follows, read from run.queried_on (dayRange() of viz.ts), and the recap note, the Dataset JSON-LD and the footer date the queries the same way. The measured on line (BenchMeta, kicker off) and the action row (BenchActions: the download pill carries its « PDF, no email » note and no arrow) are shared with /benchmarks and the duel pages since 2026-09-23; the local copy had drifted. On paper the header goes back to two columns, so the cover carries the whole recap. The takeaways block was removed in session 46, and the field with it: a field a person writes every month and nobody reads is how it drifted into publishing a false count in the first place

Sticky under the nav, scrollspy, hidden in print. A 16px gap, so the ten entries fit whenever the list has its full width (viewports from about 1140px); below that it stays a scroller and fades its right edge while entries remain to the right (.rp-rail-list--more, set by the rail script), instead of cutting « 10 The report » to « 10 T » with no sign that the list goes on

Composed track over 100 contacts. Its exact numbers are OPEN, not folded. The middle part is « Returned, not verified accurate » (legend, aria-label, claim, readAs), never « wrong »: it is coverage minus accurate, so it holds the unverifiable answers too, and the home DataQuality block names the same part the same way

coverage x accuracy = accurate per 100

Coverage against accuracy, with equal-yield curves. No red. The lede names BOTH ways a record is lost: never returned, or returned and not verified accurate (different title, stale employer, unverifiable). It said « an employer the person has left » alone, which is false for Apollo, whose largest loss is the title; the page throws if the verdict keys ever differ from those four

Titled « What a wrong answer is made of », like its twin on the duel pages. The punch line (a run-wide « one in N ») and the cost note under it were removed in session 47 at Antoine's request, and « costs you » left the title with them: the section opens on a plain, figure-free report lede naming the four verdicts, and the measure is in the figure claim. Status palette, never provider colors

A grid read by color. Four switchable measures, all rendered at build. No lede since session 47 (both the screen and the print one were removed at Antoine's request): the title opens on the claim, which COMPUTES how many providers bottom out early AND names the measure it was computed on, since a claim placed above a switch is only true on one position. The print copy shows one measure only, so the figure carries a print variant of its reading note (readAsPrint) that says so

Segments led per dimension, then a bar in every cell

6x6 grid, then ONE line of links to the duel pages. Never link from the cells: the grid holds fifteen pairs and only five have a page. The line describes a duel page from DUEL_SECTIONS (duel/sections.ts), in its order: written by hand it skipped a section and would outlive a reorder, in the PDF too

Measures in rows, sticky first column. Four rows were removed in session 46 (share of stale that is a real previous role, start month agreement, share matched by identifier, billed units per accurate record): they describe how the test WORKS, not what a buyer gets. Check the prose around any row you remove: the cost note (itself removed since) pointed at « the billed units above » and outlived them by one build

Warnings are published, never smoothed

• one solid action

The four CSV, the aggregate JSON and the evidence pack form were removed in session 45: the free PDF is the only file that leaves the site. A print variant replaces the offer, because the reader of the PDF already holds it, and it gives the EDITION URL (built from the slug, like the PDF footer), never the directory. Neither variant says « nothing is written twice »: the page holds screen and print twins, and the PDF is printed by hand

Bounded to THREE, then a link to the directory: unbounded, a monthly report adds a card to the foot of every earlier report, in the HTML and in the PDF

Trademarks, corrections address, test id. KEPT in the PDF

cta id benchmarks.cta_section.primary

• measured on line
• disclosure before the first number
• link to the methodology

The query date of the lede is read from each provider’s run.queried_on, and the lede points at the shared-contacts section by its SECTIONS id, never by a number. The measured on line (BenchMeta) and the action row (BenchActions) are shared with /benchmarks. The measured on line keeps the reference date; when the pair was not queried exactly on it, a « Queried on » item follows, from the same run.queried_on as the lede (model.queriedMeta). A dated banner appears when the edition shown is not the latest, because the competitor left the run. Everything below the header is driven by DUEL_SECTIONS (duel/sections.ts): each entry is an id and a short summary, the page maps each id to its section component from components/benchmark/duel/ and throws on an id without one, the number is derived from the position and the anchor is the id. Every section reads the same model (duel/model.ts), so moving one needs no prop rewiring

• both sides always rendered
• the ledger sentence, generated

Five measures: accurate records per 100 (featured, the measure the report ranks on, NOT a composite), then its two factors, coverage and accuracy, then the wrong title rate and the stale employer rate. The row list lives in FACE_OFF (duel/model.ts). A side is marked ahead only where the gap clears both intervals; rows without an interval print the gap and « separation not tested ». The former section « Reach, where we come Nth » is folded in under the figure: the publisher’s coverage rank on the whole run, and the segment ledger sentence, which is the transparency device of the page and must not be removed without a replacement

Two bars instead of six. n is the two returned counts added, like the report figure: the shares are of returned contacts, not of the list. The claim is computed in duel/model.ts (errors.claim) from the stale share the figure draws, and throws when it is missing: it used to fall back to 0% in the section

• the base n written on the figure
• the counter line, rendered unconditionally

The four buckets partition the shared contacts, and the component THROWS if they do not. Never mix this accuracy with the face-off accuracy: two different bases. The counter line says « right where we are not », never « wrong » (the bucket holds our unverifiable answers), and comes from counterLine() of viz.ts, the same sentence as HeadToHeadRail. Its separation threshold is the two PUBLISHED accuracy intervals added together, which are computed on everything each provider returned, and the sentence names that base: no interval is published on the shared contacts

Two series, four switchable measures. The claim is computed on the default measure and NAMES it: a claim written above a switch is false on the other three positions, and the grid rounds to integers, which creates ties the source does not have. The lede is a reading instruction only, no universal claim

The Separation line is DERIVED from the face-off measures that carry an interval, and names them « the main measures on this page », never by their position. The work history line is conditioned on provider_exp_per_profile, never on a list of names, and points at no figure: work history is not on this page. The cost line says only that neither provider publishes a unit price that can be sourced and dated (the model throws if either run says price_is_public_list); its « units are not the same object » clause was false on the MoltSets page, where both bill records per returned profile

Two entries. Colors computed on the WHOLE edition then filtered, never on the pair

Unnumbered: it is an exit, not a measure