All project documentation has been consolidated under docs/README.md. Refer there for component-specific details.
The Manifold Node computes and stores a deterministic “metric surface” for each instrument: entropy (H), coherence (C), stability (S) across fixed forward horizons. It runs locally on a GPU, writes Parquet, and optionally publishes a thin summary to Valkey for the live site.
1) Export OANDA creds in etc/oanda.env
.
2) Build:
bash
cmake -B build -DSEP_USE_CUDA=ON && cmake --build build -j
3) Download raw candles:
bash
./bin/data_downloader --instrument EUR_USD --granularity M1 \
--from 2024-01-01 --to 2024-01-31 \
--out /sep/data/raw/EUR_USD/2024/01/01/2024-01-01.jsonl.gz
4) Generate manifold:
bash
./bin/manifold_generator --instrument EUR_USD \
--input /sep/data/raw/EUR_USD/2024/01/01/2024-01-01.jsonl.gz \
--out /sep/data/manifold/EUR_USD/2024/01/01/metrics.parquet \
--horizons 1,2,3,5,10,15,30 --tau-seconds 60
5) (Optional) Publish last 7 days to Valkey:
bash
python3 scripts/manifold_publish.py EUR_USD --days 7
Outputs are deterministic. Each Parquet row includes engine commit, build id, and CUDA arch.
Total run time: less than 5 seconds
Total cpu time used: less than 5 seconds
Total disk space used: 0 Bytes