Reference results

Reference results (synthetic data)

These numbers come from the bundled synthetic replay sample. They demonstrate that the accounting is correct. They do not show an edge, and the sample is far too small to infer one.

D4 paper day (reports/examples/paper-day-example.md, regenerated by scripts/run_paper_day.py):

Metric	Value
Input rows / kept after QA	8,640 / 8,640 (0 duplicates, crossed, gaps or out-of-order rows)
Span	2.999 days
Start → final equity (marked to bid)	10,000 → 9,996.3482 USDT
Net PnL after fees and slippage	−3.6518 USDT
Fees paid	3.5936 USDT (assumed 10 bps fee rate, unverified)
Average slippage	0.50 bps
Fills / round trips / winners	24 / 12 / 2
Tickets settled / rejected (no quorum) / vetoed	12 / 74 / 0
Reconciliation drift	0
LLM tokens / LLM cost in the engine	0 / $0.00

Historical in-sample grid example (superseded; retained only as a historical artifact, reports/examples/backtests-example.json):

EMA fast/slow	Regime slope	Net PnL after costs (USDT)	Fills
20/60	0.0001	−3.6518	24
20/60	0.0005	0.0000	0
10/30	0.0001	−10.9396	60
10/30	0.0005	−5.7180	32
30/90	0.0001	−1.9840	8
50/200	0.0001	−0.5241	22

The current S15 evaluator instead runs all six fixed variants across independent chronological 60/20/20 train, validation and holdout windows. Feature warm-up uses only preceding rows, each fold starts with independent paper equity, and every fold includes fees, mark-to-market open positions, QA counts and reconciliation drift. The Backtests page displays all 18 variant/fold results and the source SHA-256. This remains a synthetic replay; no metric is evidence of an edge.