01 — THE RECORD

Every prediction
logged. Nothing
forgotten.
Data never lies.

When you send /research, Polytragent makes 8–12 predictions. Each one is instantly logged to predictions.json with: market ID, AI forecast %, Polymarket price at prediction time, Manifold price, confidence level, recommendation, reasoning snippet, and timestamp. This is the permanent record. Every trade is tracked.

Immutable Log

Every prediction goes to predictions.json. Cannot be deleted or edited. This is the source of truth for learning and accountability.

Full Context Stored

Not just the prediction—the market state at that moment. Price, volume, liquidity. This context helps the AI learn why it was right or wrong.

Reasoning Snippet

A 1–2 sentence summary of why the AI made this specific prediction. Invaluable for post-mortem analysis.

polytragent — prediction log
polytragent — background resolver
02 — THE LOOP

Every 6 hours:
check, resolve,
calculate,
improve.

The background resolver is a cron job that runs every 6 hours. It queries Polymarket's API for all resolved markets. For each resolved market, it finds matching predictions in predictions.json. It records the outcome: CORRECT or WRONG. It calculates hypothetical P&L. It updates the AI's accuracy metrics. The loop never stops. The AI gets smarter every day.

01

Background Resolver Trigger

Cron job runs every 6 hours (00:00, 06:00, 12:00, 18:00 UTC). Non-blocking—doesn't interrupt user queries.

02

Query Polymarket API

GET /markets?status=resolved. Fetch all newly resolved markets and their outcomes (YES or NO).

03

Match to Predictions

For each resolved market, search predictions.json for matching market_id. If found, compare AI's forecast to outcome.

04

Mark CORRECT or WRONG

If AI said YES and market resolved YES = CORRECT. If opposite = WRONG. No partial credit.

05

Calculate Hypothetical P&L

If correct: add (final_price - prediction_price) × confidence_weight to P&L. If wrong: subtract loss.

06

Update Metrics

Recalculate overall win rate, topic-specific accuracy, calibration biases. Store in predictions.json metadata.

07

AI Sees Updated Stats

Next /research command: Claude Sonnet sees new metrics and adjusts future predictions accordingly.

Prediction tracking
FAQ

What happens if a market is still open when the resolver runs?
The resolver only records outcomes for markets with status=resolved. Open markets are skipped. The resolver will pick them up in a future run once they resolve.
How is hypothetical P&L calculated?
If you had traded $1 at the prediction price, what would your profit/loss be? P&L = (final_price - prediction_price) × $1. This assumes you held to resolution.
Can I see my prediction history?
Not yet. Future update: /history command will display all past predictions, outcomes, and running P&L. This will be viewable via Telegram dashboard.
What if I make a wrong prediction?
It's logged as WRONG. The loss is recorded. The market and topic are flagged for calibration adjustment. Mistakes are how the AI learns.
How often is the win rate updated?
Every 6 hours when the background resolver runs. Win rate is also calculated on-demand when you send /stats command.
Can predictions.json be deleted?
No. It's immutable. Deletion would break the learning system. This is intentional—it ensures accountability and enables long-term improvement.

Track every
prediction.
Never stop
improving.

Polytragent logs everything and improves continuously. Your winning system starts with perfect tracking.

Start tracking — $79.99/mo →
polytragent — prediction lifecycle
Logged instantly on predictionResolved every 6 hoursFeeds back to improve AI

Risk Disclaimer: Prediction market trading involves significant financial risk. AI-powered research does not guarantee profits. Past performance is not indicative of future results. Polytragent is a research tool, not a financial advisor. Only trade capital you can afford to lose. This website does not constitute financial advice.