electrical-takeoff

# Electrical takeoff Five steps. Each writes files next to `--out` (e.g. `outputs/<project>`). Never skip step 2's approval. ``` python scripts/qa.py <set.pdf> --out outputs/<p> # 1 sheet register + problems python scripts/assemblies.py <set.pdf> --out outputs/<p> # 2 plan: what will be counted -> show <p>-plan.md, get approval, set "status": "approved" in <p>-plan.json python scripts/takeoff.py <set.pdf> --out outputs/<p> # 3 count + measure -> xlsx + json python scripts/verify.py <set.pdf> --out outputs/<p> # 4 crops per type -> look at them python scripts/viewer.py <set.pdf> --out outputs/<p> # 5 HTML drawing viewer ``` ## 1. Drawing pack QA (`qa.py`) Writes `<p>-drawings.md`. Read the "Fix before takeoff" list to the user: scanned sheets, sheets with no number, progress/VOID/PRELIMINARY stamps, stale revisions. Lighting plans are picked from this register, never from file names. ## 2. Assemblies (`assemblies.py`) - the planning gate Reads the fixture schedule and writes `<p>-plan.md`: - every scheduled type, measured EACH or LENGTH (LENGTH only when the schedule prices per foot / "lengths as indicated"), and the assembly it maps to in `assemblies/electrical-library.csv` - if text-layer schedules explain under 60% of the labels on the plans, the real schedule is probably drawn as outlines: it OCRs the schedule sheets (needs Tesseract). Photometric tables and cut sheets are not the schedule - their rows are dropped or marked "verify" - letter-only marks (EX, EXD exit signs) are accepted only if the schedule's type column lists them - types described as EXISTING are left out by default (relamp scope - the user can include them) - labels on the plans that are NOT in the schedule are listed, not counted Show the user the table and the not-counted list. Only then set `"status": "approved"`. The library ratios marked PLACEHOLDER are examples: tell the user to replace them. ## 3. Takeoff (`takeoff.py`) - EACH types: counted by **symbol** (vector fingerprint learned from tagged fixtures) and checked against the **tag** count. If the learned symbol doesn't match most of the type's tags (fixture graphics pasted as images, tags drawn inside boxes) it falls back to counting tags. - Look-alikes rejected: dashed boundary lines (chains of identical dashes), single strokes, and symbols whose nearest tag belongs to another type. - LENGTH types: each tag is paired with the fixture body beside it; length = body length / sheet scale. If every body is the same tiny length, it measured a symbol, not the fixture: counted EACH, one per tag. - Confidence: HIGH = tag and symbol agree; MEDIUM = one source only; LOW = check it. - Output: `<p>-takeoff.xlsx` (BOQ, Derived, Marks, Not counted) and `<p>-takeoff.json`. ## 4. Verify (`verify.py`) - required before reporting numbers Look at `<p>-verify/<TYPE>.png` for the big types and anything MEDIUM/LOW. Blue = tag, red = what was counted. If a box sits on the wrong thing, fix the cause in code and re-run. Never edit numbers by hand. ## 5. Viewer (`viewer.py`) `<p>-viewer/index.html`: fixture list on the left, the sheet on the right, every fixture marked and coloured by confidence. Click a type to see all of them, step with arrow keys, C = correct, W = wrong, F = whole sheet, "To check" filter, Copy CSV. ## Optional: independent cross-check (`compare.py`, `docs/crosscheck-brief.md`) Give the brief to a second AI for a blind count, then `python scripts/compare.py --out outputs/<p> --pdf <set.pdf>`. Settle every disagreement on the drawing - neither count wins by default. ## Report to the user Totals per type, how each was counted, the LOW list, QA problems, what was not counted and why. State the limits: placeholder assembly ratios, types counted by tags only, unresolved schedule rows. ## Limits Vector PDFs only (scanned sheets are flagged, not counted). Fixtures must carry a schedule type mark on the plans. Untagged fixtures are found only where the symbol is vector, not an image.