# docs/PROGRESS.md — phase journal (actual commands and results)

## Phase 0 — Audit (2026-11-02) — GATE PASSED
Plan: audit the UQO repository, FinData, hardware, competitors, literature and licensing; write the four audit documents.

Executed (laptop, 14:05–14:40 ET):
- Read `Recherche-UQO/README.md`, `HARMONISATION.md`, `_modele-working-paper/*`, WP 12/39/33 READMEs and `PROGRESS.md`.
  Result: WP 40 is free; conventions recorded in `docs/audit.md` §1.
- FinData: `fdq health`, `fdq datasets`, `GET /openapi.json` (saved), `GET /market/bars?format=arrow` (SPY, DST weekend),
  `GET /market/sessions` (SPY, Thanksgiving week 2024 → 211 expected bars on 11-29), `POST /exports` (SPY+QQQ Jan 2024:
  34 607 rows, 1.5 s, version 5162, SHA-256 recorded), `GET /query/views`, `GET /coverage/symbols`. Results in §2.
- Hardware: `sw_vers`, `sysctl`, `system_profiler SPDisplaysDataType`, `xcodebuild -version`, Metal device query
  program (`docs/evidence/` test), 29-node parallel SSH probe (22 reachable; files in `docs/evidence/nodes/`).
- Metal: trivial kernel compiled at runtime (34.6 ms) and executed (1 M squares, 0.029 ms GPU); `double` kernel rejected
  by both runtime and offline compilers; Metal Toolchain 27B47 downloaded (`xcodebuild -downloadComponent MetalToolchain`).
- Toolchain installs: `brew install eigen apache-arrow catch2 nlohmann-json ninja` (background), R
  `install.packages(c("highfrequency","rugarch","xts","lmtest"))` → highfrequency 1.0.3, rugarch 1.5.6, xts 0.14.3
  (lmtest failed: Java runtime missing for a dependency; not needed), `uv venv --python 3.12` + pip install of the
  Python reference stack (versions in `docs/capabilities.md`).
- Literature: 39 key references resolved on Crossref (`docs/literature/crossref_resolved.json`); preprint hits were
  re-resolved to journal versions; three DOIs verified directly (Econometrica 2008 realized kernels, SPA 2009
  pre-averaging, OUP 2010 semivariance). Still to resolve: Hansen–Lunde–Nason 2011 Econometrica MCS DOI; software
  citations (Arrow, Eigen, nanobind, Metal docs) to be added with URLs/versions.
- Legal: FirstRate private-use license (no redistribution), MIT for RAID, Metal proprietary → optional backend.

Gate memo (scope and feasibility): feasible on the laptop with the measured toolchain; FP64-on-GPU impossible → FP32
strategy; FinData provides all Tier A inputs (1-minute OHLCV bars, versions, sessions, exports); no trade/quote data →
Tier C order-flow methods excluded; cluster usable for E7 with a static bundle; 7 nodes unreachable. Decisions D-001…D-010.

## Phase 1 — Foundation (2026-11-02, 14:40–15:40 ET) — GATE PASSED
Plan: CMake/Ninja build, C++20 core types, CPU backend, data abstractions (calendar, bars, returns, aggregation),
Arrow/Parquet reader, OLS/robust covariance, realized measures, GARCH-family QML (needed by the backend contract),
Catch2 tests, nanobind Python package, CLI, docs, CI.

Executed:
- `cmake --preset macos-release && cmake --build --preset macos-release` (Apple clang 21, Eigen 5.0.1 + Accelerate,
  Arrow 26.0.0, Catch2 3.16.1, nlohmann-json 3.12). Metal sources absent → the build reports "Metal backend sources not
  present yet (Phase 3)" and continues CPU-only (`raid devices` lists the CPU only).
- `ctest --preset macos-release` → **26/26 test cases, 314 assertions** (unit: core, calendar, stats, OLS,
  realized/GARCH; reference: statsmodels 0.15 OLS/HC0/HC3/HAC/cluster at 1e−9, scipy t p-values at 1e−12,
  exchange_calendars 4.13 XNYS 2000–2026 exact session/holiday/early-close sets, numpy realized truth at 1e−12, arch 8.0
  GARCH/GJR/EGARCH/Student-t log-likelihood within 5e−4 and parameters within 5 % + 0.01, forecasts within 5 %).
- Python: `cmake -DRAID_BUILD_PYTHON=ON` builds `raid._core` (nanobind 3.1.0, Python 3.12); `pytest python/tests` →
  7 passed (OLS HC3 vs statsmodels, realized vs numpy, GARCH convergence, calendar, bar validation).
- Real-data integration (file not committed): `raid validate /tmp/raid-export-test.parquet` (SPY+QQQ 1m January 2024,
  FinData export v5162) → 21 sessions each, coverage 1.0 (391 regular bars incl. the 16:00 print), 0 duplicates,
  0 errors; `raid estimate realized … --backend cpu` → 21 sessions × 390 returns, mean RV(SPY) 2.71e−5, 0.25 ms.
- Bugs found and fixed during the phase (recorded for the paper's engineering section): thread-pool completion race
  (local mutex destroyed before the last job released it → SIGABRT under load); NLL rescaling sign; EGARCH ω
  wrongly log-transformed; `tgamma` overflow for ν > 340 (→ `lgamma`); fixture truth computed on unrounded prices.
- Environment manifest: `experiments/manifests/environment_<host>.json` via `scripts/record_environment.sh`.
- Repository: bare `gitsrv:srv/git/wp40_uqo.git` created with the SPB Git post-receive hook; `origin` set; pushed.

Gate evidence: clean build on the macOS preset, all tests green, docs (`docs/architecture/overview.md`,
`docs/methods/registry.md`, `docs/api/cpp.md`, README) written. Open items carried to later phases: `bars.aggregate`
vs FinData derived 5m (Phase 2), GARCH backcast option (Phase 4), HAR/bootstrap (Phase 4), Metal (Phase 3).

## Phase 2 — Data (2026-11-02, 15:00–15:30 ET) — GATE PASSED
Plan: secure FinData adapter (C++/libcurl), `raid data` CLI, Python access, offline + opt-in live tests, provenance,
calendar/aggregation checks against the data provider.

Executed:
- `include/raid/data/findata.hpp`, `src/data/findata.cpp`: bearer auth from env/file only, TLS, redacted logs,
  token bucket (5 req/s, burst 10), retries with backoff + `Retry-After`, Arrow-page pagination (`X-Next-Cursor`),
  31/186/3650-day chunking, async exports (POST → poll → download, `.part` rename), SHA-256 cache with `.meta.json`
  (dataset version, request ids, key fingerprint), provenance export; `raid data catalog|health|sessions|coverage|bars|
  export|fund|query|log`. Python: `raid.findata.{bars, export, health, catalog, dataset, sessions, coverage,
  fundamentals, query, read_parquet}` (GIL released).
- Offline tests (`[findata]`): chunking caps, token-bucket timing, offline refusal / missing key, synthetic Arrow IPC +
  Parquet page decoding (schema v1 metadata) → 4 passed, 2 live cases skipped. Fixture bug found: pyarrow IPC stream
  truncated when the file object was not closed → explicit `with open(...)` (recorded).
- Live tests (`RAID_LIVE=1`, key file, cache in `datasets/cache/`, AAPL 2024-11-25…29): health ok; catalog contains
  `market_bars.stock.1m.raw`; `expected_bars` 211 on 2024-11-29 == RAID calendar; bars pages (2 634 rows, 4 sessions,
  dataset version 17375) == export rows (62 933-byte Parquet, sha256 recorded) bar for bar; second call served from
  cache; RAID calendar `expected_regular_bars` == FinData `expected_bars` for every session; **aggregate 1m→5m:
  extended policy reproduces FinData's derived 5m exactly (280/280 OHLCV)**; regular policy differs only in the 16:00
  bin (FinData adds 16:01–16:04 post-close minutes) → convention documented in the registry (D-008 unchanged: the
  regular session ends with the closing print). Total 60 assertions, all passed; ≈ 25 requests.
- Python smoke: `raid.findata.bars` → realized variance per session from live bars; export + `read_parquet`
  (`zero_copy_possible = true` for single-chunk files).

Gate evidence: offline tests in `ctest`; live run transcript above; no vendor data in git (`scripts/check_no_secrets.sh`).
Open: least-privilege key for benchmarks (needs `findata create-key` on M3U96a — owner action, optional).

## Phase 3 — Numerical primitives (2026-11-02, 15:40–16:30 ET) — GATE PASSED
Plan: Metal backend with FP32 "double-float" (error-free-transform) accumulation, fused realized kernel, rolling sums,
regression sufficient statistics, batched GARCH likelihood; CPU/GPU agreement tests; measured break-even analysis.

Executed:
- `backends/metal/shaders/raid_kernels.metal` (TwoSum/Fast2Sum double-float, deterministic threadgroup tree reductions,
  chunked partial sums combined in FP64 on the CPU) and `backends/metal/host/metal_backend.mm` (device discovery,
  prebuilt `raid.metallib` with runtime-MSL fallback, pipeline cache, buffer pool by power-of-two size class, GPU
  timing from `GPUStartTime/GPUEndTime`, `-fno-fast-math` / `fastMathEnabled = NO` so the compensation is not optimised
  away). Metal has no `lgamma`/`log1p`: Student-t constants are computed in FP64 on the host and passed per parameter
  vector.
- Tests `[metal]` (6 cases, 6 411 assertions): fused and single-statistic accumulators vs CPU FP64 within 1e−6
  (Σx⁴ 1e−5) for n = 390/3 900/39 000 across chunk boundaries (4 096); bitwise determinism across runs; rolling sums
  1e−5; regression statistics 1e−6 (and the solve from GPU statistics agrees with CPU QR to 1e−5); batched GARCH NLL
  2e−5 relative (GJR + Student-t fixed ν 5e−5), argmin over a 16-point grid identical; +inf on invalid parameters.
  Full suite: `ctest` 38/38 (2 live skipped), `pytest` 7/7.
- `bench_bench_primitives` (laptop M5 Max 40-core GPU, 18 CPU threads, load ≈ 3–4 from other sessions, no thermal
  warning, 5 repetitions, shuffled order, 2 warm-ups; raw records `experiments/results/raw/bench_primitives.jsonl`
  + manifest; summary `experiments/results/bench_primitives_summary.{json,csv}`, figure
  `fig_bench_primitives_throughput.pdf`). Medians of wall-clock including FP64→FP32 conversion and buffer traffic;
  verdicts use non-overlapping bootstrap 95 % CIs of the median:
  - realized_fused: Metal wins from 39 000 elements (100 × 390: 2.4×), 2.0–3.4× vs 18-thread CPU at 3.9 M–39 M
    elements (max 3.4× at 1 000 × 39 000); ties at 390 000 for 10 × 39 000 and 100 × 3 900 shapes; CPU 1 thread is
    27–40× slower than Metal at 39 M.
  - regression_stats (k = 4): break-even ≈ 390 000 elements for a single long series (3.9×); ties for most shapes
    between 390 k and 3.9 M; 1.1–1.4× at 39 M. CPU wins below 39 000 elements.
  - rolling_sumsq (w = 22, direct O(w) kernel): **CPU wins at every size** (Metal 0.4–0.9× of CPU-MT); the GPU kernel
    is memory/arithmetic-bound by its O(w) design and the CPU running sum is O(1) per element — a negative result kept
    for the paper (a scan-based kernel is future work).
  - garch_nll: with one parameter vector per series (P = 1) the **CPU wins everywhere** (0.01–0.33×; the recursion is
    sequential and the GPU is launch-bound); with P = 64 parameter vectors per series Metal wins from 100 series
    (1.9×) up to 7.9× at 1 000 × 3 900 — batched likelihood evaluation is where the GPU pays off.
  - Device time is 5–20× smaller than wall time for the reductions: FP64→FP32 conversion and buffer fills dominate
    ("unified memory is not free computation"); E6 will add the FP32-resident ablation.
  - Shader library: prebuilt metallib loaded in 0.23 ms; first kernel call 14 ms (pipeline creation); runtime MSL
    compilation of the bundle is measured separately in E7 (toolchain-less nodes).

Gate evidence: accuracy tests and the measured break-even table above; no claim beyond the tested shapes.

## Phase 4 — Econometrics, Tier A (2026-11-02, 16:30–17:40 ET) — GATE PASSED
Plan: HAR family, block bootstrap, R references (highfrequency, rugarch), Monte Carlo validation, gradient checks,
registry completion, Python/CLI exposure.

Executed:
- `estimators/har.{hpp,cpp}`: HAR-RV / log-HAR / HAR-X / HARQ (causal lags, average or single targets, HAC with
  ≥ h−1 lags), `har_rolling` (causal re-estimation + QLIKE/MSE), `har_batch` (GPU sufficient statistics, FP64 solve).
  `statistics/bootstrap.{hpp,cpp}`: stationary / circular block / iid, deterministic streams, scalar and row variants.
- R references (`scripts/reference_r.R` → `datasets/fixtures/reference_r.json`, highfrequency 1.0.3, rugarch 1.5.6):
  RV/BPV/RSV±/RQ equal to 1e−10 after two documented conventions (highfrequency `rQuar` uses (M+1)/3; `rBPCov` with
  makeReturns=FALSE applies no M/(M−1) factor); `HARmodel` h=1 coefficients 1e−8 and h=5 (average target) 1e−6; HARQ
  equal once the regressor is (√RQ_t − √mean RQ)·RV_t as in highfrequency; rugarch sGARCH/gjrGARCH/eGARCH/sGARCH-std
  log-likelihoods within 2e−4 and parameters within 5 % (eGARCH alpha1/gamma1 swap documented; ν at the bound for
  Gaussian data in both packages).
- statsmodels references (`make_fixtures.py`): HAR h=1/5/22, log-HAR, HARQ with HAC → coefficients 1e−9, SE 1e−8.
- Calendar: the exchange_calendars comparison was previously **not running** (API error swallowed into a WARN) — fixed;
  the exact comparison 2000-01-03…2026-12-31 (6 790 sessions, 254 weekday holidays, 60 early closes) revealed three
  historical one-offs (early closes 2002-07-05 and 2003-12-26; full day 2002-07-03), now encoded. All sets equal.
- Monte Carlo (`tests/simulation`, printed to the test log): GARCH(1,1) n=2000, R=120: bias μ −0.0004, ω +0.0033,
  α −0.0020, β −0.0026; RMSE 0.019/0.009/0.014/0.019; 95 % QMLE coverage 0.950/0.942/0.900/0.933; 0 failures.
  HAR h=5 HAC coverage 0.895/0.945/0.910/0.905 (R=200, T=1200). RV RMSE vs √(2/M)·IV: 5.48e−5 vs 5.66e−5 (M=39),
  1.85e−5 vs 1.79e−5 (390), 5.78e−6 vs 5.66e−6 (3 900). Stationary-bootstrap mean coverage 0.89 (R=100). Numerical
  NLL gradients agree across step sizes 1e−4/1e−5/1e−6.
- Python: `raid.har_fit/har_rolling/har_batch/bootstrap` (pytest 9/9, HAR+HAC equal to statsmodels); CLI `raid
  estimate har <bars> [--horizon] [--log] [--harq] [--window --step]`.
- Full suite: `ctest` all green (see summary in RESEARCH_STATUS).

Gate evidence: reference tables (R + Python), Monte Carlo logs, registry entries with tolerances.

## Phase 5 — Extensions, Tier B (2026-11-02, 17:40–18:10 ET) — GATE PASSED
Selection by research value (E4 needs them): VAR(p), orthogonalised and generalized FEVD, Diebold–Yilmaz connectedness
(static + rolling), Ledoit–Wolf shrinkage. `models/var.{hpp,cpp}`. Reference: statsmodels VAR on a simulated VAR(2)
panel (`datasets/fixtures/var_panel.csv`): coefficients/Σ_u/llf/orthogonalised FEVD 1e−9, generalized FEVD (hand
computation in Python) 1e−9, DY totals 1e−7, BIC order, LW shrinkage 1e−9 (`tests/reference/test_var_reference.cpp`,
184 assertions). Rolling connectedness parallel over windows (0 failures on the fixture). Other Tier B/C items are
deliberately not implemented (registry).

## Phase 6 — Benchmarking E1–E7 (2026-11-02, 18:10–20:40 ET) — GATE PASSED
Raw records + manifests under `experiments/results/raw/` (one JSONL per experiment, `.manifest.json` with host, versions,
load, thermal, seeds, input hashes); summaries and every paper table/figure produced by
`experiments/scripts/make_paper_assets.py` (hashes in `experiments/manifests/paper_assets.json`; prose numbers only from
`experiments/results/key_numbers.json`).
- E0: universe = seeded sample of 1 000 of the 6 046 symbols active through 2024 (14 600 coverage rows); 20 exports,
  43 346 886 bars, 631 MB, dataset version 17 375, 437 s; E0b: 1.82 M daily bars 2015–2024.
- E1 (estimator path, laptop): 926 symbols × 252 sessions = 40 M returns — CPU 18T 29–30 ms, Metal 20–28 ms
  (1.07–1.41×), ties/CPU wins ≤ 4.8 M elements; fusion gain 1.4–3.2×; max |ΔRV|/RV vs FP64 1.6e−7 (daily segments),
  1.4e−8 (yearly). Synthetic 10 000 × 1 month (85.8 M): CPU 54 ms, Metal 35 ms (1.53×).
- E2: HAR rolling (120-session window) on 100 symbols: median QLIKE HAR 0.196/0.146/0.134 vs RW 0.251/0.239/0.301
  (h = 1/5/22); HAR beats RW for 84/86/78 % of symbols. Runtime 781 series × 252 obs: per-series QR 18T 0.68 ms,
  batched CPU 3.9 ms, batched Metal 4.7 ms (tiny designs: GPU launch-bound).
- E3: 888 symbols × ~2 459 daily returns: convergence GARCH 100 %, GJR 98.4 %, EGARCH 92.8 %, GARCH-t 100 %
  (after ≤ 2 restarts); 2.9–18 ms per fit; median QLIKE of fixed-parameter 1-day forecasts vs RV 0.39–0.40 against
  0.27–0.28 for the RV random walk. GPU grid 888 × 256 parameter vectors: Metal 23 ms vs CPU 18T 117 ms (5.1×),
  argmin identical 888/888, max rel. NLL difference 2e−3 at boundary grid points.
- E4: realized covariance (5-min synchronised) N = 10/30/60/100: median condition number 81/422/5.9e3/singular (N > 78
  bins); LW shrinkage 0.16–0.33; rolling DY connectedness 133 windows in 0.5/2.3/8.9/29 ms, 0 failures.
- E5 (R = 1 000): GARCH n = 2 000 coverage 0.93–0.97, |bias| ≤ 0.005, 1 000/1 000 converged; n = 500 β bias −0.026,
  α coverage 0.90; HAR HAC coverage 0.94 (h = 1), 0.90–0.94 (h = 5); bitwise reproducibility across threads and Metal
  runs; Metal vs CPU RV 3.7e−8.
- E6: per 2.37 M-bar export file — RAID CPU 0.118 s (20 M bars/s), RAID Metal 0.125 s, Polars 0.165 s, DuckDB 0.164 s,
  pandas 2.56 s; read 0.065 s, prep 0.021 s, estimate 0.031 s (CPU) / 0.037 s (Metal). Column projection took the
  pipeline from 0.35 s to 0.12 s. Baselines: approximate early-close rule (no exchange calendar in those engines);
  bugs fixed on the way: DuckDB 2.0 API, Polars i8 overflow in hour×60.
- E7 (11 nodes, M2 → M6): all Metal tests pass with prebuilt metallib and runtime MSL; cold compile 37–49 ms (110 ms on
  the 8 GB M2), warm 0.3–0.6 ms; RV fused 1000×3900 speedup 4.0× (M2) … 1.1× (M5 Max/Pro); GARCH P=64 7.8× (M2 Max) …
  1.8× (M5 Pro); rolling 0.2–0.65× everywhere.
- Engineering findings during Phase 6 (recorded): serial tripower/post-processing diluted E1 (fixed: parallel, cbrt);
  Metal backend recreated per call through an expired weak_ptr (fixed: process singleton); E3 evaluation index bug with
  the longer daily history (fixed); per-series FP32 scaling in the GARCH kernel (max rel. error on real grids 6.8e−4 →
  2e−3 at boundary points only, argmin agreement 887 → 888/888).

## Phase 7 — Manuscript (2026-11-02, 20:40–21:30 ET) — GATE PASSED
`paper/main.tex` + 13 sections + appendices; 30 pages; 57 BibTeX entries all resolved on Crossref (DOI content
negotiation) or hand-verified; `latexmk` exit 0, **0 undefined citations/references**; all tables `\input` from
`paper/tables/` and figures from `paper/figures/`, both generated by `make_paper_assets.py` from tracked artifacts.
Known cosmetic overfull boxes in two listings; no placeholder, no hand-typed number.

## Phase 8 — Independent review (2026-11-02, 21:30–22:10 ET) — GATE PASSED
- Clean clone of `wp40_uqo.git` → `cmake --preset cpu-only` (Arrow/Metal off) builds, `ctest` 49/49; `macos-release`
  builds, `ctest` 55/55; `raid estimate realized … --backend metal` runs on the clone (fp32, 5 sessions); `latexmk`
  exit 0, 30 pages, 0 undefined; `make_paper_assets.py` regenerates the 9 tables byte-identical (the 3 PDF figures
  differ only by matplotlib's embedded creation timestamp).
- **Defect found and fixed by this phase**: the `.gitignore` pattern `data/` silently excluded `include/raid/data/` and
  `src/data/` from every earlier commit (the working tree built, the repository did not); now `/data/`. The raw
  benchmark records (`experiments/results/raw/`, 9.6 MB) were also ignored and are now tracked, as the charter requires.
- Numerical audit: `scripts/audit_manuscript_numbers.py` — 32/32 quantitative claims of the prose found verbatim from
  `key_numbers.json`. Secret/data scan: `scripts/check_no_secrets.sh` clean; only synthetic fixtures tracked.
- Release checklist: `docs/RELEASE-CHECKLIST.md` (gaps stated: Linux build not executed, pip wheel not built from a
  clean environment, E1–E6 need a FinData key). Tag `v0.1.0-wp40`.