analyzer_b/results/first_contact_ledger.csv

building,date_verified,config,pages_verified,tp,fp_deleted,fn_added,recall,precision,notes
72-01 QUEENS BLVD - Mechanical 2022-03-21,2026-07-02,v26-era default (pre-v30),18,289,216,22,0.929,0.572,"first-contact = pre-review predictions vs human review; source: review_mining.py"
554 West 46th St MH,2026-07-06,v31 ship (inclusive + taxonomy prompts),2,8,19,0,1.000,0.296,"RECALL 100% — green tier perfect AND complete (8/8 green precision+coverage); junk fully contained in yellow(11)/gray(8); zero-touch would have been flawless if greens auto-accepted + rest dropped. Page 6 = x5 multi-floor sheet (F7 multiplier applied). Caveats: classifier found 0 plan pages (title-block OCR failed building-wide; pages hand-picked); schedule extraction near-useless (2 junk-ish tags) yet detection+snap self-sufficient; all units ducted (wall-mounted exist on unreviewed pages); n=2 pages, read trend not delta. Cost $0.64."
2022.11.02 527 Grand Ave_M,2026-07-06,v26-era CACHED detections (June 28 run; not v31),2,33,0,0,1.000,1.000,"user-verified 100%; only box-position nudges (for training tightness � product tolerance fine per user). CAVEAT: detection was the June-28 cache so this row measures the old config on an easy AC-A/AC-B building; delete evidence unrecoverable (editor splice bug, task #10). Floor-name bug found: fallback regex glued FLOOR + stray 44 across newlines � fixed."
5 Withers St.MH,2026-07-08,v31 ship + thinking-fix + LeverB,2,28,7,0,1.000,0.800,"USER-VERIFIED. p3 2nd Fl + p6 5th Fl, $0.97. BEST first-contact yet: INDOOR recall 100% (28/28, 0 missed), precision 0.80 pre-review, and 0 RELABELS + 0 yellow pile — near-zero-touch. Review was just: confirm 28 greens (auto), delete 7 gray UNKNOWN junk (Lever B pre-tiered them). Validates the full fixed stack on an unseen building: thinking-fix -> tags right first time (E082 needed 28 relabels; Withers 0); Lever B -> junk pre-graybinned. All units AH-1/AH-2/AH-4 (air handlers); no CU/ERV on these floors (those are on the bulkhead p10, OLD June cache, not run fresh). Cost $0.97."
20260415-E08201001-Mechanical_100pct Construction Documents,2026-07-07,v31 ship (inclusive + taxonomy prompts),2,38,90,0,1.000,0.297,"RECALL 100% (0 boxes added). p13 Cellar: 7 real (AC-C + AC-C-1..6) of 39; p18 10th: 31 real of 89. Tier report: green 9/9 + orange 4/4 = PERFECT again; noise band 11/11 junk (safe); yellows = 104 boxes, only 25 real (24pct). Junk families: UNKNOWN x37 (fixtures/fittings - adaptive-UNKNOWN target), word-fragment tags AC-A x26 / AC-R x11 / AC-T x3 / bare AC-C x9. KEY: building uses per-floor sequential tags (AC-10-1..31 on 10th; AC-C-x in cellar). TAG ACCURACY AS-DETECTED ONLY 12/38 (32pct): user relabeled 26 units, dominated by AC-A schedule-echo (19 boxes AC-A->AC-10-x) - worst mistag_junk_vocab instance yet; positional recall unaffected. 2 final relabels user-confirmed 07-07: p18 id49 = AC-10-21 (sequence now COMPLETE 1..31 - gap prediction validated) and p13 id7 = CRAC-C-1 (computer-room AC, NEW family; cellar hosts TWO families - multi-family pages are a demotion-rule edge case). GUI relabels logged as result=kept (new tag) NOT relabeled - originals recovered from _prereview snapshot; editor should record relabels (new task). Pattern-conformance demotion would have cut yellows ~75pct. First live run of delete-persistence: worked (90 delete rows persisted). Cost $1.74."