Prexora Whitepaper

A disciplined, transparent AI agent for weather prediction markets. This document explains the problem — backed by data, not opinion — why weather is the right battleground, exactly how the engine forms and risk-manages a position, the fifty-plus subsystems behind it, and how the $PRIX token fits in.

Version 1.1 · 23 May 2026 · Living document

Executive Summary

Prediction markets are the most honest financial instrument retail has ever had: a market that settles to objective truth instead of sentiment. Yet the data is brutal — the large majority of participants lose money, and the profits concentrate in a tiny, sophisticated, increasingly automated minority. The reason is not that the markets are rigged. It is that winning requires speed, quantitative calibration, and emotional discipline sustained across thousands of small decisions — three things humans are structurally bad at and software is good at.

Prexora is an autonomous agent built to do exactly that, on one tractable corner of the market: daily temperature markets. It pulls a three-model weather forecast ensemble, converts it into a probability for every Polymarket bracket, bets only where its estimate diverges materially from the market price, sizes each position with half-Kelly against a hard daily budget, defends capital with multiple independent risk guards, and settles every bet against the exact Weather Underground station reading Polymarket resolves with. It does all of this in public — signing signals, anchoring a daily integrity hash on-chain, and reflecting on every resolved bet, wins and losses alike. The $PRIX token turns this working product into an access and utility layer.

The Five Pillars

Strip everything back and the engine stands on five load-bearing ideas. Break any one and it stops being profitable.

That is the whole machine: align, price the edge, demand consensus, learn, and stay honest — repeated thousands of times a day across our cities.

Part I — The Discipline Gap

This is not a vibe. It is what the trade data shows.

1.1 Markets that settle in truth

A prediction market pays out based on whether a real-world event occurs. Unlike a perpetual future or a memecoin, it does not depend on the next buyer’s sentiment — it depends on reality. When the event resolves, truth is paid and noise is wiped out. That makes prediction markets uniquely suited to a quantitative operator: estimate the true probability of an event better than the market prices it, and you have a durable, repeatable edge.

1.2 The latency and emotion tax

The catch is that “better than the market” is a race. Information arrives, prices adjust, and the gap between a fair price and a stale price closes in seconds on liquid markets. A human reading a forecast, opening an app and placing an order is structurally late. Worse, humans size emotionally — bigger after wins, frozen after losses — which turns even a real edge into a losing equity curve. The market does not tax intelligence; it taxes slowness and indiscipline.

1.3 Why the winners are increasingly machines

The same research that shows most people losing shows who wins: a thin layer of disciplined, fast, quantitative participants — many of them automated. That is the gap Prexora fills. We are not promising to beat the elite at everything. We are putting elite-grade discipline onto the one market where the true probability is genuinely estimable from public physics: the weather.

Part II — Why Weather

2.1 A physically grounded market

Weather is the rare prediction-market category governed by physics, not narrative. A daily high temperature is the output of atmospheric dynamics that national agencies model continuously and publish openly. The “true probability” of a temperature bracket is therefore genuinely estimable from public data — a real signal to extract, not a popularity contest to front-run.

2.2 Structural inefficiency and recurring liquidity

Weather brackets are underserved by quant players relative to politics or sports, so mispricings persist longer. And because every covered city produces a fresh market every single day, the opportunity set recurs — thousands of independent, repeatable bets rather than a handful of one-off events. Recurrence is what lets a small per-bet edge compound.

2.3 Ground truth you can verify

Weather resolves against published observations anyone can check — there is no ambiguous oracle to dispute. Critically, Polymarket settles each weather market on a specific Weather Underground (WU) airport station, so Prexora’s number-one job is not to predict “the weather” in the abstract but to predict that exact station’s recorded high. Every data source and gate exists to align the agent’s estimate with the WU number that will actually settle the market. And because that number is public, a signed, on-chain-anchored prediction can be checked against reality by a third party with zero trust in us.

2.4 The markets we cover

Prexora trades the daily highest-temperature markets for 26 cities across the world, each resolved against a specific weather station, in °F (US) or °C (international):

Each city is graded easy / medium / hard by how predictable its climate is, and the engine weights its conviction accordingly. Coverage expands as new Polymarket weather markets come online.

Part III — The Data Layer

Prexora is only as good as its inputs, so it triangulates across independent, free, no-auth public sources with automatic fallback at every leg.

Part IV — How a Pick Is Made

The pick logic is described here in full at the structural level. The exact tuned thresholds, penalty tables and calibration coefficients — the parts that actually constitute the edge — are kept private.

4.1 From three forecasts to one distribution

For each city/date the engine collects the daily-max forecast from all three NWP models and summarises them into a mean (μ) and a spread (σ), where σ is the larger of the model disagreement and a baseline horizon uncertainty. The result is a Normal distribution describing the day’s likely outcome — tight when the models agree, wide when they don’t.

4.2 From distribution to bracket probability

Each Polymarket bracket is an interval. The probability the outcome lands in [a, b] is the area of the distribution between those bounds: P = Φ((b−μ)/σ) − Φ((a−μ)/σ), computed with the standard normal CDF (via erf, no heavy dependencies). The parser handles open intervals correctly — an “above X” ceiling runs to +∞, a “below X” floor to −∞ — so probabilities stay consistent across ceiling, floor, range and single-value markets.

4.3 Edge, gates and strength

The model probability minus the market price is the edge. Candidates then pass through seven independent quality gates — an edge floor, a confidence floor, an entry-price band, a 24h-volume floor, a minimum-model-count, a directional block on already-expensive contracts, and a sigma-sanity check. A signal that fails any gate is dropped, not weakened. Survivors are tiered by conviction:

4.4 Sizing

Funded picks are sized with a half-Kelly rule scaled by strength, bounded by a per-bet minimum and maximum, and capped by the daily budget. Conviction buys size — within hard limits that the agent physically cannot exceed.

The Intelligence Layers — Demanding Consensus

A single forecast is a guess. Prexora only acts when several independent kinds of evidence agree, each contributing a vote that the consensus gate can boost or block. These layers were forged in Prexora’s research engine and the proven ones are productized in the shipped agent.

The forecast ensemble & real-time METAR

The three NWP models (ECMWF, GFS, ICON) provide the probability distribution; the WU forecast is the resolver-aligned anchor; and hourly METAR observations add a same-day reality check — if the running high has already cleared a bracket (or can no longer reach it), that overrides the forecast.

The AI swarm — a second opinion that is not a weather model

Surviving candidates are sent to a 12-agent AI swarm that debates each pick: every agent independently estimates a win probability, and their averaged verdict plus an agreement score is blended into the decision. Because it is not a weather model, it catches failure modes the models share. Its influence is auto-learned — it earns weight only if it proves more accurate over time, and fades if it doesn’t.

Whale tracking & copy-trading

The engine watches a set of top weather-market wallets and the public leaderboard. When sharp money is on the same side as a pick, the score gets a bonus; when it is against, a penalty. When several top traders converge on the same city and direction, that convergence becomes its own signal. A trader-intelligence layer goes further, profiling how winners win — their zones, timing and sizing — and feeds that back as a heavily-weighted learning dimension.

The multi-brain & the learning engine

The engine is split into crash-independent specialized workers — a forecast brain, a market brain and an intel brain — coordinated by a pure-math master that scores, gates and sizes without making a single network call, so the decision core is nearly unbreakable. Over the top sits a multi-dimensional learning engine that turns bet history into a per-pick multiplier across dimensions like per-city profile, edge honesty, model trust, time-of-day quality, consensus pattern, bracket memory and streak momentum.

Part V — How Risk Is Minimized

A real edge is destroyed by a single undisciplined drawdown. Prexora runs several independent risk guards, each a pure function of the resolved-outcome history, that sit between a candidate signal and a placed bet. Any one of them can stop a trade.

5.1 Anti-revenge city pause

If a city loses two or more bets in a single UTC day, the agent stops publishing in that city until midnight rolls over. Revenge betting — pressing harder after losses — is the single most common way retail blows up. The agent simply refuses to do it.

5.2 Global drawdown stop-loss

If the trailing seven days of resolved bets show losses worse than −30% of staked capital, all publishing halts until the rolling window recovers. Hold cash; do not press.

5.3 Per-detector auto-learning

Every (city, side, strength) combination is a “detector.” The agent tracks each detector’s recent win rate and, when one is running below par, multiplies its reported edge by a penalty below one — quietly pushing soft signals from that detector below the publish threshold. Detectors that keep losing effectively pause themselves, with no human intervention.

5.4 Event loss limiter, budget cap & live exit

A single event can lose at most a few brackets before that event is paused for the day; the daily budget is a hard ceiling on total deployment; and an open position is auto-exited if a fresh forecast run pushes our side’s probability below threshold for consecutive cycles. Together these bound the worst case from every angle — per city, per event, per day, and per open trade.

Part VI — The Brain: A Learning Loop That Earns Its Keep

A static model is a depreciating asset. Prexora is built to compound its edge from its own resolved bets — safely, with every change graded before it can move real money.

6.1 Per-bet reflection

Every resolved bet produces a written reflection — what was expected, what happened, what it implies — feeding the public lesson stream and the higher-order learning layers. The agent also sets and grades its own weekly performance targets.

6.2 Strategy lifecycle: shadow → graded → active

Candidate trading rules are proposed as broad, cross-city patterns and enter in shadow mode, evaluated against live decisions without risking capital. Promotion to active requires enough graded evidence and a positive, statistically meaningful effect, and is gated behind explicit approval. Rules that stay dead are auto-retired. The principle: the AI may propose, but only graded, rule-based logic ever moves money.

6.3 Calibration & bias correction

The engine measures systematic per-city forecast bias and computes corrections, validated in shadow against a faithful replay of historical bets before they are allowed to influence live sizing.

6.4 The impact ledger

Learning is only worthwhile if it earns. The impact ledger correlates each learning action with the bet-outcome trend before and after it, and against the weekly targets, for an honest read on whether learning is paying off — labelled correlational, not causal.

Part VII — Self-Healing Infrastructure

An autonomous trader is only as good as its uptime. Prexora is built so a human never has to log in at 3am to clear a stuck state or restart after an API hiccup.

• Watchdog + circuit breaker restarts the engine if it dies, without flapping.
• Per-source health probes track market data, weather data and the language layer; the cockpit shows exactly which source is degraded and why.
• Automatic fallbacks: WU↔METAR for observations, NOAA/NWS when Open-Meteo is rate-limited, alternate market endpoints, and template narration if the language model is down.
• Stuck-state recovery retries pending verifications older than six hours and voids overdue open signals.
• Database integrity checks run on a cadence and halt strategy promotion on any drift.
• Crash-loop detection pages a human on repeated cycle failures, and a 02:00 UTC self-test cycle catches silent problems.

Part VIII — Transparency Architecture

Transparency is built into how signals are produced, not bolted on. Three independent legs make the track record checkable by anyone:

On top of these, the public read-only API serves signals, recent activity, lessons, on-chain commits and the track record; performance is reported on funded signals only.

Part IX — The System, End to End

Prexora is not a script — it is a layered system of well over fifty cooperating components. The brain-control map below shows how they fit together.

Part X — The Decorrelation Thesis

Weather outcomes are uncorrelated with crypto market direction. A heat dome over Madrid does not care whether Bitcoin is up or down. For a token holder that matters: Prexora’s edge — and therefore its case — does not depend on a bull market. In a category where most “AI trading” products are implicitly levered long on crypto beta, a strategy whose returns come from physics is a genuinely different risk profile.

Part XI — Token Design: $PRIX

$PRIX is the access and utility token for the Prexora agent, launching on Base via the OpenServ launchpad. Total supply is a fixed 1,000,000,000 with 500,000,000 initially circulating; allocation is 50% liquidity, 25% treasury, 20% team, 5% $SERV staking. Day-one utility is access — holding or staking PRIX unlocks Member and Pro tiers (pre-publication signals, full rationale, custom alerts, discounted subscriptions, reduced auto-trade fees). The 18-month team performance cliff ties the builders to long-term results. Staking-for-curation and deeper value-flow follow in a later phase. Full detail is in Tokenomics and Access & Utility.

Part XII — Risks & Honest Limitations

Prexora is a real product with real risk. Edges can compress as more players enter weather markets. Forecast models carry biases the engine corrects for but cannot eliminate. Copy trades never guarantee identical fills, timing or returns. The token is an access and utility instrument, not a claim on profits or investment advice. These are covered in full in Risks & Disclaimers.

Conclusion

The data is unambiguous: most people lose prediction markets, and the winners are increasingly disciplined machines. Prexora packages that discipline onto the one market where truth is physics and verification is trivial — and shows every move it makes. The product works today. $PRIX turns it into something the public can hold a stake in.

Sources

Figures on prediction-market profitability are drawn from 2026 reporting and studies, including CoinDesk, CNBC, Gambling Insider and analyses summarised at predictionmarkets.org and casino.org. Specific links are provided alongside this document.