# Simulation Plan

## Goal
Build a minimal but useful Hermes-MCP simulation pipeline with clean separation between data acquisition, preprocessing, and replay/analysis.

## Current structure
- `download_candles.py` downloads raw Binance archives into `simulation/data/raw/`
- `prepare_candles.py` converts raw archives into a canonical prepared CSV plus manifest
- `run_replay.py` replays Hermes against the prepared input only

## Next steps
1. Add month-append support so new raw archives can be added without rebuilding everything.
2. Add prepared-dataset merging so multiple months become one continuous test set.
3. Add basic evaluation reports for decision quality and fee-adjusted outcomes.
4. Add optional parameter sweeps for unknowns in the decision layer.

## Working rule
Only the prepared dataset should be used for replay and optimization. Raw downloads stay isolated in `simulation/data/raw/`.