0 raw events captured by the current process; durable files from earlier restarts remain in the evidence window.
PHASE 1 OF 5 · LIVE EVIDENCE COLLECTION
v6 is collecting data.
It is not predicting or trading yet.
We are building a trustworthy dataset before judging a new model. This prevents another apparently impressive model from failing when tested live.
Predict’s official winners and settlement start/end prices are being saved.
Raw capture and official outcomes continue; live settlement-distance features must wait.
Current checkpoint
Starting recorderComplete a stable 48-hour evidence window
Next step after this window: freeze the capture, run the gap/replay audit, then calculate whether any baseline beats Predict’s own market price.
48 hours integrity and replay checkpoint7+ UTC days earliest shadow-model evidence gate
ONE EXTERNAL ACTIONAsk Chainlink to authorize Predict’s BTCUSDT feed
The credentials are installed and authenticated. The account is not entitled to this exact feed ID yet. This does not stop the recorder, but it remains required before a model can be promoted.
0x0003ef42…f769d37WHAT COMES NEXT
One gated step at a time
- 1Collect synchronized evidenceRUNNING NOW
Raw books, trades, contract state and official outcomes for 48 hours.
- 2Freeze, audit and replayNEXT
Prove there are no important gaps and that the same data produces the same result.
- 3Compile dataset and train challengersLATER
Build fixed-time Predict rows, then test market-anchored models walk-forward.
- 4Paper executionLATER
Include price, fees, slippage, latency and realistic fills—still with no money at risk.
- 5Promotion decisionLATER
Only consider a tiny canary if the frozen model beats the market after costs.
Technical details For debugging and review
Replay audit
- State
- NOT RUN YET
- Events
- 0
- Files
- 0
Runs after the evidence window is frozen.
Model comparison
- State
- WAITING FOR DATA
- Brier lift
- —
- Log-loss lift
- —
No model is promoted without beating the executable market quote.