How we know what we publish
Everything on Decks comes from two things: an image we captured ourselves, and a label a vision model put on it. This page explains how each is produced, what the vocabularies mean, where the method stops, and how to tell us it got something wrong.
Published 8 September 2026 · Last updated 8 September 2026
How screens are captured
Every screen in the library is a first-party capture. We photograph the operator’s own live product with our own capture pipeline — we do not buy screens from a stock library, we do not take them from affiliate sites or press kits, and there is no public or automated way for anyone outside our team to put an image into the library.
Two fingerprints of that pipeline are visible in the images themselves, and we strip both out of the text we index so they can never be mistaken for the operator’s own words. The first is our capture chrome: a status-bar clock beside a Decks address, burned into the top of the frame. The second is our own capture account — the login we use to get past an operator’s signup wall, which the operator’s own forms and profile pages then echo back into the screenshot. That second one is why the library holds screens most people never see: the cashier, the identity check, the loyalty tiers and the logged-in lobby are all behind a registration.
Screens enter the library through our own internal tools, and creating one requires an administrator account. In practice a person decides that a set of screens belongs in the library, and a person puts them there.
What the date on a screen means
The date we show is when a screen entered our system — not when it was captured. Our database holds exactly one date for a screen: the moment its row was created in the library. There is no capture timestamp anywhere in it. So the interface in the image may have been photographed earlier than the date beside it, possibly much earlier.
This is the sentence we render beside those dates, and it is the only claim we make about them:
“Screens added to our library on {date}. That is the date they entered our system; we do not hold the date each screen was captured, so the interface may have been captured earlier.”
You will therefore never see “recorded on”, “captured on” or “as of” against a date on this site. If you do, it is a mistake and we would like to hear about it.
A second thing follows from it. Screens are added; they are not refreshed on a schedule. For most operators, everything we hold entered the library on a single day, so a page’s “library dates” line is often one date rather than a range, and an interface may well have changed since. The dates are shown on every page precisely so you can judge that for yourself rather than taking our word for how current the library is.
What a model wrote, and what 70% means
Six families of field on this site are produced by a vision model reading the screen image: device, vertical, screen type, licence marking, brand, and the written description of the screen. The persuasion-tactic tags come from the same pipeline. Nothing in that list is the operator’s own statement about itself, and none of it is a human classification.
This is the disclosure we render beside those fields, in full, wherever they appear:
“Screen types, licence markings, verticals, brand, device and the screen descriptions on this page are all produced by our vision model reading the screen image, each from a fixed vocabulary we defined. The model returns nothing when it is under 70% confident, so every count is a floor, not a census. These are our model's classifications — not the operator's statements, and not human review.”
The 70% floor is a real instruction, not a description. The classifier is told, in the prompt itself, to be conservative and to return an empty result rather than guess whenever it is less than 70% confident. That is a deliberate trade: we would rather hold a screen with no label than publish a wrong one.
The consequence is that every count on this site is a floor. If a page says we hold twelve identity-verification screens for an operator, what that means is: at least twelve of the screens we hold were confidently identified as identity verification. Some of the others may be too, and the model declined to say so. A low count means we hold few screens we could confidently label — it never means the operator’s flow is short, absent, or simple.
There is no human review pass over every classification before publication. A person on our team can correct any of them, and does when one is reported — that route is here.
The vocabularies we classify against
The model does not invent labels. For five of the six families it picks from a fixed list we wrote, and it may pick more than one term for a single screen. These are those lists and what we mean by each term. The terms are ours; the assignment of a term to a screen is the model’s.
Vocabulary sizes and membership measured 8 September 2026. We add terms rarely, and a term is added when the library needs one — so this list can lag the product by a term or two.
Screen type — 57 terms
What the screen is. This is the vocabulary the library is searched by, and it is the only one where a single screen routinely carries several terms at once — a cashier screen with a bonus pop-up over it is both.
“A screen can carry more than one type, so these counts add up to more than the number of screens.”
We classify the tiers-and-rewards surface as Loyalty. Operators themselves almost always call it VIP. They are the same surface under two names: if you search our library for “VIP” you will find far less than we hold, and the term you want is Loyalty. This is the one place our vocabulary and the industry’s differ enough to lose you results.
- Casino Lobby
- The game-browsing surface of a casino: categories, providers, search and game tiles.
- Homepage
- The operator's front page, as it loads before the visitor goes anywhere.
- Game
- An individual game open and running, in its own frame.
- Promotions
- The list or hub of current offers, bonuses and campaigns.
- Pop-up
- An overlay shown over another screen — an offer, a consent prompt, a confirmation, an interruption.
- Wallet
- The balance and payment-method surface: the cashier's home screen.
- Sports Lobby
- The sportsbook's browsing surface: sports, competitions and upcoming events.
- Live Betting
- In-play betting on events already under way.
- Odds
- A price list for a market or an event.
- Deposit
- The flow for putting money in: amount, method, confirmation.
- Profile
- The player's own account surface: their details, preferences and status.
- Registration
- The sign-up form, or one step of it.
- Daily Rewards
- A recurring claimable — a daily bonus, a streak, a wheel, a calendar.
- Responsible Gambling
- Limits, self-exclusion, reality checks and the operator's safer-gambling material.
- Menu
- A navigation surface: the drawer, the tab bar, or an expanded menu.
- Loyalty
- The tiers, points and rewards programme. Operators usually call this VIP — see the note above.
- KYC
- Identity verification: document upload, proof of address, verification status.
- Slot
- A slot game.
- Bet History
- The record of the player's own placed bets.
- Bet Slip
- The selections a bettor has staged, with stake and potential return.
- Tournament
- A competition between players — its rules, prizes or standings.
- Landing page
- A single-offer page a campaign points at, outside the site's main navigation.
- Customer Support
- The contact surface: live chat, a ticket form, a support hub.
- Dice
- A dice game.
- Leaderboard
- A ranked table of players — by wagering, winnings or points.
- Security
- The account-security surface: password, sessions, devices.
- Crypto
- A cryptocurrency surface: choosing a coin, an address, a network.
- Crash
- A crash game — a rising multiplier the player cashes out of.
- Game list
- A listing of games outside the main lobby: a provider's set, a category, search results.
- FAQ
- A list of answered questions.
- Mines
- A mines game.
- Withdrawal
- The flow for taking money out.
- Affiliate program
- The operator's affiliate or partner offer.
- Help
- Help content that is not a question list — a guide, an explainer, a how-to.
- Refer a friend
- The referral offer and the mechanics for sharing it.
- Plinko
- A plinko game.
- Notifications
- The message or alert centre.
- Buy Crypto
- Buying cryptocurrency inside the product, usually through a third-party ramp.
- Community Chat
- A chat between players, usually alongside the games.
- Statistics
- Figures about play: a game's history, a player's totals, a market's numbers.
- Roulette
- A roulette game.
- Store
- A place to spend points or currency on items, spins or bonuses.
- Loader
- A loading or splash state.
- 2FA
- Two-factor authentication: setting it up, the codes, or the challenge itself.
- Onboarding
- The guided first-run sequence after signing up.
- Keno
- A keno game.
- Vault
- A place to lock funds away from the playable balance.
- Blog
- An article, or the article index.
- Footer
- The page footer, captured as the subject of the screen.
- Feed
- A stream of recent activity: big wins, bets, posts.
- Autoplay
- A game's automatic-play controls.
- Tower
- A tower game.
- Limbo
- A limbo game.
- Giftcard
- A gift-card purchase or redemption.
- Error
- An error state.
- Feedback
- A surface asking the player for feedback or a rating.
- Forgot password
- The password-reset request, or one step of it.
Persuasion tactic — 25 terms
What the screen is doing to move the reader. Same rules as above: multi-label, model-assigned, and a floor rather than a census. A tactic tag says our model saw the pattern on the screen — not that the operator intended it, and not that it worked.
- Visual hierarchy
- Prominent calls to action, contrast, size emphasis.
- Framing
- Positive spin; benefit-focused language.
- Friction reduction
- Autofill, simplified forms, guest checkout.
- Gamification
- Points, badges, streaks, leaderboards.
- Bonuses and discounts
- Free gifts, percentage off, bundle deals.
- Trust badges
- Security seals, guarantees, verified marks.
- Urgency
- Countdown timers, limited-time offers.
- Curiosity gap
- Teasers, partial reveals.
- Progress steps
- Progress bars, step indicators, milestones.
- Social proof
- Reviews, testimonials, user counts, ratings.
- Personalization
- Use of the player's name, tailored recommendations.
- Commitment and consistency
- Saved preferences, profile completion.
- Chunked forms
- Multi-step forms, progressive disclosure.
- Default options
- Pre-selected choices, recommended picks.
- Loss aversion
- "Don't miss out", last-chance messaging.
- Authority
- Expert endorsements, certifications, awards.
- Risk reversal
- Money-back guarantees, free returns.
- Anchoring
- Crossed-out prices; original shown against sale price.
- Reciprocity
- Free content, or a gift before a purchase.
- Foot in the door
- Free trials, low-commitment entry.
- Scarcity
- Limited stock, exclusive access, spots remaining.
- Cross-sells
- Related products, frequently bought together.
- Exit intent
- Pop-ups and last-chance offers on the way out.
- Upsells
- Premium options, upgrades.
- Exclusivity
- Members-only or invitation-framed access. This term is in our vocabulary and in the data, but it is not one of the terms our current classifier is given, so this definition is ours rather than the model's.
The other three
- Device
- Whether the screen is a phone or a desktop capture. The library is overwhelmingly mobile.
- Vertical
- The kind of product the screen belongs to — casino, sports, sweepstake, bingo and so on. It is a LIST, not a single value: an operator running a casino and a sportsbook has screens tagged both, and many screens carry both at once.
- Brand
- Which operator’s product the screen came from, read off the image. The operator page you are on is the authority for that; the model’s brand reading is a cross-check, not the record.
Licence marking — an observation, never a status
A licence term records that our model saw a regulator’s marking in the image. That is all it records. It is an observation about pixels on a screen, not a statement about whether an operator holds a licence, and it must never be read as one. We do not check registers, we do not verify a marking against the authority that issued it, and an absence of a marking in our library is an absence in our library — the screens we happen to hold — and not a finding about the operator.
Where our model found no marking at all on any screen we hold for an operator, we say exactly that, and we say it as a gap in our data. Where it found several different authorities on different screens, we show all of them with their counts, because that is what we observed — and the reason it happens is the caveat in the next section.
What our counts cannot tell you
Four limits apply to every number on this site. They are rendered on the pages themselves, not buried here, and these are the exact words:
“Our library is a sample of what an operator showed, not a complete record of any flow. A low count for a screen type means we hold few screens of it — not that the operator's flow is short or absent.”
“Operators commonly show different regulatory markings to visitors in different countries. We do not record the country a screen was captured from, so these counts cannot be read as the marking an operator shows in any particular market.”
“A screen can carry more than one type, so these counts add up to more than the number of screens.”
“Screens added to our library on {date}. That is the date they entered our system; we do not hold the date each screen was captured, so the interface may have been captured earlier.”
Two more things we will not do, stated so you do not have to guess whether we are being selective. We publish no traffic, ranking, revenue or player-number figures on Decks — we do not measure them, so we do not report them, and we will not describe a “method” for a number we did not produce. And we make no claim about change over time: the library is not on a recapture schedule, so nothing here is a trend, a history or an alert, and you should not read two dates as a before-and-after.
Corrections
If something on Decks is wrong, tell us and we will fix it. Use the contact form, and give the URL of the page and what is wrong with it. The corrections page says what to include, what counts as an error, and what we do with each of the three possible answers.
We reply to every correction within five working days. Where a fix needs the stored image re-checked or a classification re-run it can take longer than that to land, and we will say so in the reply rather than going quiet on you.
Every correction that changes a published fact is listed below, with the date, the page, what was wrong and what we changed. The list starts here:
No corrections published yet. This log opened on 8 September 2026, the day this page went up — so it is empty because it is new, not because nothing has ever been wrong. When we correct a published fact we list it here with the date, the page, what was wrong and what we changed, including the ones we find ourselves.