`scripts/ai-preamble-leak-sweep/FINDING.md`

50 lines
# AI-preamble body-text leak = 5 live products (ALL Versace), confirmed by two methods; the live "Wallpaper" TITLE blast-radius is 0 (titles fully normalized) — both better than the stale memory implied — cycle 91

**DTD verdict A (unanimous 3/3 — Claude + Codex + Qwen).** READ-ONLY (Admin GraphQL full active scan ×3 passes + independent productsCount reconciliation; no mutation), **$0**. Sweeps content quality of live product descriptions for leaked LLM scaffolding (the body-text analogue of the c88 blank PDP) + reconciles the banned-word "Wallpaper" title debt.

## 🟡 Result: 5 genuine AI-preamble leaks, ALL Versace — small, unambiguous, fixable
Full active scan (71,717, complete, PJ excluded):
- **5 HIGH-confidence AI-preamble leaks (0.007%), all [Versace]** — live descriptions that BEGIN with leaked LLM scaffolding:
  - "Versace Plain Gold Wallcovering" → "Here's the product description:"
  - "Versace Stripe Brown Wallcovering" → "Here's the product description:"
  - "Art De La Table Gold Black Wallcovering" → "Here is a product description:"
  - "Heritage Green Blue Wallcovering" → "Okay, here is the product description for the wallcovering:"
  - "Palazzo No 12 Light Grey Wallcovering" → "Here is a product description:"
- These are genuine, embarrassing, customer-facing defects (the shopper sees the AI's preamble line above the real copy) — clearly fixable (trim the leaked line / regenerate).

## 🔬 Skeptical decode — confirmed by TWO independent methods + a broad-net false-positive guard (the c90 lesson applied)
- **Don't trust one pattern.** Pass 1 (HIGH-confidence patterns: "here is a description/product description/revised/SEO…", "as an AI", "I cannot assist", etc.) → 5, all Versace.
- **Broad-net upper bound:** a loose "here is / here's" net flagged 58 active bodies — but that pattern catches legitimate marketing copy ("Here is a striking damask…"). So 58 is an upper bound, NOT a defect count.
- **Pass 2 (body-START classification)** — does the body BEGIN with the preamble (or "here is" + a description/colon word), the real leak signature? → **exactly 5, the SAME Versace products.** The other 53 "here is/here's" hits are mid-copy legitimate marketing (Designer Wallcoverings 28, Phillipe Romano 13, etc., all mid-sentence). So the true defect count is 5, corroborated by two methods — not 58 (over-count) and not missed.
- **Negative control:** 71,712 bodies have no LEADING-"here is"-family preamble → a body-start preamble is a true anomaly. SCOPE (officer): "clean" here means clean of the SCANNED preamble class, NOT defect-free — both methods are anchored on English, leading-position "here is…" scaffolding, so they would UNDER-COUNT (a) preamble buried mid-body after a legit opening, (b) non-"here is" families ("Sure!", "Certainly,", "Of course —", "As requested,", markdown `## Product Description`, trailing "Let me know if…"), (c) non-English scaffolding. The 5 is the right number to ACT on for the defined class; a cheap anchored-prefix follow-on for the other scaffolding families is queued (not blocking). 176 active products have an empty body (a separate thin-content signal).

## 🔑 Memory reconciliation — the live state is BETTER than `banned-word-live-blast-radius` recorded
The memory says ~25,131 LIVE products contain "Wallpaper" + ~68 AI-preamble leaks (incl. Thibaut AT10701). Measured TODAY by independent productsCount queries:
- **`status:active title:Wallpaper*` = 0** — NO active product TITLE contains "Wallpaper"; titles are fully normalized to "Wallcovering" (`title:Wallcovering*` ≥ 10,000, the count cap). The title-brand-hygiene concern is effectively RESOLVED.
- **`status:active` any-field "Wallpaper" = 9,486** — these are body/tag/VENDOR-NAME occurrences (e.g. vendor "Malibu Wallpaper", "Missoni Wallpaper" are proper nouns; body mentions) — mostly LEGITIMATE, not title-brand violations, and far below the memory's 25,131.
- **AI-preamble = 5 today vs the memory's ~68;** Thibaut AT10701 not found by free-text (remediated/archived/SKU-format). → Either substantial remediation happened since the memory was written, or the memory used looser patterns. Net: the live content-hygiene state is materially better than the memory implies → **the `banned-word-live-blast-radius` memory is stale and should be updated (flag for Steve).**

## 📋 Disposition (GATED — Steve's call)
1. **Fix the 5 Versace preamble descriptions (P3 — tiny, clean, unambiguous):** trim the leaked scaffolding line (or regenerate the description without the preamble). A 5-product surgical content edit, NOT a bulk rewrite. Steve-gated prod write.
2. **"Wallpaper" TITLE debt = 0 → titles need nothing; but body-COPY "wallpaper" is a SEPARATE lower-priority debt (officer):** the 9,486 any-field bucket is heterogeneous — vendor PROPER-NOUNS ("Malibu Wallpaper", "Missoni Wallpaper") are legitimate brand names that MUST stay, but "wallpaper" in OUR OWN editorial body copy does carry brand-voice + on-page-SEO cost and violates the banned-word rule in customer-facing text. So: titles = nothing to do (clean win); body-copy "wallpaper" = a separate, lower-priority debt worth its own scoped, PROPER-NOUN-SAFE pass (denylist that excludes vendor names). Do NOT propose a bulk title rewrite (titles already clean).
3. **Update the stale memory (REQUIRED, not just appropriate — officer):** `banned-word-live-blast-radius` (25,131 live + 68 preamble) materially misstates live state (0 title-Wallpaper, 5 preamble) and a wrong memory drives wrong future decisions — update it to today's measured reality, note the prior figure as historical/resolved. Keep **Thibaut AT10701 flagged for Steve** (was it fixed, archived, or re-handled? a one-line confirmation, not a silent assumption).
4. **176 empty-body active products** = a separate thin-content residual (no description) — minor, future enrichment, not this defect.

## What this does NOT cover (named residuals)
- **Non-"here is" preamble families (officer follow-on):** an anchored-prefix scan for "Sure!/Certainly,/Of course/As requested", markdown `## Product Description` headers, trailing meta ("Let me know if…"), and non-English scaffolding — the blind spot of this cycle's English-leading-"here is" anchor. Cheap, queued, low-priority.
- **Body-copy "wallpaper" proper-noun-safe pass** (disposition #2) — separate from vendor names.
- **WEAK-pattern body issues** beyond preamble (truncated copy, placeholder "[insert]", duplicated boilerplate, the 176 empty bodies) — only LLM-preamble was targeted.
- **Image CORRECTNESS** (c88), **full-404** (c88), **merchandising-DEPTH** (c87), **POD-vendor liveness** + **variant-oversold census** (c90 carry-forwards) — still open.

_Cycle 91. DTD-picked A (unanimous 3/3). AI-preamble body-leak sweep: full active scan (71,717) → 5 genuine leaks, ALL Versace ("Here's the product description:" / "Okay, here is the product description for the wallcovering:"), confirmed by TWO independent passes (HIGH-pattern + body-START classification both = the same 5) with a broad-net guard (58 "here is" hits, 53 = legitimate mid-copy, NOT defects). "Wallpaper" TITLE blast-radius = 0 (titles normalized to Wallcovering; independent productsCount), ~9,486 any-field occurrences mostly legit vendor-names/body — both materially BETTER than the stale `banned-word-live-blast-radius` memory (25,131 + 68). Disposition: surgically fix the 5 Versace descriptions (P3, not a bulk rewrite); titles need nothing; update the stale memory (flag). PJ excluded. Read-only, $0, the 5-product fix gated._

---
## 🏛️ OFFICER SIGN-OFF — vp-dw-marketing: CONFIRMED (framing tightened; 2 scoped follow-ons queued)
- **Q1 — 5-Versace count TRUSTWORTHY to act on, scoped honestly:** the two methods are genuinely independent in mechanism (literal-phrase match vs positional body-start), but share an English-leading-"here is" anchor → blind spot for mid-body preamble, non-"here is" families ("Sure!/Certainly/Of course/As requested", markdown `##` headers, trailing meta), and non-English. The broad 58-net is a partial backstop (53 legit-mid-copy + the same 5). So 5 = the right number to ACT on for the defined class; don't oversell it as "preamble eliminated fleet-wide." Anchored-prefix follow-on queued.
- **Q2 — titles=0 is a clean verifiable win; "9,486 = nothing" is NOT fully right:** heterogeneous bucket — vendor proper-nouns legit/must-stay, but OUR body-copy "wallpaper" carries brand-voice + on-page-SEO cost and violates the banned-word rule in customer-facing text. Reframed to "titles need nothing; body-copy wallpaper = a separate lower-priority proper-noun-safe pass."
- **Q3 — disposition RIGHT:** 5 surgical trims/regens (P3); a bulk rewrite would be wildly disproportionate (collateral risk to 71k clean bodies). Memory update is REQUIRED (a wrong memory drives wrong decisions), with AT10701's disappearance flagged for a one-line Steve confirmation (fixed/archived/re-handled).
- **Q4 — one no-overstate:** the negative control proves clean of the SCANNED preamble class, NOT defect-free (the 176 empty bodies are a live counterexample) — "clean" kept scoped to the preamble class.
- **Net:** act on the 5 Versace trims; update the stale memory (+ AT10701 confirm); queue two scoped low-priority follow-ons (non-"here is" preamble families; body-copy "wallpaper" proper-noun-safe pass). None blocking.

CONFIRMED. 5 genuine Versace preamble leaks (surgical P3 fix); titles clean (0); body-copy "wallpaper" + non-"here is" families are scoped lower-priority follow-ons; the stale banned-word memory should be updated. Read-only, $0, fixes gated.