How Trixy turns a market into an edge
Every Polymarket trade indexed, every trader scored, 180+ signals per market — fused with the live web by an agentic loop that closes in seconds.
Trixy reconstructs every trade on Polymarket directly from the chain, scores every wallet by how right it has actually been, and computes 180+ signals on each market. Ask Trixy puts that engine behind a single box: paste a link, and a live agent loop on Cerebras plans a search, pulls the web, fuses it with the quants, and writes a cited verdict.
Everything you read on the next screen is grounded in data we own end to end — not a repackaged feed, not a model guessing from a headline. Here is the stack underneath.
We indexed all of it
Trixy reconstructs every trade on Polymarket directly from the chain — fills, splits, merges, redemptions and resolutions — from each market's inception. Not a rolling window, not a curated subset, and not dependent on anyone else's API.
The Trixy score, and the classifier beneath it
The crowd prices every opinion equally. Trixy weights them by demonstrated accuracy — segmented by category, price and time — so smart money means smart here, on this market, at this price.
Skill, from the only ground truth there is
Every wallet earns a rating from resolved outcomes. Each resolution updates it, so a score reflects how right a trader has actually been — not how loud or how large.
It is segmented: a wallet can be sharp on crypto and noise on politics, sharp near 50¢ and useless on longshots. We rate skill by category, price range and time.
Every market, understood before it's scored
A Gemma-based classifier sorts all 1.3M markets into a hierarchical taxonomy — sport, election, crypto level, macro print, and the long tail beneath each.
That taxonomy is what makes segmented skill possible — and what lets Ask Trixy pull the right siblings, resolution source and comparables the instant you paste a link.
A proprietary ranking system and a deep market classifier — the moat compounds with every resolution.
180+ live signals per market
From the indexed history we compute a dense feature set on every market — across eleven time scales — and refresh it as new trades land. These are the families underneath that number.
Smart-money positioning
Which side the skilled capital is on, weighted by demonstrated accuracy — not size. The crowd prices every opinion equally. We don't.
Skill-weighted flow
Buy and sell pressure re-weighted by trader rating, plus flow momentum and acceleration. Not just where money moved — whether good money moved it.
Price technicals
RSI, MACD, ATR and Bollinger bands per market across multiple time scales. The toolkit a desk runs, on prediction-market order flow.
Candlestick patterns
A purpose-built detector flags bullish and bearish reversal structures as they form, with the bias and how long ago each fired.
Participant mix
Holder count, concentration and trader-style breakdown. A market held by a thousand small wallets behaves nothing like one held by three whales.
Composite edge
Smart-money disagreement, skill, holder depth and room-to-move fused into one score, so the strongest setups surface on their own.
The web fills what on-chain can't see
The order book tells you what the market believes; it doesn't tell you why.
For each question Trixy searches multiple live sources in parallel, re-ranks the results, and brings the freshest relevant evidence back to the desk — the news, the catalyst, the context — and grounds every claim in it. Each thesis is cited, so you can follow any line back to where it came from.
Four steps, under five seconds
Every read is a full agent loop on Cerebras + Gemma. Each step is strict structured output, so the model fills our schema reliably instead of free-forming. On a slow model this is a batch job; at Cerebras speed it closes faster than you can read it — and you can refine it in place, like a conversation.
Plan
Gemma reads the market and generates the search queries that would actually resolve it.
Search
Multiple sources hit in parallel, then re-ranked so the strongest evidence rises to the top.
Fuse
Web evidence merged with Trixy's quant signals — the part a language model cannot invent.
Write
A cited, structured thesis: direction, bull and bear case, catalyst, EV and Kelly sizing.
Auditable by design
Every signal is auditable and every input traces back to an on-chain event or a cited source.
We don't publish a win rate we can't defend. Out-of-sample Brier validation is in progress, and the score will ship with its realized calibration. This is decision support, not financial advice — and we'd rather under-claim and be trusted than the reverse.
Build on the engine
The index, the Trixy score, the signals and the loop are all available over the API. Pull the intelligence into your bot, dashboard or strategy.