# DECISIONS.md — dated design and engineering decisions (WP39)

| ID | Date (ET) | Decision | Reason | Consequence |
|---|---|---|---|---|
| D-001 | 2026-11-02 13:00 | Compute on **M3U96a** (FinData host) with a dedicated least-privilege key `wp39_uqo`; repository lives in the UQO Research workspace (`Recherche-UQO/finance/invisible-balance-sheet`, dir name follows the series convention; project id `wp39_uqo`) and is rsync'd to `M3U96a:~/wp39_uqo` for runs | API on localhost, lake co-located, 28 cores / 96 GB; secrets never leave `~/findata-secrets` / `~/Desktop/Cluster/secrets` | results and manifests synced back; raw cache stays on the node |
| D-002 | 2026-11-02 13:05 | **Earnings release dates** (`earnings_events.event_date`) define the reveal date `A_q`; the filing date `F_q` defines structured-fact availability (`available_at = F_q + 1`) | press releases precede 10-Q filings by a median of 1 day; cutoffs must precede the *first* public reveal | cutoff rule `c < min(A_q, F_q)`; release rows read from Postgres (no API route) — documented exception |
| D-003 | 2026-11-02 13:10 | Daily bars are **rebuilt from the 1-minute raw export** (regular session), not taken from `market_bars.stock.1d.raw` | vendor 1d dataset covers about half the symbols before 2016 and no delisted names | one source, one adjustment path for all four frequencies |
| D-004 | 2026-11-02 13:30 | Streaming feature extraction (export → features → delete raw Parquet; manifest keeps SHA-256) | disk budget on the FinData host (hard floor 60 GB free) | re-runs re-export from the pinned dataset version |
| D-005 | 2026-11-02 14:00 | Universe = all FinData fundamentals entities (survivor-conditioned); delisted-fundamentals backfill **proposed, not run** | vendor ingestion covered active tickers only; backfill needs owner approval + capacity plan (rule 3 of the brief) | survivorship caveat stated in abstract, data section, limitations; price side is survivorship-free |
| D-006 | 2026-11-02 14:05 | **No neural (B7) nor text/multimodal (B8) models; no LLM anywhere in the pipeline** | owner instruction received mid-session on 2026-11-02 ("limite le plus possible l'utilisation AI"); also keeps the frequency comparison about information sets rather than capacity | ladder is B0–B6; LightGBM provides the nonlinear benchmark with equal budgets |
| D-007 | 2026-11-02 14:10 | Primary target = adverse **operating-income** surprise vs a seasonal random walk with drift, scaled by total assets, threshold −0.5 % | no point-in-time consensus exists in FinData; operating income is less affected by one-off items than net income; Foster (1977) model is the classic honest expectation | consensus-based miss (vendor `eps_estimate`) is exploratory only |
| D-008 | 2026-11-02 14:15 | Frequency sets B3/B4/B5 use **identical feature families** (one column per family) computed from bars of 30 m / 5 m / 1 m | compare information sets, not dimensionality | a supplementary "all frequencies" model is run in Phase 4 |
| D-009 | 2026-11-02 14:20 | Literature verified with **Crossref metadata** (no AI summarisation); references without a verifiable DOI are cited without one (Grossman–Stiglitz 1980, Holm 1979) | rule 7 of the brief (no fabricated references) | `artifacts/logs/crossref_batch*.json` are the evidence |
| D-010 | 2026-11-02 14:30 | Phase gates that require "owner approval" are recorded as **self-approved under the autonomous mandate** with the memo attached; the owner can veto any of them before Phase 6 | the owner asked for the whole project to be built in one autonomous session and is not available mid-task | each gate entry in PROGRESS.md names the memo and the date |