`scripts/continue-masked-segmentation/FINDING.md`

47 lines
# The c89 (b) "CONTINUE-masked never-onboarded" defect population = ZERO — resolved; buyability rests on real stock (97.5%) + intentional untracked POD (2.5%), both benign — cycle 90

**DTD verdict A (unanimous 3/3 — Claude + Codex + Qwen).** READ-ONLY (Admin GraphQL full active scan; no mutation), **$0**. Closes the c89-officer-named residual: segment the active catalog (all buyable) into buyable-via-real-stock vs buyable-ONLY-via-CONTINUE-oversell-at-qty≤0 (the hypothesized soft onboarding-defect population).

## 🟢 Result: the true (b) defect set is ZERO — buyability is 97.5% real stock + 2.5% intentional untracked POD
Full active scan (71,717, complete=true, PJ excluded), classified by best buyable path per product:
| Buyable by | Products | Share | Verdict |
|---|---|---|---|
| **REAL-STOCK** (tracked, qty>0) | 69,916 | **97.49%** | benign |
| **UNTRACKED** (not tracked, always-available POD) | 1,801 | **2.51%** | benign (by design) |
| **CONTINUE-OVERSOLD only** (tracked, qty≤0, CONTINUE = the c89 (b) set) | **0** | **0.00%** | — none exist |
| **BLOCKED** (tracked, qty≤0, DENY = unbuyable) | 0 | 0.00% | consistent with c89 (0 unbuyable) |
- **The hypothesized (b) "buyable only because CONTINUE oversells an empty tracked variant" population does NOT exist (0).** The 1,801 always-available-without-stock products are all UNTRACKED inventory (print-on-demand dropship: Pixels, Printy6, Contrado, Steve Abrams Studios, Fentucci POD merch) — made-to-order by design, not an onboarding defect.

## 🔬 Skeptical decode — I caught + corrected my OWN classifier flaw (the discipline working)
- **v1 reported 831 "CONTINUE-masked-only" — that number was WRONG and I did not ship it.** v1 treated null inventoryQuantity as -1, so UNTRACKED-inventory products (qty=null) fell into the "qty≤0" bucket and got mislabeled as oversold. My own sanity counter caught it: v1 flagged **457 "DENY-variant-buyable-at-qty≤0"** — an impossible state UNLESS inventory is untracked. The DENY/0 examples were all POD vendors (Printy6/Pixels/Contrado/Steve Abrams Studios), the tell.
- **v2 fix:** read `inventoryItem.tracked` per variant and rank paths real-stock > untracked > continue-oversold > blocked. Result: the 831 decomposes into ~1,801 untracked (benign) + 0 true-oversold. **value-matching-a-flawed-classifier is a tautology** — verifying the mechanism (tracked vs CONTINUE) independently is what surfaced the truth.
- **Negative control:** real-stock bulk 97.49% → a true CONTINUE-oversold-at-empty set would be a distinct anomaly. There are none.
- **Resolves the deeper c89 "73.9% CONTINUE" worry too:** c89's 73.9% was a VARIANT-level policy count; this is PRODUCT-level best-path. 97.49% of products have a qty>0 variant → CONTINUE oversell is layered ON TOP of real stock (belt-and-suspenders), NOT propping up empty inventory. So "buyable because we never say no" does not describe the tracked catalog — only the 2.51% intentional POD is stockless, by design.

## 📋 Disposition: c89 (b) residual RESOLVED at zero; no fix
1. **No CONTINUE-masked onboarding-defect population exists.** The c89 (b) concern is measured and closed at 0. No fix.
2. **The 1,801 untracked POD products are benign by design** (made-to-order: Pixels/Printy6/Contrado/Steve Abrams Studios/Fentucci merch — print-on-demand, no stock to hold). NOT a defect. The c89 fulfillment/lead-time residual applies to them only in the trivial, intended sense (POD is always made-on-demand) — no action.
3. **Buyability thread now fully closed:** c89 = 0 active-but-unbuyable (census); c90 = 0 CONTINUE-masked-at-empty, buyability = 97.5% real stock + 2.5% intentional POD. The inventory/buyability surface is structurally healthy.

## Carry-forward watch items (officer — net-new threads, gated, neither blocks this finding)
- **Variant-level CONTINUE-oversold census:** product-level data proves every product has ≥1 real-stock OR intentional-untracked path, which closes the (b) residual — but it does NOT prove per-VARIANT buyability. A multi-variant product could have one qty>0 variant + dead CONTINUE-oversold-at-0 siblings. That's a separate, lower-priority oversell-exposure question (and a known, accepted DW belt-and-suspenders posture), not a gap here.
- **POD-vendor liveness / discontinuation check:** the 1,801 untracked POD products are "always buyable by design," but that IS a fulfillment-exposure surface to TRACK (not absorb into "benign"): a POD vendor going dark / a discontinued POD design / an upstream-pulled Contrado SKU stays buyable on our store with no inventory signal to stop it. A read-only liveness check so untracked-always-buyable can't silently outlive its upstream.

## What this does NOT cover (named residuals — the open non-inventory surfaces)
- **Image CORRECTNESS** (loads-but-wrong-colorway, c88 higher-severity; Gemini-capped → needs a non-vision signal or a different key).
- **Full-404 image pass** (heavy fan-out, low probability after the c88 clean sample).
- **Merchandising-DEPTH** (c87-queued; gated on cracking the Boost facet param).
- **Title/feed hygiene** (the banned-word "Wallpaper" live blast-radius + the c86 misnamed-catch-all feed mislabel) — a separate compliance/feed surface.

_Cycle 90. DTD-picked A (unanimous 3/3). Segmented buyability by mechanism: full active scan (71,717) → 97.49% real-stock + 2.51% untracked POD + **0 true CONTINUE-oversold-at-empty** + 0 blocked. The c89-officer-named (b) "CONTINUE-masked never-onboarded" defect population does NOT exist. CRITICAL self-correction: my v1 classifier conflated UNTRACKED inventory (null qty → -1) with oversold and reported a contaminated 831; my own sanity counter (457 impossible "DENY-buyable-at-0") caught it; v2 added inventoryItem.tracked → the 831 was ~1,801 benign POD, true-defect = 0. Also resolves the c89 73.9%-CONTINUE worry: CONTINUE is layered on real stock (97.5% have qty>0), not masking empty inventory. Buyability thread fully closed (healthy). Residuals: image-correctness, full-404, merchandising-depth, title/feed hygiene. PJ excluded. Read-only, $0, no fix._

---
## 🏛️ OFFICER SIGN-OFF — vp-dw-commerce: CONFIRMED (boundary-case sound; 2 net-new watch items carried forward)
- **Q1 — v2 settles (b) at zero, boundary-case CORRECT:** v2 splits on `inventoryItem.tracked` BEFORE reading qty, so untracked POD (qty=null) can never be miscounted as CONTINUE-masked-tracked (the v1 bug). The qty==0 worry opens no hole: real-stock is `qty>0` (not `>=0`), so a tracked qty==0 CONTINUE variant lands in the oversold bucket by definition — 0 there is real.
- **Q2 — "layered on, not masking" SOUND for closing (b):** every active product has a qty>0 tracked path (97.5%) or is intentional untracked POD (2.5%) → answers "is CONTINUE hiding empty inventory" regardless of per-variant policy mix. Caveat: product-level does NOT prove per-VARIANT buyability (a multi-variant product could have a qty>0 variant + dead oversold-at-0 siblings) → a separate variant-level census, carried forward.
- **Q3 — self-correction handled HONESTLY (strongest part):** the contaminated v1 831 was never shipped as truth; caught by an internal sanity counter (457 impossible "DENY-buyable-at-qty≤0" = a genuine tell); root cause named precisely (null→-1 coercion); fix mechanistic (add `tracked`, rank paths). Validate-by-mechanism, not by vibes.
- **Q4 — buyability thread closed; the 2.5% POD is a fulfillment surface to TRACK, not wave off:** untracked POD is always-buyable by design (correct, not a defect), but a POD vendor going dark / discontinued design / upstream-pulled SKU stays buyable with no inventory signal → carried forward as a liveness watch item, not absorbed into "benign."
- **Disposition:** CONFIRM (b)=0 + buyability thread closed. Two net-new READ-ONLY watch items (gated, not fixes, neither blocks): (1) variant-level CONTINUE-oversold census; (2) POD-vendor liveness/discontinuation check.

CONFIRMED. The c89 (b) residual is measured and resolved at zero (v2 tracked-aware, boundary-correct); the buyability thread is closed healthy; two net-new threads carried forward. Read-only, $0, no fix.