Microstructure/AGENTS.md

Repository instructions

These instructions apply to the entire repository.

Mission and safety boundary

Build a reproducible research and simulation system for short-horizon market microstructure. The repository must never place live orders, authenticate to a trading account, or imply that a simulated result is executable profit.

Research integrity

• Never invent observations, performance, statistical significance, or a data source. Label synthetic, fixture, smoke-test, partial, and full-data results.
• Keep predictive quality, execution assumptions, and strategy results separate.
• Every generated result must include the configuration hash, input manifest hashes, UTC data interval, code version or explicit UNBORN, and dirty state.
• Preserve raw observations. Put transformations in normalized or derived data and log exclusions; do not silently repair suspect events.
• Treat a timestamp at decision time t as unavailable unless its event and receipt ordering prove it was observable at t. Features use information at or before t; labels begin strictly after t.
• Do not tune against the final test period. Use time-ordered splits, purge overlapping label horizons, and embargo adjacent folds when configured.

Engineering conventions

• Target Python 3.12 and a local Apple Silicon machine with 16 GB RAM.
• Keep core logic in src/microstructure; notebooks may call but not duplicate it.
• Prefer Polars lazy/streaming scans and partitioned Parquet. DuckDB may query partitions without loading the full data set.
• New data sources implement the adapter interfaces; exchange-specific fields do not leak into normalized research modules.
• Store timestamps as UTC epoch nanoseconds and prices/quantities as decimal-safe integer ticks/lots where the adapter supplies metadata; floating research columns must document their units.
• Randomized procedures require an explicit seed.
• External raw data belongs under ignored data/raw; only small, documented test fixtures belong in Git.

Verification

Run focused tests after each meaningful phase and make check before handoff. Tests must cover sequence gaps and book invariants, temporal leakage, purged splits, deterministic simulations, partial fills, fees, and latency. A test may not call the public internet; mock adapters at the HTTP boundary.

Primary commands:

make setup
make download-sample
make validate-data
make smoke
make test
make reproduce-sample
make report
make dashboard

Documentation discipline

Record material assumptions and reversals in docs/DECISION_LOG.md. Keep STATUS.md honest and current. Update docs/PROJECT_PLAN.md acceptance evidence when a milestone moves state. Do not manually paste model metrics into prose; reports must read machine-generated run artifacts.