# PLAN.md — Operative plan as of 2026-07-01

**Supersedes** the day-by-day calendar in `SPRINT_BRIEF.md` (the brief's rationale,
metrics, and locked parameters still stand — only the schedule and scope details
below replace it). `HANDOFF.md` Part 2 remains valid technical reference.

## 0. Goal (confirmed with user, 2026-07-01)

- **This is a commercial product.** The product is the **web app** (current
  FastAPI GUI): contractors upload a blueprint PDF, get units/capacities/CSVs,
  review flagged boxes in the browser.
- **North star: zero-touch.** Human review of flags is acceptable only during
  the current maturing phase; the long-term bar is fully automatic. This makes
  the recall-first rule *stronger*, not weaker — every silent deletion is a
  product failure, and "boxes needing review per floor" is now a first-class
  leaderboard metric (it measures distance from zero-touch).
- **Full A/B/C bake-off proceeds** (not descoped), on a **~1 month** timeline.
- **Budget is not the constraint.** Track spend per run (ledger), but spend what
  the diagnosis/training needs. The $150 ceiling is retired.
- **Division of labor:** the user does gold-CSV hand-review and Kaggle GPU
  training runs; Claude does everything else (code, harness, data prep,
  cascade, scoring, writeups, GUI/productization).

## 1. Root cause of the Carroll p13 blocker (diagnosed 2026-07-01, free/local)

The failure chain, confirmed by reading the run artifacts and code:

1. **Schedule extraction for p13 is corrupted** — `outputs/_p13_ultra/...schedule_cache.json`
   contains OCR garbage: `AC-28H` qty 50, `AC-U` qty 24, `AC-CESS`, `AC-O`, `AC-C`.
   Nothing matching the real `AC-9-x` tags on the floor.
2. The corrupted schedule **poisons the tag vocabulary** used by detection, so
   every candidate box carries a wrong tag string.
3. **Snap fails 0/45** — `_local_ocr_snap` searches for tag text that doesn't
   exist on the page (wrong string).
4. **Verify rejects 45/45** — `_refine_hit_box` (analyze_blueprint_b.py:2314)
   demands the **exact text `'{hit.tag}'`** be visible in the crop and returns
   `None` (silent deletion) otherwise. A tag misread becomes a deleted unit.
5. The review CSV ends up **empty** (header only) — the 45 rejections leave no
   reviewable trace (`HVAC_VERIFY_DEBUG` exists but was off).

Tile size was never the lever (the runbook's own decision table already said
so). The bug is **architectural**: the verify stage conflates two questions —
"is there a unit here?" and "does the tag match?" — and lets the weaker signal
(tag text, downstream of a corruptible schedule) veto the stronger one.

## 2. The fixes (Week 1 core work)

- **F1 — Never silently delete.** Verify-rejected candidates become
  fuzzy/yellow rows in the review CSV instead of vanishing. (~5-line change
  around analyze_blueprint_b.py:2152.) This alone un-blinds every future run.
- **F2 — Two-question verify.** Split the verify prompt: (a) *class-level* —
  "is there an HVAC indoor unit in this crop?" (controls keep/flag);
  (b) *tag-level* — "what tag text is printed near it?" (controls the tag
  field; mismatch with the detected tag → orange, never rejection). This also
  pre-builds the multi-class future (outdoor/CU as a distinct class answer).
- **F3 — Schedule sanity gate.** Score every extracted schedule before use:
  qty outliers (>20 for a single tag), malformed tags (`AC-U`, `AC-O`,
  `AC-CESS`), fraction of empty model numbers. A suspect schedule must NOT
  constrain the detection/verify tag vocabulary; instead flag it in the GUI
  (the manual schedule-region picker already exists as the human fallback).
- **F4 — Verify model configurable.** `gemini-2.5-flash` is hardcoded in
  `_refine_hit_box`; make it a recipe knob and test `gemini-3.5-flash`
  (brief step 7 — expected cheaper AND better).
- **F5 — Debug trail on by default** for all bench runs (`HVAC_VERIFY_DEBUG=1`
  in `BEST_ENV` / bench harness), plus a **cost ledger**: append per-run
  model/tokens/$ (from `_record_gemini_usage`) to `analyzer_b/results/cost_ledger.csv`.

## 3. Other plan defects found in review (fixed by this plan)

- **Two unreconciled benchmark harnesses.** The 80-version leaderboard
  (`bench.py`, 355+Dean, 4 pages) never validated `v26_consensus` against the
  gold set (Lewis/Carroll). Winner risk: a config tuned to easy pages.
  → Standardize on `bench_detectors.py` + gold set as canonical; keep
  `bench.py` pages as a secondary regression set.
- **Gold set too small for the decision rule.** With 2 buildings, "within 2
  points" differences are noise. → Promote **2 more reviewed buildings to
  gold** (user reviews them during Week 2), report per-building numbers, and
  never trust sub-2-point gaps.
- **Doc sprawl.** `.gsd/` is dead, publication docs are parallel-track. This
  file is now the single operative plan; anything conflicting is stale.
- **Repo hygiene / secret risk.** The stray `HVAC-Blueprint-Analyzer\` junk
  directory (broken-copy artifacts with U+F05C in filenames) contains an
  untracked, **un-gitignored `.env`** — a `git add -A` would commit an API
  key. Nothing has been committed yet (verified against full git history).
  → Verify the stray `.env` matches the root `.env` (no unique secrets),
  delete the junk directory, apply/merge `.gitignore_new`, commit the real
  backlog deliberately (not with `-A`).
- **OneDrive IO.** Training/eval must run from `C:\hvac_work` (non-OneDrive),
  foreground — unchanged rule, now part of Week 2 setup rather than a footnote.

## 4. Week-by-week

### Week 1 (now → ~Jul 8) — Fix Track A, make the benchmark trustworthy
**Claude:**
1. Implement F1–F5.
2. Re-run p13 (~$1–2): expect the 6 `AC-9-x` units to appear (green/orange/yellow —
   any color is a win vs. silent 0/6).
3. Wire gold Lewis + Carroll into the canonical harness; re-score
   `v26_consensus` and the fixed pipeline on gold → new Track A baseline row
   (center-recall, review-rate, $/floor via ledger).
4. Run `error_taxonomy.py` across ALL reviewed buildings; encode recurring
   failure modes into `guardrails.py` (multi-class-friendly rules, not
   tag-string post-filters).
5. Repo hygiene items from §3.

**User:** re-review gold pages that gain new flags after the fix (~30–60 min);
confirm p13 gold boxes are still right.

**Exit criteria:** p13 recall > 0; Track A has an honest gold-set number.

### Week 2 (~Jul 8–15) — Track B data + training
**Claude:** harvest YOLO labels from the ~20 reviewed buildings into
`dataset` with a **multi-class-ready schema** (class 0 = indoor unit; ids
reserved for outdoor/CU, diffuser, fan); finish symbol bank; run flip-n-slide;
build the synthetic data factory (~35% real / 65% synthetic, tune); assemble
`kaggle_upload.zip` + turnkey `kaggle_run.py` notebook (DocLayNet-pretrained
base + yolov8n-doclaynet control).

**User:** launch 2–3 Kaggle T4 training runs (everything pre-packaged; kick
off, download weights); hand-review 2 more buildings → gold set grows to 4.

**Exit criteria:** proposal-recall@conf0.05 on gold measured.
**Kill-switch (unchanged):** < ~0.85 → Track B shelved, Track C falls back to
"Gemini proposes, Flash patch-verifies".

### Week 3 (~Jul 15–22) — Track C cascade
**Claude:** proposer (YOLO or Gemini per kill-switch) with SAHI tiling +
Weighted Boxes Fusion; verifier = the F2 two-question patch check on Flash;
union-check vs. full-page Gemini on sample floors (the real recall question);
measure $/floor and sec/floor.

**User:** spot-review cascade output on 1–2 non-gold buildings.

**Exit criteria:** Track C metric row complete.

### Week 4 (~Jul 22–29) — Bake-off + productize the winner
**Claude:** full A/B/C metric table on the (now 4-building) gold set:
center-recall, review-rate (the zero-touch metric), proposal-recall@0.05,
$/floor, sec/floor, tag accuracy. Apply the decision rule (max recall; cost
breaks ties only within 2 points AND consistent across buildings). Writeup.
Then productize: wire winner into `modules/pipeline.py` DEFAULT_RECIPE, add
per-run cost/latency telemetry to the GUI, HF Spaces deploy prep, and a
follow-the-sun path from "review flags" → "zero-touch confidence thresholds".

**User:** sign off on the winner; HF Spaces account/secrets for deploy;
pricing thinking can start from the measured $/floor.

## 4b. Progress log (2026-07-01, evening)

- **F1–F5 implemented and validated.** p13 re-run #1 proved the kill chain: sanity
  gate fired, tags read correctly off the page, rejects flagged yellow.
- **New failure family discovered: `leader_callout_separation`.** Carroll p13's
  units are UNLABELED hatched bodies; tags live in circled callouts ~1000px away.
  Fix F6 (detect unlabeled duct-connected bodies as tag `UNKNOWN`, no snap-drag,
  verify reads the tag): p13 center-recall **0% → 67%** (4/6 located, 2 FN left,
  FP/review burden high — 39 flagged). Runs cost $1.67 + $1.67.
- **Gold set now 4 buildings** (user reviewed 1326 + 23042 on 2026-07-01; CSVs
  locked in `analyzer_b/ground_truth/_gold_1326/` and `_gold_23042/`).
- **Gold box convention (confirmed with user):** box covers unit body + tag when
  adjacent; unit body ONLY on leader-callout plans. Tag string lives in the tag
  column, location in the box.
- **New engine requirement F7 — multi-floor sheets.** One sheet can serve several
  floors ("3RD-4TH FLOOR PLAN", "THIRD THRU SIXTH", Carroll's "10TH & 11TH"/
  "12TH-16TH"). Two bugs: (a) the page classifier sometimes marks these `other`
  and SKIPS them entirely (Carroll pages 14/15/17/18 — whole floors silently
  missing from output); (b) building unit counts/capacity must multiply by
  floors-covered. Gold convention: boxes count once per sheet;
  `floor_multipliers.csv` sidecar in each gold dir carries the multiplier.

## 4c. Track A baseline on the 4-building gold set (2026-07-02) — THE headline finding

Production config (`v26_consensus` equivalent: tile 1800, runs 2, consensus 2,
strict prompt, no fp-filter) on 17 never-tuned-on gold pages, $6.13 total:

| Building | Center-recall | Gold units | Main failure story |
|---|---|---|---|
| 11 Lewis | **13.3%** | 15 | over-eager schedule gate emptied a mostly-good vocabulary → strict prompt blind |
| 417 Carroll | **51.8%** | 83 | p13 = 0 (leader-callout units need inclusive prompt); offsets on dense floors |
| 1326 (2016) | **66.2%** | 71 | p1 Cellar 0% with 14 preds/0 matches (uninvestigated — possibly another callout/offset family) |
| 23042 (2024) | **3.9%** | 76 | real units are `AH-2.x.y`; schedule lacks those rows entirely → strict prompt blind |
| **TOTAL** | **≈39%** | 245 | |

**The leaderboard's F1 0.967 was overfit to its 2 tuning buildings.** Root cause
across all four: the strict prompt treats the (unreliable-in-a-different-way-every-
time) schedule as a hard vocabulary. **F8** (implemented 2026-07-02): quarantine
suspect schedule ENTRIES instead of nuking whole schedules; if >30% of entries are
junk or none survive, force the inclusive prompt automatically.

**A/B RESULT (2026-07-02): inclusive prompt is the new production default (`v30_inclusive`).**

| Building | Strict recall | Inclusive recall | FP strict→incl |
|---|---|---|---|
| 11 Lewis | 13.3% | **60.0%** | 5 → 46 |
| 417 Carroll | 51.8% | **90.4%** | 4 → 151 |
| 1326 | 66.2% | **80.3%** | 17 → 102 |
| 23042 | 3.9% | **69.7%** | 3 → 146 |
| **TOTAL** | **39%** | **79.2%** | 29 → 445 |

Cost: $6.13 strict vs $14.66 inclusive for 17 pages (~$0.86/page). Wired into
`modules/pipeline.py` DEFAULT_RECIPE (`recall_prompt: True`) and registered as
`v30_inclusive` in versions.py. Open fronts, in attack order: (1) review burden
~26 yellow boxes/page → dedupe/merge + FP-family guardrails mined from the
Queens deletes (216 cropped FPs available); (2) 51 remaining FNs on gold → error
taxonomy per miss; (3) 81 mistags → tag-read hardening (Qwen second opinion);
(4) $/floor ~2.4× — Track C patch-verify is the cost lever. Verify-stage F1–F6
remain in effect throughout.

## 4d. Review-loop notes (2026-07-02)

- **72-01 Queens Blvd human review done** (partial floors — only floors with a
  saved/applied CSV count as verified). Feedback from the reviewer: yellow
  flags on fl 28 were 7 real units / 15 FPs — recall net is working (7 units
  saved from being missed) but FP burden must come down. Lever order:
  (1) dedupe/merge near-identical candidates before review, (2) harvest her
  delete decisions into FP-family guardrails via error taxonomy, (3) Track C
  patch-verifier is the real precision machine.
- **Bug: floor-name misread** — a Queens sheet titled "PARTIAL 5TH FLOOR" was
  labeled "35th Floor". Same family as F7 (page classify/title parsing).
- **Direction confirmed by user:** YOLO detector classes = unit TYPE from the
  schedule (wall-mounted / ceiling-ducted / cassette / vertical …), NOT tag
  strings; tag comes from OCR/verify. Symbol bank harvesting should key crops
  by unit type.

## 4e. Precision work + taxonomy (2026-07-03)

- **Cross-location dedupe: null result** (implemented, kept as safety net).
  Offline validation on saved _incl_* detections removed 0 boxes — duplicates
  are OFFSET near-dupes (100–300px, different sizes), not stacked; geometric
  merging would risk real back-to-back twins.
- **CU/CUH guardrail measured on Carroll:** FP 151→131, recall unchanged. Wire
  into engine path (currently only in ab_guardrails.py).
- **Queens harvest done:** 216 FP crops + 311 bank candidates under
  `analyzer_b/harvest/`, metadata CSV included. FP families identified
  (grille callouts, room-type text, keynote circles, shafts, 'A/C' text,
  fixtures) → added as verify-prompt negatives. Taxonomy in
  `analyzer_b/results/error_taxonomy_notes.md`.
- **Biggest FN family fixed in prompt:** 23042's small SOLID-BAR wall units
  with stacked dual-floor labels (15 of 51 gold FNs) — new clause in the
  inclusive prompt. ⚠ CONTAMINATION NOTE: this clause was derived from GOLD
  failures, so gold is no longer pristine for Track A prompt-tuning. Fair
  bake-off requires B/C to see the same taxonomy; ideally validate future
  prompt tweaks on non-gold buildings first.
- **v31 measurement run scored (2026-07-05): 83.3% recall — NEW SHIP CONFIG.**

  | Building | v30 | v31 | note |
  |---|---|---|---|
  | Lewis | 60.0% | 60.0% | flat |
  | Carroll | 90.4% | 90.4% | flat (p13 33% on production recipe vs 67% ultra) |
  | 1326 | 80.3% | **70.4%** | p1 Cellar collapsed to 0/11 — now 27% of ALL gold FNs, top diagnosis target |
  | 23042 | 69.7% | **92.1%** | solid-bar clause worked (p8+p3 = 100%) |
  | **TOTAL** | 79.2% | **83.3%** | FN 51→41; FP 445→499; mistag 81→74; F1 ~0.43 flat |

  Cost: ~$8.65 (remainder) + ~$5 (pre-crash portion). Status split across v31
  CSVs: 599 yellow / 209 gray-noise (bulk-deletable) / 145 green / 10 orange —
  noise share understated (Lewis + Carroll p4/6/8 predate the noise band).
  Verdict per decision rule: v31 recall +4.1 pts → stays (prompts already
  live). Precision/F1 recovery still owed: F1 target ≥0.52 unmet — next
  levers are the 1326-p1 diagnosis, outdoor guardrail (in flight), mistag
  hardening, and Track C.
- **Review-burden product fix identified:** flag-don't-drop means verify can't
  shrink the yellow pile, only label it. GUI must band yellows: conf≥0.5
  "needs a look" vs conf 0.3 verify-rejected "likely noise" (collapsed,
  bulk-delete). Task #9.

## 4f. First-contact evaluation protocol (user-proposed 2026-07-03 — adopt)

Every NEW building is an unseen test set exactly once, before any fix or
training derives from it. This replaces "tune against gold" as the honest
generalization measure (gold keeps the regression-test job).

1. Run 1–2 pages of the new building (one typical + one weird/dense) with the
   CURRENT ship config — ~$1–2, `--pages` / GUI floor select.
2. Human verifies every box (keep/delete/add) — the score isn't real until a
   human confirms the unit count.
3. Score via `review_mining.py` (TP/FP-deleted/FN-added → recall+precision)
   and append to `analyzer_b/results/first_contact_ledger.csv`:
   building, date, config/version, pages, recall, precision, notes.
   THE LEDGER'S TREND LINE IS THE HEADLINE PRODUCT METRIC — expected accuracy
   on a brand-new upload, config version by config version.
4. Mine errors → guardrails/prompts/training data → building joins the
   training pool. Never re-count a building's first-contact score after fixes
   derived from it (that's a tuned score, not a first-contact score).
5. Caveat: 2-page samples are noisy (±1 unit on an 8-unit floor = 12 pts) —
   read the trend across buildings, not single-building deltas.

## 4g. Title-block reader v2 (2026-07-06/07 — task #11, user's 3-step plan)

1. **High-DPI clip OCR** of the title-block regions (right strip, bottom-right
   corner, bottom strip at 350 DPI via fitz clip — cheap, renders only strips).
2. **Sheet-number parsing** (`M-303.00` pattern) — new `sheet_no` manifest
   column; the M-30x sequence flags the floor-plan series even when titles fail.
3. **Gemini-Flash fallback** (`HVAC_CLASSIFY_LLM=0` to disable) — fires ONLY
   when no INFORMATIVE standard title emerges (bare "MECHANICAL"/"MECHANICAL
   PLAN FILED" now counts as uninformative — that gate bug initially masked
   the fallback entirely). ~$0.01/failing page.

Validated on 554 W 46th (the 0-plan-pages building): **0 → 12 plan pages**,
every floor correct Cellar→Bulkhead, p6 auto-read as "3RD TO 7TH" ×5 (matches
the user's hand-verified truth), p2 stamp-title regression fixed. Canonical
554 manifest upgraded. Floor-name fallback regex also fixed (no longer glues
"FLOOR" to numbers across newlines — 527 Grand's "Floor 44" ghost).

## 4h. First-contact run: E082 (2026-07-07)

v31 ship config on "20260415-E08201001-Mechanical_100% Construction
Documents.pdf" pages 13 (Cellar ductwork — dense pick) + 18 (10th-floor
ductwork — typical pick). Out dir = the GUI stem dir (classify artifacts
copied from `_fc_e082`; June schedule cache reused). Pre-review CSVs
snapshotted to `_prereview/`. **Cost: $1.74** (737 Gemini calls).
Boxes awaiting user review: p13 = 39 (2 confirmed / 3 recovered / 31 review /
3 noise), p18 = 89 (7 / 1 / 73 / 8). Ledger row 4 pends her verification.
Same day: Kaggle Track B weights returned (ctrl 0.943 / doclaynet 0.981
same-corpus holdout recall); gold tile set built (3,136 tiles, 245/245 GT
sanity-checked). **Gold bench result: KILL-SWITCH FIRED** — proposal-recall
@0.05 = 0.20 (ctrl) / 0.14 (doclaynet) vs the 0.85 bar; R@0.25 = 0.14/0.05.
Same-corpus 0.94–0.98 → gold 0.14–0.20 = severe overfit to training
buildings (data volume/diversity is the blocker, as the recalibration memo
predicted). Track B shelved per pre-agreed rule (user sign-off pending);
Track C proceeds as "Gemini proposes, Flash patch-verifies". Full leaderboard
in C:\hvac_work\gold_tiles_v1\bench_log.txt.

## 4i. Plan revision (2026-07-07) — post-Track-B, research-informed

Full research reports in session logs; sources verified. Market context: NO
commercial takeoff tool has solved MEP/HVAC symbol counting (97-98% claims
are architectural elements; vendors explicitly punt MEP to human QA) — our
recall-first + tiers approach matches or leads the market.

**Metric change: report ZERO-TOUCH F1** (greens+oranges auto-accepted, grays
auto-dropped, yellows per current policy) per first-contact + as a bake-off
column. That's the product number; raw F1 stays for regression comparison.

**Tier 1 (free/cheap, do first):**
1. **Offline verifier bench** from ~350 labeled crops (216 Queens + 90 E082
   deletes + 38 E082 units + gold sheets): score the F2 verify prompt per FP
   family offline; iterate at pennies instead of $14 gold sweeps. On the same
   bench: **A/B Flash-Lite vs Flash** for verify (research: small-crop
   classification is Flash-Lite territory, 5-20× cheaper) and a **tag-grammar
   regex demotion** (read tag matches no plausible `FAM-…-n` grammar →
   gray) — kills AC-CESS/KE-2 word-fragment families.
2. **Detection-prompt audit vs Gemini best practices**: coordinate order
   [y0,x0,y1,x1]; UNIQUE per-instance labels for repeated units (published
   finding: Gemini undercounts dense identical objects — plausible factor in
   1326-p1's 0/11 and dense-floor offsets); thinking budget off for detection.
3. **GreedyNMM/NMM box fusion** for fp_offset_near_dupes (plain IoU dedupe
   measured zero effect; NMM merges transitive chains) — test offline on
   saved detections first.
4. **OCR snap audit**: rotated-text orientation handling (known PaddleOCR
   rotated-text bug family; verify what our snap uses and its rotation path).
5. Verify prompt tag-grounding: ✓ already implemented (audited 2026-07-07);
   template parser will make the tag context floor-correct.

**Tier 2 (spikes, timeboxed):**
6. **Exemplar-similarity scorer** from the symbol bank (DINOv2/Siamese
   embedding vs approved crops). Strongest research result: JBE 2025 MEP
   paper, few-shot graph+Siamese matching, **F1 0.95 on UNSEEN equipment
   types** — similarity generalizes where classifiers overfit (exactly
   Track B's failure). Use as tier-ranking feature + demotion protection
   (CRAC-C-1 case).
7. **Zero-shot visual-exemplar detectors** (T-Rex2/DINO-X/YOLO-World) on the
   already-built gold_tiles_v1 — UNVALIDATED on line art per literature;
   strict timebox; success = proposal-recall ≥0.85 → Track C gets a cheap
   local proposer.
8. **Vector-PDF subpath clustering** (pymupdf cluster_drawings) as auxiliary
   recall proposer on the 2 vector buildings — free spike, temper
   expectations (academic SOTA needed trained GNNs).
9. **Calibrated tier thresholds** (conformal-style) once verify+similarity
   scores exist; recalibrate as buildings accumulate; guarantees are
   approximate under building-shift — advisory, not certified.
10. Caution (research): if fusing Qwen+Gemini as voters, CALIBRATE each
    model's confidence first (MoCaE finding: naive fusion degrades).

**Unchanged/reaffirmed:** structure-prior stack (template rows → pattern
lever → gap flags → cross-floor consistency) is the highest-leverage Track A
work; no Track B retraining this sprint; no more gold prompt-tuning;
consensus-run marginal-value audit; 1326-p1 + Lewis FN diagnosis before
bake-off (dense-cluster undercount is now the leading 1326-p1 hypothesis).

## 4j. Fable-as-proposer probe (2026-07-08) — promising, reshapes Track C

User observed Claude Fable 5, one-pass in chat (no tiling/schedule), marking up
whole floors with apparent 100% recall + correct per-apartment tags + category
labels (indoor AC vs CRAC vs fans), ~$5/floor in chat. Tested on the TWO gold
floors that most broke the tiled-Gemini pipeline, blind, scored vs gold:

| Floor | Gold | Gemini pipeline | Fable one-pass |
|---|---|---|---|
| 1326 p1 (cellar) | 11 | 0/11 | **11/11**, tags 11/11, 0 FP |
| Carroll p13 (leader-callout) | 6 | 4/6 (only w/ F6 code; 0 without) | **6/6**, tags 6/6, 0 FP |

PROVEN: recall + tag identification are excellent on the hardest cases; also
correctly excluded TEF/BS/ERV/CV callouts on p13 (the FP family Gemini boxes).
NOT YET MEASURED: coordinate precision — the strict center-match failed only
because centers were ESTIMATED BY EYE off a downsampled Read image then scaled
x5-7 (my measurement error, not Fable's; loose tolerance = 11/11 & 6/6). Cost:
API single-pass est $0.15-0.40/floor (one call replaces the pipeline's hundreds
of tiled/verify/consensus calls) — comparable-to-cheaper than pipeline ~$0.65,
and the $5 chat figure is conversation-overhead-inflated.

ARCHITECTURE IMPLICATION: Fable becomes a serious PROPOSER for Track C —
"Fable proposes (near-perfect recall + correct tags) -> existing OCR-snap
grounds each unit to its printed tag for a pixel-tight box" — potentially
deleting tiling/consensus/verify entirely. Recall-first rule favors a proposer
that misses nothing.

OPEN RISKS before committing: (1) coordinate precision of Fable's OWN emitted
coords (needs a scripted API call, not eyeball); (2) dense floors — Claude caps
images ~1568px, so a 39-unit floor (Carroll p8) in one downsampled pass may
lose tiny symbols -> may need a few zoom sub-calls (still << hundreds of tiles).
NEXT STEP: real API test needs ANTHROPIC_API_KEY in .env (currently MISSING;
OpenRouter is set but may lack claude-fable-5 + has a small daily cap). Build
fable_propose.py -> send floor image, get structured coords, run hard + dense +
a clean floor, score precision, pair with snap, log true cost. Then it enters
the A/C bake-off as a proposer variant.

## 5. Standing rules

- **Recall-first, now enforced in code:** no code path may silently delete a
  candidate — reject means flag.
- **Review-rate is on the leaderboard** — it is the zero-touch progress metric.
- Every paid run appends to the cost ledger.
- Training/eval from `C:\hvac_work`, foreground.
- Gold buildings (now growing to 4) are never trained or tuned on.
- Dead end stays dead: no graph/topological approaches (raster PDFs only).