Patina

Proof of tenure for accounts a person already owns
Anyone can make a new account. Nobody can make an old one.

Live App • Documentation • MCP Server • Verify a score

Overview • The Score • Privacy • Architecture • Quick Start • Integration • Test Results

---

Live Deployments

Surface	URL
App	patinadata.xyz
Docs	patinadata.xyz/docs
MCP	patinadata.xyz/api/mcp
Verifier	patinadata.xyz/verify
Registry	io.github.ramakrishnanhulk20/patina
Signing address	0x620dDbEceaD28Bbf1b979bfaB8e3a7B893aa54A1

---

Overview

Patina reads the history in accounts somebody already owns and turns it into portable, verifiable evidence that a real human has been there for years. It is built on Vana, the data portability protocol: the data lives in the user's own Personal Server, is read once under a grant they can revoke, and never passes through an OAuth handoff that keeps it.

Sybil resistance is unsolved and expensive. An airdrop, a quadratic funding round, a DAO vote or a free trial all need to know whether they are talking to a thousand people or to one person with a thousand wallets. The usual answers are poor: biometrics are invasive, KYC destroys privacy, and most behavioural signals can simply be bought.

Somebody running an account farm can buy followers, bulk-upload posts and fill in a convincing profile. What they cannot buy is a decade.

Why Patina?

	KYC	World ID / proof of personhood	Patina
Proves years of history	No	No	Yes
Proves you are one unique person	Partly	Yes	No
Works without a wallet	Yes	No	Yes
Works without biometrics	Yes	No	Yes
User can revoke access afterwards	No	No	Yes
Another app can reuse the proof	No	Yes	Yes

Patina answers tenure, not uniqueness. A freshly verified unique human can be produced today; sixteen years of corroborated activity cannot. The two are complementary, and the tool descriptions say so rather than implying otherwise.

---

Features

Ownership proved, not assumed

Collection is meant to run through Vana Desktop, which opens a browser on the user's own machine and asks them to sign in. Credentials never leave the device and Patina never sees a password.

But Vana has two collection paths and does not report which one it used. The other reads a public page from a profile URL, which proves an account exists and nothing about who holds it. So "desktop only" cannot be asked for; it has to be forced.

Every source therefore requires one scope a public page cannot serve, and the read is refused if it does not come back. LinkedIn must produce connection dates, Spotify saved tracks, YouTube its Watch Later list. A request containing something private cannot be answered from a public profile, so the person ends up on Desktop and has to log in. sources.test.ts fails if a source is ever added without one.

Steam was withdrawn for the same reason. Its connector never signs anybody in: it takes a Steam Web API key and a Steam ID, and a Steam ID is public. Anyone could have handed Patina a stranger's account age and friend dates as their own, and unlike YouTube there was no private scope to require instead.

Nine sources, nineteen scopes

GitHub, LinkedIn, Spotify, Instagram, YouTube, Amazon, Uber, DoorDash and Shop. All of a source's scopes go in one approval, so the whole manifest is nine approval trips rather than nineteen.

Timestamps in, content out

Every scope is requested for its dates. Captions, addresses, emails, track names, game titles and other people's names are discarded before anything is stored, and a test fails if any of them reach the store.

Signed, portable, and dated

Every non-provisional score carries an EIP-191 attestation anyone can verify offline against Patina's published address. No key, no OAuth, CORS open.

Attestations expire after thirty days, and the expiry sits inside the signed message rather than beside it, so checking freshness stays offline too. A revocation list would have fixed the same problem by making every verifier call Patina, which is exactly the dependency people integrate to avoid. Scores are signed with a key separate from the one holding the escrow balance, so a leak of one is not a leak of both.

Answerable about itself

/api/health does a real write-and-read against the database and a live read of the escrow balance, and returns 503 when the deployment cannot serve the next person, so ordinary uptime monitoring catches an outage without being taught what to look for. A password-gated /admin shows the funnel, the user numbers and how many more connections the balance will fund. Funnel counting is done server-side: no third-party script, no cookie banner, nothing that contradicts the privacy page.

Readable by agents

An MCP server exposes the score to AI agents mid-conversation, with a handle resolver that returns tenure without revealing identity.

---

The Score

Six components, weighted by how expensive each one is in time rather than in money.

Component	Max	What it measures
Age	30	The oldest date provable across every connected source. Full marks at twelve years.
Continuity	25	Distinct months the person was actually present for, as absolute count times coverage.
Corroboration	15	Independent sources agreeing on the date, each weighted by its own age.
Vouches	12	When other people chose to connect, not how many. Weighted by how long ago.
Depth	10	Things actually made, discounted when the volume arrived in one burst.
Breadth	8	Independent accounts backing each other up.

Age and the two time components are earned outright. Depth, Vouches and Breadth are gated behind them: an attacker can manufacture volume, friends and breadth in an afternoon, so those only count to the extent that real elapsed history backs them up. A floor of 15% stops a genuinely young person being flattened to zero for the crime of being nineteen.

Scored against reference profiles, asserted in score.test.ts:

Profile	Score
Eleven active years across four accounts	95
Ordinary: eleven years, two accounts, real gaps	71
Genuinely young: three years, two accounts, real	26
Account farm: 3,900 bought followers, 120 posts in one week	2

The gap between an ordinary person and a farm is 69 points. That gap is the product.

Honest limits

Stated on the site itself, not buried here:

• It cannot prove personhood. It reports how much real history it can see. That is evidence, not a certificate.
• A low score is not an accusation. Young, private or quiet accounts score low and are doing nothing wrong.
• Aged accounts can be bought. That turns a free attack into an expensive one whose price climbs with age. It does not make it impossible.
• Desktop proves session control, not ownership. A bought account arrives with its password. The bar moves from "knows a username" to "holds the credentials", which is a large jump and not the same word.
• The proof scope is a floor, not a guarantee. Requiring something private forces a real sign-in, and a real sign-in is still only a sign-in. What it rules out is somebody scoring a stranger's public profile, which is the failure that would make the whole number meaningless.

---

What is read, and what is kept

Every scope below is requested for its timestamps. This table is the trust argument, so it is on the privacy page in roughly this form too.

Scope	Kept	Discarded on arrival
linkedin.connections	dateConnected[], count	names, headlines, profile URLs
instagram.posts	taken_at[], count	captions, images, likes, the entire who_liked[] array
spotify.savedTracks	added_at[], total	track names, artists, albums
github.history	createdAt[], hashed repo, summed engagement	PR and issue titles and bodies
linkedin.experience / education	parsed date ranges	companies, job titles, schools, grades
uber.trips	requestTime[], count	pickup and dropoff addresses, fares, cities
amazon / doordash / shop orders	dates, count	items, merchants, totals, delivery addresses
youtube.profile	joinedDate, counts	email address
youtube.watchLater	nothing at all	everything; it is requested only as proof of a signed-in session

Timestamps collapse to month buckets before storage. The scorer only ever asks about months, so holding anything finer would be holding it for no reason.

---

Architecture

Connect flow