docs/data_construction.md

Data construction and lineage

This document summarizes the released data lineage.

Released lineage

Belief

wallet holdings / position records
  -> dataset/Belief/source/data.parquet
  -> fixed wallet-level, time-ordered splits
  -> benchmark/Belief/{direct,profile,retrieval}

The source table summarizes a wallet's dominant held outcome and normalized holding imbalance per market. Split metadata records 2,000 experiment wallets, a 70/15/15 split ratio, ordering by stance block fields and market ID, rolling history, and a 30-item prompt-history cap.

Trade

trade and market records
  -> trade_price_matched.parquet + market_semantic.parquet
  -> behavior_fixcost.parquet
  -> dataset/Trade/splits/*.parquet
  -> benchmark/Trade/{direct,profile,retrieval}

Dataset and benchmark layers

• dataset/ contains structured source, split, and support data suitable for inspection or task reconstruction.
• benchmark/ contains model-ready prompts/contexts and targets.
• Parquet and .jsonl.gz files in the same benchmark directory are alternative serializations where both are present.