The Engine & Mission Control
Under the surfaces is a single always-on agent built from well over fifty cooperating components. This page describes its architecture, how it keeps itself alive, how it learns, and how Mission Control lets you watch it think.
Architecture: brain control
Prexora separates duties strictly. The modelling core never invents inputs; the language layer never invents conclusions — it only narrates what the rule-based core decides. A self-healing wrapper keeps every layer alive.
| Layer | Responsibility |
|---|---|
| Data layer | Weather Underground (the resolver), METAR real-time cross-check, ECMWF/GFS/ICON forecast ensemble via Open-Meteo, NOAA/NWS fallback, Polymarket Gamma markets + resolution. |
| Model layer | Ensemble → μ,σ → Normal-CDF bracket probability → edge → seven gates → strength tier → half-Kelly sizing. |
| Risk layer | Anti-revenge pause, drawdown stop-loss, detector penalties, event limiter, budget cap, live forecast-revision exit. |
| Execution layer | Publish, place, resolve; funded vs display; one-click copy and autonomous delegated trading. |
| Learning core (the brain) | Per-bet reflection, strategy lifecycle, calibration/bias, impact ledger. |
| Narration layer | A language model that rewrites decisions into readable rationale — a narrator, not a decision-maker, with template fallback. |
| Transparency layer | ed25519 signed feed, daily Solana anchor, public read-only API. |
| Self-healing wrapper | Watchdog, health probes, automatic fallbacks, self-test. |
Design principle. The AI may propose — new rationale, candidate strategies, calibration shifts — but only graded, rule-based logic ever moves real money.
The intelligence layers
The model is one voice among several. Prexora blends independent layers and only acts when they agree — the consensus principle.
| Layer | What it adds |
|---|---|
| Forecast ensemble + WU | ECMWF/GFS/ICON give the distribution; the Weather Underground forecast anchors it to the resolver. |
| Real-time METAR | Hourly observations confirm whether a bracket is already hit or now impossible — a same-day override. |
| AI swarm | A 12-agent swarm debates each pick and contributes a consensus probability + agreement score — a non-weather second opinion whose weight is auto-learned. |
| Whale & copy-trade signals | Alignment with top weather wallets and leaderboard convergence boosts a pick; opposition penalizes it. |
| Multi-brain workers | Forecast, market and intel brains run crash-independently; a pure-math master scores, gates and sizes with zero network calls. |
| Learning engine | A multi-dimensional multiplier from bet history — city profile, edge honesty, model trust, timing, consensus pattern, bracket memory, streak momentum, and learning from top traders. |
Self-healing operations
The watchdog restarts the engine if it dies. This layer handles the harder case: the engine is up but a dependency is misbehaving or stuck state has built up. On every sweep it automatically:
- Re-verifies Polymarket verifications older than six hours, surfacing any disagreement with our recorded outcome.
- Voids stuck open signals whose expiry is well past with no resolution, and logs the recovery.
- Probes the language model and, if it has been failing for over 30 minutes, flips to template narration and shows the cockpit why.
- Runs a database integrity check on a cadence and halts strategy promotion on any drift.
- Detects crash-loops (repeated cycle failures) and pages a human rather than silently flapping.
- Runs a 02:00 UTC self-test to catch silent failures before they cost a trade.
The learning loop
Strategy lifecycle
Candidate rules are proposed as broad, cross-city patterns and run in shadow mode, evaluated against live decisions without risking capital. Promotion to active requires graded evidence plus a positive, statistically meaningful effect, and is gated behind explicit approval. Dead rules auto-retire.
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 influence live sizing.
Impact ledger
Each learning action is correlated with the bet-outcome trend before and after it, and against the agent’s self-set weekly targets — an honest, correlational read on whether learning is paying off.
Mission Control
Mission Control is the public, real-time window onto the engine: what it is scanning, the decisions it is making, the bets it is placing, source-health indicators and the live decision flow — from the same state that drives the trades. It is the difference between “trust us” and “watch us.”
The fifty-plus components, by layer
For the technically curious, the real subsystems behind Prexora:
| Layer | Components |
|---|---|
| Data | Weather Underground fetcher (resolver), METAR tracker, Open-Meteo client, ECMWF IFS, GFS Seamless, ICON Seamless, NOAA/NWS fallback, rate-limiter + weather cache, Polymarket Gamma client, event-slug builder, bracket parser, resolution fetcher. |
| Model | Forecast summariser (μ,σ), Normal-CDF interval probability, edge calculator, strength tiering, seven quality gates (G1–G7), half-Kelly sizer, strength multipliers. |
| Intelligence | WU resolution-source gate, consensus gate, METAR same-day boost, 12-agent AI swarm, whale tracker, copy-trade leaderboard signal, trader-intelligence profiler, forecast/market/intel/master multi-brain. |
| Risk | Anti-revenge city pause, global drawdown stop-loss, per-detector auto-penalty, per-event loss limiter, daily budget cap, forecast-revision live exit, EV gate, Kelly-sensitivity guard, zone (widow-maker) filter, news veto. |
| Learning | Per-bet reflection, daily journal/lessons, weekly self-targets, strategy proposer, shadow→graded→active lifecycle, dead-rule pruner, σ calibration, mean-bias correction, impact ledger. |
| Narration | LLM narrator (OpenAI/Anthropic auto-detect), template fallback. |
| Transparency | ed25519 signer, daily Solana memo anchor, funded/display separation, public track-record API. |
| Self-healing | Watchdog + circuit breaker, source-health probes, WU↔METAR fallback, NOAA rate-limit fallback, stale-cache last resort, stuck-signal voider, LLM health probe, DB integrity check, crash-loop detector, daily self-test. |
| Surfaces | Live signals board, Mission Control cockpit, lesson stream, Telegram dispatcher, one-click copy, autonomous execution, track-record dashboard, public API. |