Wallet Health

The overall score is a fixed weighted blend of five pure-math components. The weights — and what each measures — are:

Diversification uses a Herfindahl concentration index and flags your largest position against a recommended max (50% stablecoin / 40% blue-chip / 25% anything else). Yield efficiency quantifies your idle capital and the annual yield you're leaving on the table.

Every score is computed by pure math from normalized portfolio data — the LLM cannot recalculate or override a score. It receives the deterministic results as context and only writes the narrative: a headline, an overview, actionable opportunities, and risk warnings.

This separation means the numbers are auditable and stable. The prose may shift between runs, but the scores never move unless the underlying portfolio did. A score breakdown shows the signed impact of each factor (e.g. "+8 Good diversification", "-10 42% in PEPE").

Raw balances aren't enough to judge a portfolio. Each meaningful asset runs through a pluggable enrichment pipeline:

An asset's conviction risk tier feeds the risk score; its safety level feeds the security score. New enrichment modules can be added without touching the scoring engine — they simply plug in upstream.

A wallet report is only as trustworthy as the data behind it. Arvio tracks a data-completeness figure (how many data sources resolved) and an overall confidence that blends completeness with enrichment coverage.

When data is thin, confidence is lowered and the narrative names the gap honestly — rather than presenting a confident score built on missing information.

The same smart-money tracking that powers token reports feeds wallet health: when profitable early wallets are accumulating or distributing an asset you hold, that signal informs the report's risk read. You can also follow wallets to track them over time — watching a known smart-money address the way you'd watch a token.

A token's report card carries its own smart-money flow; a wallet's health report folds that same signal into how risky your holdings are.

The synthesis layer turns the deterministic scores into specific, actionable output: opportunities ("Move 14,200 idle USDC to Aave — est. +$640/yr") ranked by estimated impact, and risk warnings (concentration, scam exposure, protocol risk) with severity and a recommended action.

Both are grounded in the real numbers from the scoring engine — the LLM is told to quote the figures it was given, never to invent amounts.

Every Wallet Health Report is logged as a directional prediction: a score of 60+ is bullish, below 40 is bearish. The "price" logged is the report's frozen held basket — the sum of every holding's value at report time. Seven days later the same fixed basket is repriced at current market prices, and the call hits if it moved the predicted way.

This is flow-immune by design: deposits or withdrawals after the report can't fake the result, because only the frozen holdings are repriced — never a live balance that flows would move. Resolved calls fit a wallet_health calibration model, so the portfolio engine earns a measured track record alongside the token engines.