`scripts/image-integrity-sweep/FINDING.md`

45 lines
# Image integrity: only 2 active products have NO image (genuine fixable defects) — near-clean bill; 0/39 sampled featured-image URLs broken; 63.9% single-image PDPs is a soft merchandising signal — cycle 88

**DTD verdict B (unanimous 3/3 — Claude + Codex + Qwen).** READ-ONLY (Admin GraphQL full active scan + a bounded 39-URL 404 spot-check; no mutation), **$0**. First image-quality sweep ever; the twice-deferred (Qwen c86/c87 dissent) surface. Picked over merchandising-depth (A) for clean execution + high probability of real defects after two clean discoverability cycles.

## 🟡 Result: 2 genuine no-image defects in 71,720 active — the unambiguous defect is essentially absent, but the 2 are real and orderable
Full active scan (71,720, complete=true, PJ excluded):
| Images | Products | Share |
|---|---|---|
| **0 (DEAD PDP — the defect)** | **2** | **0.003%** |
| 1 (single-image PDP, soft) | 45,837 | 63.91% |
| 2–5 | 20,593 | 28.71% |
| 6+ | 5,288 | 7.37% |
- **2 active products have NO image AND no featuredImage** (storefront shows a blank/placeholder), both ORDERABLE:
  - "Petite Zig Zag White on White Vinyl Screen Print" [China Seas] — type Wallcovering, inventory 2026.
  - "Pipelines Brick Wallcovering | Tres Tintas" [Tres Tintas] — type Wallcovering, inventory 4052.
- These are genuine, fixable, customer-facing defects: a shopper who lands on either sees a blank PDP yet can add to cart.

## 🔬 Skeptical decode — the negative control + a bounded 404 layer
- **Negative control holds:** with-images = 71,718 (100.00%). So a 0-image product is a TRUE anomaly against an image-rich bulk, not the norm — the 2 are real outliers, not a measurement artifact.
- **404 CDN layer (the hidden defect a null-check misses):** a non-null featuredImage URL can still be a broken 404. I HEAD-checked a SPREAD sample of 39 featuredImage URLs (every ~1800th product) → **39/39 reachable (2xx/3xx), 0 broken.** No signal of a broken-CDN epidemic. NAMED LIMIT: this is a 39-of-71,720 SAMPLE — it rules out a widespread CDN-rot problem, NOT every individual broken URL. A fuller 404 pass is a future option if warranted.
- **Lifecycle:** both no-image products are ACTIVE + inventoried → live customer exposure (not draft/archived), so the P2 severity is warranted despite the tiny count.

## 📋 Disposition (GATED — Steve's call)
1. **Fix the 2 no-image products — the immediate action is a RULE-MANDATED draft-flip, not a discretionary add-image (officer):** "active + orderable + zero image" is a hard breach of the standing "NEVER ACTIVE without image" rule (a shopper can place a paid order on a blank PDP right now). So the prescribed move is NOT "add an image OR unpublish" as equal options — it is: flip both to DRAFT + tag `Needs-Image` immediately (cheap, reversible, Steve-gated prod write), THEN source imagery from the repped vendors (Tres Tintas = confirmed trade line per `newwall-trade-pricing`; China Seas = Quadrille brand) as the actual P2 follow-up. Count is P2; the state is a clean rule-breach → draft-flip first, image-sourcing second.
2. **Single-image PDPs (63.9% / 45,837) = a SOFT merchandising signal, NOT a defect (P3/quality):** the majority of the catalog has exactly one image. NOT a broken state (the product is sellable, titled, priced, imaged) and 63.9% is non-actionable on its own — a flat colorway swatch legitimately needs one image; a statement mural wants room/scale shots. **Mass-flagging would manufacture 45,837 false "defects" and bury the 2 real ones (officer) — do NOT.** If any future cycle targets this, the actionable cut is "single-image AND high-traffic AND high-price," not single-image writ large; home it in targeted enrichment (room-setting-generator) prioritized by consideration value, not image-count.
3. **No broken-CDN epidemic** (0/39 sampled). Image integrity is otherwise HEALTHY fleet-wide.

## What this does NOT cover (named residuals)
- **Full 404 CDN verification = the natural NEXT image cycle (officer):** a full HEAD pass over all 71,720 featuredImage URLs (read-only, ~$0, rate-paced) — only a 39-URL sample was checked here (rules out an epidemic, not every URL); a real broken-CDN tail would surface there if one exists.
- **Image CORRECTNESS — higher-severity residual than count-depth (officer):** a featuredImage that LOADS but is the wrong colorway/pattern is invisible to a presence/count sweep and is the more expensive customer error (shopper orders the wrong product). Not assessable via this null/count method — needs visual/perceptual matching. Flag higher than the single-image-depth item.
- **Merchandising-DEPTH** (curated-facet coverage, the c87-officer-queued surface) — still open, gated on cracking the Boost facet param.
- **inventory=2026 state** — still un-audited fleet-wide.

_Cycle 88. DTD-picked B (unanimous 3/3). First image-integrity sweep: full active scan (71,720) → only 2 products with NO image (China Seas "Petite Zig Zag" + Tres Tintas "Pipelines Brick", both active+orderable = genuine P2 fixable defects); negative control holds (100% with-images, so the 2 are true anomalies); a bounded 39-URL 404 spot-check found 0 broken (no CDN-rot epidemic, sample-limited). The unambiguous defect (no image) is essentially absent. Secondary signal: 63.9% (45,837) of active products are single-image PDPs = a soft merchandising/conversion-quality opportunity (Glitter Walls + PR grasscloths heavy), NOT a defect. Disposition (officer-sharpened): the 2 = rule-mandated draft-flip + `Needs-Image` then source imagery (P2, not a discretionary add-image); single-image 63.9% is soft-not-defect (do NOT mass-flag). Residuals: full-404 pass (next image cycle), image CORRECTNESS (higher-severity), merchandising-depth, inventory=2026. PJ excluded. Read-only, $0, the 2-product fix gated._

---
## 🏛️ OFFICER SIGN-OFF — vp-dw-commerce: CONFIRMED (4 refinements, folded in)
- **Q1 — P2-for-2 right, but SHARPEN to a rule-mandated draft-flip (folded in):** "active + orderable + zero image + zero featuredImage" is a hard breach of "NEVER ACTIVE without image," not a soft backlog item. Immediate action = flip both to DRAFT + tag `Needs-Image` (cheap, reversible, Steve-gated), THEN source imagery as the P2 follow-up. Don't frame add-image and unpublish as equal options — the rule says draft-until-imaged.
- **Q2 — "soft, not a defect" CORRECT; do NOT spin a cycle on 63.9% (folded in):** a single-image PDP is sellable, not broken; mass-flagging 45,837 would manufacture false defects and bury the 2 real ones. Future target, if any, = "single-image AND high-traffic AND high-price," homed in targeted enrichment by consideration value — not image-count writ large.
- **Q3 — the 39-URL sample is honestly SCOPED, not overclaiming:** the cycle named the limit explicitly and listed full-404 as a residual, so "no epidemic" off 39/39-clean is a defensible negative-screen, not a per-URL certification. The full HEAD pass over all featuredImage URLs (read-only, ~$0, rate-paced) is the natural next image cycle.
- **Q4 — scoping (folded in):** image CORRECTNESS (loads-but-wrong-colorway) is the more expensive customer error and is invisible to a presence/count sweep → flagged HIGHER-severity than count-depth. Confirmed the denominator is active-only (71,720, PJ excluded); the inv values confirm both no-image products are genuinely live-sellable (reinforces Q1 urgency).
- **Net:** near-clean bill accurate; 3 dispositions correctly prioritized; the one tightening = the 2-product fix is an immediate rule-mandated draft-flip, not a discretionary add-image. All actions Steve-gated prod writes.

CONFIRMED. 2 genuine no-image defects (draft-flip-then-image, P2); 63.9% single-image is soft-not-defect (no mass-flag); no CDN epidemic (sample-scoped); image-correctness + full-404 are the named higher-value next image residuals. Read-only, $0.