Remaining Tasks — CVCE / Option A / Model Brief

Last updated: 2026-06-25

Tracking the four-task roadmap (Option A porting → PDF pipeline → EV PC data → gap-analysis brief).

✅ Task 1 — Port remaining visual panels into Option A prototype — DONE

All 9 atlas visuals now render from per-case data (no per-file HTML).

• Ported trade_map, activity_map, sankey, tech_tree into the prototype.
• scripts/proto/build_datadriven_proto.R: added firms per case (S&P) + per-tech side-files _tech_<Tech>.json (country-independent tree + sankey).
• analysis/_proto_datadriven/panel.html: topojson + ISO3N + 4 new cards/draw fns.
• Full regen: 253 case files + 11 tech side-files, 15 MB. All-158 projection ≈ 85 MB / 3.2 min.

✅ Task 2 — PDF rendering (landscape slides) — DONE (polish pending)

Root cause of the old broken PDF: iframes loaded/drew async → print fired before they settled (“Loading…” + distorted). Fixed by rendering each slide from Option A data synchronously.

• New analysis/_proto_datadriven/viz.js — shared, selector-parameterized rendering core (9 visuals
  • highlights() auto-bullets).
• New analysis/_proto_datadriven/report.html — landscape slide deck: ToC left · visual · bullets below, @page landscape, one slide per page. PDF validated end-to-end (9/9 pages render).

Polish backlog

1. Treemap / tech-tree labels clip at panel edges (e.g. “STREAM” ← UPSTREAM, “Solar Glass; E”).
2. Visual selection is URL-param only (?vis=radar,map,…) — add toolbar checkboxes.
3. Theme: currently light page + dark figures — decide on one direction.
4. Wire report_builder.html “Save as PDF” → report.html for the current selection (closes the loop between the interactive composer and the print pipeline).
5. Bullets are auto-generated → later become the analyst’s editable narrative (IR’s reserved text strip).
6. Dedupe panel.html onto viz.js (currently panel.html keeps its own copy of the draw fns).

⛔ Task 3 — Add EV PC data — BLOCKED (awaiting data)

EVs aren’t in the PC ML dataset, so the EV radar + PC scatter render blank.

• The EV model exists (RCA EV Analysis.ipynb, RandomForest + SHAP) but its outputs were not shipped — RCA Construction/EV/ is an empty folder in ML_vars.zip. See model_data_inventory.md.
• Action: request the 3 CSVs the notebook writes to RCA Construction/EV/ from Ishana/Alon:
  • ev_rca_by_category_year.csv → radar axis (pc_rca)
  • predicted_comp by country-year → PC scores (pc_scores)
  • shap_df / shap_by_cat (per-HS6 SHAP) → pc_features
• Once received: slot into data/pc/* for EVs, rebuild radar + scatter (incl. Option A regen).

🟢 Task 4 — Gap-analysis handbook (“Model Implications”) — v1 DRAFT DONE

Handbook for SHAP–RCA gap analysis, written as a manual (per user’s notes), one worked example per WP-4 cell.

• Output: qmd/report/model_implications_handbook.qmd → renders HTML + PDF + DOCX (docx = the Google Docs target). “Reading the Capability Map.”
• Decisions locked: audience = one source serving internal + public; format = QMD → HTML/PDF, then docx for Google Docs.
• Framework: the WP-4 7-cell grid (SHAP level × RCA level) at two zooms — radar (category) and PC scatter (product). Live R classification from pc_features.csv + pc_rca.parquet (ref year 2024).
• Worked cells (verified): Leverage→Brazil·Biofuel (product) · Build-up→Canada·Batteries · Critical gap→India·Solar · Mature→Canada·Nuclear · Gap→India·Solar·Electronics · Bonus→Brazil·Heat Pumps · Not-priority→India·Solar·Ind.Materials. Full India·Solar diagnostic plot in §5.

Follow-ups

1. Product-level scatter examples are currently narrative (e.g. Brazil ethanol “leverage”) — back them with actual HS6 product-RCA × SHAP data from the scatter pipeline.
2. Fold in Alon & Ishana’s optimization tool (LLF sensitivity / RCA-impact-by-category) as a forward-looking section (“from gaps to investment sizing”).
3. Team review pass (Bentley/Tim/Alon/Ishana), then public-polish from the same source.
4. Optional: NZIPL brand styling (dark-green theme, logo) for the public version.