One market. Two different questions.
Movement from now — and the chance of finishing above a fixed round reference.
Self improvement WAITING
| Horizon | Baseline CRPS | Champion CRPS | Challenger CRPS | Champion skill | Challenger skill | Vs champion |
|---|
Simple experimental promotion policy. Passing does not establish statistical significance or trading eligibility.
Deployed champion vs baseline
Actual live time blocks. The baseline follows current volatility. This curve can fall; selected winners from different market windows are not a cumulative improvement percentage.
Model generations
| Cycle | Training history | Available / loaded rows | Decision | Vs champion |
|---|
Each row retains its own evaluation window. These selection-window results are not an untouched performance benchmark.
Automatic learning controls — owner token required
Enter the same owner token in the round/model controls below. Pause stops new training requests, not collection or an already-running evaluation.
Rust RTDS price inputs · official Chainlink TWAP
| Source | Price · USD | Freshness | Observation age | Delivery bound |
|---|
Waiting for the Rust stream.
Forecast prices remain Binance BTC/USDT microprice. TWAP60 is an optional training input from the official RTDS E18 payload. The untrained baseline does not use it. TWAP30 is disabled.
Current round WAITING
Probability at the fixed round end
Generic Binance receive-clock rounds. No Polymarket/Kalshi settlement mapping is assumed.
The price champion forecasts 3, 5, 10 and 30 seconds ahead. With more than 30 seconds remaining, this card uses a separate terminal model if one is loaded, otherwise the volatility baseline. During the final 30 seconds it uses the selected price model. Automatic Model A training currently promotes price models only.
Round probability history
Price now vs the fixed reference
Short-term microstructure · a different target
These estimates concern t+3/5/10/30 seconds. “Above reference” here is NOT the round outcome unless that horizon is exactly the remaining time.
Feed integrity
| Venue | Book | Quote age | Depth age | Gaps |
|---|
One collector; every strategy consumes the same canonical sample.
Model vs baseline · microstructure
| Horizon | Rows | CRPS | Baseline | Skill |
|---|
Lower CRPS is better. Baseline losses use the same timestamp and realized target as the model. Correlated rows are not independent experiments.
Prospective round evaluation
| Strategy | Settled rounds | Point-score rows | Bounded rows | Brier | Baseline | Skill | Log loss | ECE |
|---|
Scored only after the exact target quote and a live-feed witness arrive. Each round gets equal total weight. Use the compare API for challengers started at different times. Overlapping 5m/15m rounds still share information. Tail-only predictions retain score bounds in the API.
Round & model controls — owner token required
Token stays in this page’s memory only. Read it from runtime\v04\control.token (or the token_file in your configuration). Never share it or put it in a URL.
An existing reference is never edited. A manual reference creates a new round and is labeled manual. Generic rounds are not exchange-listed contracts.