1# docs/DECISIONS.md — decision log23| ID | Date | Decision | Rationale |4|---|---|---|---|5| D-001 | 2026-11-02 | Project lives at `Recherche-UQO/finance/wp40_raid/`, WP number **40**, remote `wp40_uqo.git` | Charter asks for `wp40_raid/`; highest existing WP is 39; finance is the closest domain (WP 12 / 6 / 39 precedents) |6| D-002 | 2026-11-02 | Primary development and benchmark machine = the owner's laptop (M5 Max, 40 GPU cores); cluster used only for E7 with a self-contained bundle | Most nodes lack Homebrew/cmake in SSH PATH; production nodes (M3U96a, M4M64a/b, M1M32) must not be loaded; laptop has the newest toolchain |7| D-003 | 2026-11-02 | GPU numerics are FP32 with compensated/pairwise reductions; FP64 only on CPU; every Estimate reports `precision` | Metal rejects `double` (verified) |8| D-004 | 2026-11-02 | Shaders are shipped as MSL source **and** a prebuilt `.metallib`; runtime `newLibraryWithSource:` is the fallback and the "first compilation" benchmark case | Nodes without the Metal toolchain can still run; the charter asks for cold vs cached compilation measurements |9| D-005 | 2026-11-02 | Build: CMake ≥ 3.28 + Ninja, Apple clang, C++20, Catch2 v3 tests, nanobind bindings, scikit-build-core wheels; Arrow C++ optional (`RAID_WITH_ARROW`, default ON on macOS) | Portability of the CPU core to Linux CI; Arrow is heavy |10| D-006 | 2026-11-02 | Data access only through the RAID FinData client (bars Arrow IPC pages ≤ 31 d, exports Parquet for panels, query only for universe selection); cache under `datasets/cache/` (git-ignored) keyed by dataset version | Charter §5; license; reproducibility via pinned `X-Dataset-Version` |11| D-007 | 2026-11-02 | Reference software pinned: highfrequency 1.0.3, rugarch 1.5.6, arch 8.0.0, statsmodels 0.15.0, numpy 2.5.4, scipy 1.18.1 | Installed and recorded during the audit |12| D-008 | 2026-11-02 | Session policy default = **regular hours 09:30–16:00 ET inclusive of the 16:00 closing print** (391 bars), extended optional; early closes from the exchange calendar (211 bars); no bar is ever manufactured | FinData conventions (`expected_bars`) and charter §5 |13| D-009 | 2026-11-02 | Tier A experiments use market data only (no fundamentals) to avoid PIT/survivorship issues in a methods paper | docs/risks.md R15 |14| D-010 | 2026-11-02 | Energy is reported only when measured with `powermetrics`; otherwise "not measured" | charter §8, risks R17 |15