Privacy policy

Coasty Systems, Inc. ("Coasty", "we", "us") operates Coarena at coarena.ai and is responsible for your information. This version took effect on November 1, 2026. The licence you grant at sign-in and the rules for using Coarena are in the Terms, which control if the two disagree about the licence.

Our Notice at Collection under the California Consumer Privacy Act (Cal. Civ. Code §1798.100). Sections 3 to 5 give the detail.

Purposes: running the arena, your account and your tasks; scoring labels; preventing abuse; measuring how people find and join Coarena; emailing you; and building datasets we license to AI labs, which is a sale under California law (section 8). Retention: most categories are kept indefinitely (section 4).

Email founders@coasty.ai about anything in this policy. The founders read every request; there is no request form, phone line or ticketing system yet.

3.1 Your account, from Google

Sign-in is through Google only. Google gives us your email address (required), name, picture link and account id; we record when your account was created and last seen. We never see your Google password.

3.2 Email

We send email through Resend, in five kinds:

• Welcome: one note from the founders at sign-up; it cannot be switched off.
• Run complete: a task you posted finished. Sent only if you have left the site, grouped, at most three a day.
• Weekly standings: only in weeks the leaderboard moved.
• Comeback: after time away, only if a new model has joined since your last visit.
• Announcements: occasional product news, each sent once.

Every email but the welcome has one-click unsubscribe and a preferences link that needs no sign-in. We record delivery, opens and clicks. A fixed random half of accounts gets no run complete notices, so we can measure whether they help. No email contains your prompt, screenshots or which models ran your task. Only Resend receives your address for sending; we do not sell or rent it.

3.3 Your consent record

Which consent version you accepted and when (currently v12), an activity log entry, and an exact archive of every consent sentence we have shown. Older version names look like dates but are not; the time you agreed is recorded separately.

3.4 Cookies

• Judge cookie: a random id set on your first visit, with or without an account, kept one year and readable by our page scripts. It links your activity, so we treat it as personal information.
• Session cookie: keeps you signed in for up to 30 days.
• Referral cookie: the code from an invite link, kept 30 days and readable only by our server, to credit whoever invited you.
• Sign-up cookie: marks a new account for the welcome screen and sign-up measurement (section 3.7), kept 10 minutes, readable only by our server.
• Google advertising cookies: set by the Google Ads tag (section 3.7).

3.5 Activity before sign-in

When you sign in, everything made earlier in that browser under the judge cookie (votes, annotations, battles, activity, arrival records and page views) is attached to your account for good, and is no longer anonymous.

3.6 IP address

Rate limiters hold your IP address in memory for up to an hour; it is never written to storage. A salted hash of it is stored with each label and battle, for rate limiting and abuse prevention. A hashed IP address is pseudonymous, not anonymous.

3.7 Page views, arrivals and the Google Ads tag

Our own code records page views on our servers: the page (from a fixed list, otherwise "other"), the referring site's name, the link's utm tags, advertising click ids and invite tags (the rest of the link is dropped unread), your country code, your device, browser and operating system family, and your language. Not your IP address, full browser string, full links or screen size. Published result cards are never logged.

At sign-up we keep the browser's first and most recent arrival and its latest Google click id, so the sign-up can be reported to Google Ads as a conversion. Arrival records never linked to an account are deleted after 180 days.

Google Ads. Our pages load Google's advertising tag with ad storage allowed; ad user data, ad personalisation and analytics storage denied; and personalisation signals off. On the welcome screen it sends Google a sign-up conversion with a random transaction id and no name, email or account id. Google sets its own cookies through the tag. There is no other third-party analytics, session recording or tag manager.

3.8 How you use the replay while judging

Recorded around each vote without asking: time taken, how far you scrubbed, how often you played the replay, where you paused and what you copied. It ships in the evaluation tier (section 8) without your name or email.

3.9 Tasks you write, and the files you attach

The prompt you type, an optional starting link, and a category we derive from it.

Before any agent runs, a prompt with no attachments is sent to OpenAI to decide whether to run it as a task or answer it directly; a follow-up also sends the original prompt. This happens whichever models then run, up to a limit per account. A prompt that runs goes word for word to the providers running its agents and into datasets as the task. Prompts are stored as typed and not filtered for personal information, so put nothing in one you would not want licensed.

Files you attach are saved in private storage with the task, with no scheduled deletion. You can remove a file before it is attached to a task that runs. What an agent reads from a file reaches its model provider, and files are within the licence you grant at sign-in.

3.10 Replies from Coarena

When a prompt is answered directly instead of run (a greeting, a question about Coarena), we keep your message and the reply so the conversation survives a reload. Only you can read them; they never enter datasets.

3.11 Saved sign-ins and agent memory

Signed-in accounts can store both at /vault, under the current consent, v12.

Saved sign-ins: a username, password or token for a named site, at most 25 per account, encrypted before storage. The agent asks for a sign-in by name and our system types it; the agent never learns the value. The screenshot of that moment is never kept, and on a desktop the rest of the run's screenshots are withheld. A battle using a saved sign-in, and every follow-up, is private, never offered to judges and never in a dataset; its screenshots are for you alone. The machine's signed-in state is kept sealed for up to 7 days so follow-ups can continue. A revoked sign-in is unusable at once; its encrypted record is kept, marked revoked, and never opened.

Memory: a note of up to 4,000 characters, added as it stands to both agents' instructions when you start a battle. It reaches the model providers, and agents may repeat it in reasoning recorded with the battle. Clearing it deletes it.

3.12 Votes, annotations and verdicts

Your preference between two runs, its strength and any written reason; flaws you mark; step verdicts; and your verdict and note on a single agent run. In datasets your preference ships as a label, flaws ship combined (type, label, judge count, no judge id) and step verdicts within the judged record. Written notes, reasons and single agent verdicts are not delivered.

3.13 Your judge quality profile

Your agreement with known answers, your answers to test items, a trust state with a reason, and retracted labels. It assesses you as a judge, is linked to your cookie and account, and decides whether your labels count. Retracted labels are kept so discard rates can be recalculated. Buyers get only combined quality statistics, never your profile.

3.14 Screenshots, page text, agent files and links

One screenshot per step, a masked copy where something was masked (section 6), and a text snapshot of the page. The snapshot holds everything the page rendered, which can exceed the screenshot, and is not masked; when used to check an agent's result, its text can reach buyers after personal information filtering. Files an agent produces or copies (text, tables, code) carry whatever was on the page and are filtered when saved. Visited links are filtered before delivery, but the registered domain always ships, so buyers can see which sites your task visited.

3.15 Feedback, invitations and dataset access requests

• Feedback: your message (obvious secrets removed), the page and your account or cookie id. A person reads it; it never enters datasets.
• Invitations: who invited whom, and when the invited account first took part.
• Dataset access requests: organisation, email, intended use and tier, kept for sales. There is no payment in the product; a person issues access after terms are signed.

Most categories have no automatic limit. If we set one, this table changes.

The full list, with what each company receives, is at /subprocessors.

• Model providers: Anthropic, OpenAI, Google, Fireworks AI, Meta, xAI and OpenRouter. The provider running an agent receives your prompt word for word, every screenshot unmasked, what the agent reads from your files, and your memory note. Fireworks AI hosts open-weight models by Moonshot AI, Alibaba (Qwen) and DeepSeek and receives the data itself; it also hosted Thinking Machines’ Inkling until we retired that model from Coarena on November 3, 2026. OpenRouter forwards to a fixed upstream: Z.ai for GLM, Xiaomi for MiMo, Moonshot AI for Kimi K3.1 Pro, Tencent for Hy4 and Together AI for MiniMax M3.1 Flash (MiniMax itself served that model for its first battles on November 5, 2026). Yutori received the same data while it ran models here and is no longer in use. Anthropic keeps requests on our account for 30 days, as Claude Fable 5 and 5.1 require, and those models run safety checks that can stop a run.
• OpenAI, as prompt router: prompts without attachments, before any agent runs (section 3.9).
• Google: the source of your account details and, as advertiser, the recipient of the sign-up conversion, its cookies and the click id (section 3.7).
• Infrastructure: Supabase (database, private file storage), Amazon Web Services (hosting, network, security and country code, and the hosts for the emulated Android phones that phone battles run on, each phone restored to a clean image after each run), Daytona (isolated agent machines, destroyed after each run) and Resend (email). Render (hosting) and Cloudflare (network) carried the site before it moved to Amazon Web Services and are no longer in use. The Decodo egress proxy is configured but switched off; a computer run's traffic goes out directly, and a phone's traffic leaves its host through our own filter.
• Dataset buyers (AI labs): section 8.
• Anyone: the public sample (section 8), with no screenshots, reasoning or judging behaviour; and any result card you publish, which shows your prompt and outcome to anyone with the link, without the replay, and is never in our sitemap or any index.
• Our operators: can export the database for backup and support, without screenshot files.

6.1 How detection works

Detection is structural, with no text recognition or vision model. On a web page we ask the page where its personal information sits (filled-in form fields such as passwords, emails and phone numbers, and text matching the published patterns), cover those areas with solid boxes, and check every box in the saved image before use. Without a web page there is no detection. Most battles now run on a desktop, so most recent screenshots are unmasked.

6.2 The five outcomes

1. Masks found, applied and checked: the masked copy is shown and delivered.
2. The detector ran and found nothing: the original is used; only that scan stands behind it.
3. Desktop battle: a whole-screen capture where masking is impossible; the original is used.
4. Phone battle: a whole-screen capture of an emulated Android phone; there is no page structure to mask; the original is used.
5. The detector failed, a mask could not be checked, or the capture failed: withheld in the product from everyone but the battle's owner.

So a screenshot can show whatever was on screen, including names, addresses, account details or a face. Screenshots of a saved sign-in being typed are never stored (section 3.11).

6.3 Model providers always see the raw screen

Masking covers what we store, show and deliver. The agent must see the screen to act, so every step's unmasked screenshot goes to its model provider.

6.4 What the detector cannot see

Desktop and phone screenshots, text inside images, canvas or WebGL drawing, video, PDFs in the browser viewer, embedded frames, closed components, text generated by style sheets, and personal information split across page elements (a card number drawn in four pieces). The text filter misses personal names, postal addresses, unlabelled identifiers, and unlabelled passwords or keys. No person has audited delivered material for what slipped through.

6.5 Who can see a screenshot

• You, in the replay of any battle you posted.
• Other signed-in judges, once you offer a battle for judging (ticked by default on the task form; untick it and only you can open the battle). Judges see the replay, including unscanned desktop and phone screenshots, without being told which models ran or who posted it.
• Anyone holding a viewing link, briefly. Each screenshot loads through a link with a pass valid for 15 to 30 minutes; anyone given the link can open it until then. Screenshots cannot be listed or fetched without a pass.
• Dataset buyers. Every dataset package includes each step's screenshot: the masked copy where a web page was masked, the original otherwise, including desktop screenshots. The self-serve download tiers include no screenshot files, and screenshots from saved sign-in battles are never delivered. Phone battles are not in any dataset.

Saved sign-ins are sensitive personal information under California law, collected only if you save them (section 3.11). We use them only to perform your task, never sell, share or deliver them, and never use them to infer anything about you.

Other sensitive information can appear by accident in a screenshot or page text, and can reach a delivered record, so the right to limit its use applies (section 9.5):

• Log-in details on screen: masked on scanned web pages, not on desktop or phone screenshots; the text filter misses unlabelled credentials.
• Precise location: none derived; a map or address can appear in a screenshot.
• Racial or ethnic origin, religion, health, sex life, union membership: not collected, inferred or used, but can appear on a page an agent visits.
• Biometrics: faces can appear; we run no face detection and derive no face data.

Yes, we sell it. Our contracts call it a licence, but under California law licensing datasets to AI labs for money is a sale. By tier:

• Public sample (free, no account or key; a public disclosure, not a sale): task prompts, blind preference outcomes, the first three actions of each run, and which agent ran and how it ended.
• Preference: the above without limits, plus the models' reasoning and the flaws judges marked, with judge counts.
• Trajectory: the above plus what the agents observed at each step and any errors.
• Evaluation: the above plus judging behaviour (section 3.8) and full battle details.
• Packaged deliveries: prepared for a buyer, including every step's screenshot (section 6.5).

Not sold: your email, name, picture or Google id. Delivered records have no field for who posted or judged, only how many judges labelled something. Written notes, saved sign-ins, memory notes and anything from a saved sign-in battle are not delivered.

No sharing for cross-context behavioural advertising. We advertise on Google Ads with personalisation denied (section 3.7) and run no other advertising technology.

Under 16: we have not knowingly sold anyone's information under 16; with no age signal (section 12), this states what we know, not a control.

Datasets already delivered cannot be recalled (section 9.2).

Send any request to founders@coasty.ai.

9.1 Right to know and access

You can ask what personal information we hold about you, by category and in detail, where it came from, why we use it and who receives it. Your account page shows your counts and recent activity; a person assembles the rest, as there is no self-serve export yet. We answer within 45 days, extendable once by 45 more if we tell you.

9.2 Right to delete, and withdrawal

Withdraw yourself on your account page. It acts on your account and the judge cookie it adopted:

• Every task you submitted becomes private at once, leaves every dataset tier including the public sample, and is no longer offered for judging. This cannot be undone.
• Your votes, annotations and step verdicts leave future dataset deliveries and the public rankings and statistics. They stay in our database, excluded, so discard rates can be recalculated.
• Signed-in machine state kept for sign-in battle follow-ups is erased.
• Datasets already delivered cannot be recalled.
• Your account record is not deleted.

You can also email founders@coasty.ai and we will withdraw for you.

Full erasure of your account record, activity log, telemetry, rate limit records and judge profile is done by a person on request. You can revoke saved sign-ins and clear memory yourself at /vault.

9.3 Right to correct

Email us and a person corrects the record. Your name, email and picture are copied from Google at each sign-in, so change them there. Archived consent text is never rewritten, because it records what you were shown.

9.4 Right to opt out of sale

There is no "Do Not Sell or Share My Personal Information" link or opt-out setting yet; email us, or withdraw (section 9.2). Taking part and granting the licence are one act today, so you cannot keep posting or judging while opted out. Unsubscribing from email does not affect the licence.

9.5 Right to limit use of sensitive personal information

Saved sign-ins are used only for your task and can be revoked at /vault. For sensitive information that appears by accident, the only control is removing the affected records: email us the battle or screenshot.

9.6 Right to non-discrimination

Using your rights never costs you service, price or quality of service. We pay nothing for data and offer no discount for it, so there is no financial incentive programme; arena points have no monetary value.

9.7 Authorised agents and appeals

You may use an authorised agent: send written authorisation to founders@coasty.ai and we will verify your identity with you directly. There is no formal appeals process yet; reply to our decision and a different person will review it.

With an account, email from the Google address you sign in with, or use the self-serve withdrawal, which acts only on the signed-in account. Without an account (for example, you appear in a screenshot), see section 13: we cannot verify you against a record we cannot find, and will say so rather than delete the wrong thing. We use no identity verification service, ask for no government identification, and will never ask for your password.

We do not yet honour Global Privacy Control or other opt-out preference signals, although California requires a business that sells personal information to. A browser sending one changes nothing. Until we do, email us to opt out (section 9.4).

Coarena is not directed to children, and the Terms require you to be at least 18. We do not verify age (no birth date, age check or age signal), so we cannot tell if a user is under 16. We have not built the opt-in consent California requires to sell the information of people aged 13 to 15, or parental consent for those under 13. If you believe someone under 18 has an account, email us and we will withdraw its contributions and remove it.

Agents use real websites, desktops and phone apps, so screenshots and page text can hold personal information about people who never used Coarena: a name in a review, an address on an order page, a face, someone else's message. Masking does not reliably catch it (section 6).

What we do: the task form tells people never to point a task at anything private or signed in (a rule, not a technical control); screenshots need a short-lived pass (section 6.5); and every screenshot records its masking outcome, so a buyer can exclude unscanned desktop screenshots.

What we cannot do: search our data for a person (there is no text recognition over screenshots and no face or name index), or notify people in screenshots, because we do not know who they are. If you tell us you appear in one, we can act only on details that lead to a record: a battle or result card link, a site the agent visited, or an approximate date. If we cannot find the records, we will say so.

Your rights still apply: California residents whose information is in our systems can use section 9 with or without an account.

Screenshots and files are kept in private storage, and every read goes through our application, which checks who is asking. Saved sign-ins are encrypted before storage, and secrets are scrubbed from activity log entries as they are written. Sessions last 30 days; the judge cookie grants no access on its own. No method of storing or sending information is completely secure. Details: /security.

Coarena is operated from, and processes information in, the United States. Our position under the GDPR, the UK GDPR and Canadian privacy law is in the Terms; this policy is the California notice.

We update this policy when what it describes changes, and change the date at the top with it. We do not announce changes by email or banner. A change that widens what we sell or what the licence covers also changes the sign-in consent, which asks every account to agree again.

Coasty Systems, Inc., founders@coasty.ai. Email is the only channel for privacy requests; there is no postal address, phone number or web form yet.