How scoring works

​

The idea

A TrustGate score is a read of behavior. Not of who someone says they are, not of what a contract claims to be, but of what an address has actually done onchain and what that pattern usually means. There are two lanes. One scores wallets. The other scores tokens, NFTs, and contracts. They look at different things, but they land in the same place: a tier, sometimes a number, a confidence level, and any flags worth raising. That shared output shape is what makes the signal easy to consume. The engines are not identical across chains. Arc uses the behavioral oracle plus local contract/NFT/temporal scorers. Sui TokenShield uses SuiScan and market signals on mainnet coins; Sui wallet standing is written to a testnet ScoreRegistry. Do not assume an Arc score and a Sui score are interchangeable numbers.

​

What we believe about trust

A few principles shape every read. Behavior over identity. An address earns its standing through what it does over time, not through a badge, a follower count, or a form it filled out. Current, not frozen. Scores are living. They move as behavior moves, so a read reflects where an address stands now, not where it stood at some snapshot months ago. Earned, not bought. The patterns that lift a score are ones that are expensive to fake at scale and cheap to produce honestly. That asymmetry is the whole game. Hard to game on purpose. A read does not hinge on one number that an attacker can farm. It weighs several kinds of behavior together, so gaming one dimension does not move the result much on its own.

​

What goes into a read

At a high level, a wallet read looks at the depth and shape of its history, how long it has been active, what it has built and interacted with, whether real value moves through it, and whether its activity looks human or automated. A token read looks at who holds it and how they got in, how it actually trades over time, wash and honeypot patterns, coordinated exits, and the standing of the wallet that deployed it. Temporal Token Shield signals (hold duration, exit ratio, wash pairs, sell-to-buy after distribution) now feed the token lane. Staking behavior can boost a deployer when the stake is real, aged, and not self-dealing. The Signals page goes through these by category. The Flags page covers the specific patterns that get raised on a read.

​

Two different outputs, two different maps

Do not treat every number you see as the same scale. HIGH_ELITE and ELITE are the same top behavioral tier. Arc oracle and widget.js still emit HIGH_ELITE. Discovery batch and widget-discovery.js normalize that to ELITE. Treat them as aliases. Onchain, HIGH_ELITE shares the HIGH payment band (>= 75). The elite distinction is a UI and API label, not a fourth contract enum.

​

Scoring version

Every hardened oracle payload carries scoringVersion. Oracle and batch use v1.0. Attestations label the lane and environment as well: testnet-wallet-v1.0 or testnet-token-v1.0 (or mainnet-… when that environment is set). Pin this in any gate. A recalibration that changes who passes a band is a breaking change and will bump the version. Protocols that ignore it will silently drift.

​

Interpretable, not reproducible

Here is the line, and it is deliberate. We tell you what we look at and what the output means. We do not publish the exact recipe, the precise weights, the timing windows, or the cutoffs. That is on purpose, because TrustGate is adversarial infrastructure. Users read these docs. Builders read these docs. And so do the people trying to beat the system. If the exact thresholds were public, they would stop being a trust signal and start being a checklist to game. This is the same posture mature anti-abuse systems take. Categories and meaning are public. The quantitative internals are not. So you can build on the signal with full confidence in what it represents, without us handing anyone a map to fake it.

​

What TrustGate is not

A second, optional lens exists on the TrustGate site only: Knowledge Graph trust from Intuition. It answers a different question (what others have attested) and is not mixed into widget, batch, oracle, or gating responses. Integrators never call it. A few more things scoring deliberately is not. It uses no off-chain data, no personal data, and no KYC. There is no manual review deciding who is good and who is not. And it is not a credit bureau, a price oracle, or a safety guarantee.

​

Go deeper