How we reach a read.

Every claim is labelled FACT (retrieved from a named provider with a timestamp), INFERENCE (a reading built on those facts), or UNKNOWN (evidence is missing or providers disagree).

Where a value cannot be retrieved we print Unavailable. We never substitute zero, and mock values are never presented as live: demo fixtures carry a DEMO DATA badge everywhere they appear.

When two providers disagree, both readings and both retrieval times are shown and the confidence level drops.

• Top-holder concentration excluding AMM pools and burn addresses.
• Mint and freeze authority state.
• Liquidity depth and estimated slippage for example order sizes.
• Creator / deployer history where it can be sourced.
• Funding-source clustering and bundled supply where reliably detectable.
• Volume-to-liquidity ratios and unique-buyer trends as wash-trading tells.

Sniper and bundle detection is heuristic. When detection is not reliable for a launch type we say so rather than guessing.

A wallet is never called “smart” from win rate alone. The score is context, not endorsement, and it is withheld entirely when sample quality is too low — Insufficient data instead of a number.

• Jupiter

  Token search and metadata, pricing, trading metrics.

  Limitations: Routing-derived prices can differ from pool-derived prices on thin pairs.

• Birdeye

  Market overview, OHLCV, trades, liquidity, holder distribution, security context.

  Limitations: Endpoint availability depends on the subscribed plan.

• DEX Screener

  Pairs, token profiles, boosts and community takeovers, pair context.

  Limitations: Documented rate limits apply; boosts are paid placements, not quality signals.

• Helius

  RPC, parsed transactions and events, wallet history, webhooks and streams.

  Limitations: Historical depth and parser coverage vary by program.

• X API / licensed social provider

  Public social signals and narrative velocity.

  Limitations: Only public, licensed data. We never scrape authenticated surfaces.

• TRENCHGPT knowledge base

  Terminology, launchpad mechanics, community context, sourced trader profiles.

  Limitations: Curated and reviewed; every entry carries a review date and source.

• On-chain data shows behaviour, not intent.
• Past realized performance does not predict future results.
• Social data is partial and can be manufactured.
• Provider outages and rate limits produce partial answers, which we label.
• No answer here is a recommendation to buy or sell anything.

TRENCHGPT is provided for research and education. You are responsible for your own decisions. Provider data is used under each provider's terms; we do not scrape authenticated sites or redistribute licensed feeds.

Accounts are optional and use magic-link email sign-in. We store your conversations, watchlists, preferences, and answer feedback, readable only by you under row-level security. Public wallet addresses you look up are treated as public data; we do not attempt to identify people behind them.

All third-party API calls run server-side. No provider key ever reaches the browser. TRENCHGPT will never request a seed phrase or private key, and v1 executes no trades.