How Contrarian works

Contrarian is a decision firewall. Instead of confirming what you want to do onchain, it actively argues against it: every transaction intent is run through four adversarial review lenses, scored for concern, and surfaced as a threat dashboard. You then choose to Execute Anyway or Cancel.

The premise: most onchain losses don't come from bad infrastructure, they come from unchallenged decisions. Wallets ask “confirm?” — Contrarian asks “are you sure, and here's why you might not be.”

The fastest way in is the hosted dashboard — no wallet required to run a review:

Open the dashboard →

To run it locally instead:

npm install
npm run dev   # http://localhost:3000

Then open the dashboard and describe an action in plain English, for example swap 100 USDC to SOL or stake 50 SOL with Marinade. Connecting a wallet (injected connector, Arbitrum) is optional and only affects the header — reviews work without it.

Every intent moves through a small state machine. If the input can't be parsed into a known action, the review ends in an error state and you can rephrase.

intent ──► analyzing ──► complete ──► executing ──► executed
   │                       │
   ▼ unparseable           ▼ cancel
 error                  logged as CANCELLED

• Analyzing — the four lenses run in parallel; results arrive together after a short delay.
• Complete — scores, counter-arguments, and a verdict are on screen. The review waits for your decision.
• Executing → Executed — choosing Execute Anyway simulates the transaction and logs it to the activity feed with a mock hash.
• Cancelled — the review leaves the transcript and is logged as CANCELLED. Cancelling is recorded on purpose: avoided mistakes are wins worth counting.

Each lens answers one question and scores its level of concern from 0 to 100 — higher means more reason to pause.

Scores map to three severity bands:

The verdict is derived from the average of the four scores: above 55 the review reads RECONSIDER; at or below it reads PROCEED WITH CAUTION. There is no third verdict and no hard block — the final decision is always yours.

The intent parser recognizes eight onchain action types from plain English. Anything it can't parse returns an error state with a prompt to rephrase.

• Wallet connection uses wagmi + viem with the injected connector (MetaMask, Rabby, and similar).
• Supported chains are Arbitrum One and Arbitrum Sepolia; the header offers a one-click switch if you're elsewhere.
• Execution is simulated. Execute Anyway produces a mock transaction hash — no calldata is built and nothing is signed or broadcast.
• Review history lives in memory only and resets on reload.

• Analysis is a deterministic local mock — phrase pools and weighted random scores, not a real risk assessment.
• No persistence: history, scores, and verdicts vanish on reload.
• The intent parser is regex-based and intentionally narrow; it covers the eight actions above and nothing else.
• Nothing in this product is financial advice.

On the roadmap: a real four-persona LLM call with structured output, transaction construction with pre-sign simulation, and persistent review history.