TK-11449: fold codex adversarial pass — downgrade 'dominant signal' to association, scope the money-exposure claim to the tested path, mark the Mayenne staleness mechanism INFERRED, add the appeal-then-disambiguator fallback

aa7a9f6be85570c44dd9eb52af61b67a8d398c8d · 2026-09-10 18:18:10 -0700 · Steve Abrams

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016yzBkdi8E67XZz1TJNmvca

Files touched

M data/tk11449-pricemismatch-jsonld-test.json
M data/tk11449-tobacco-REPORT.json

Diff

commit aa7a9f6be85570c44dd9eb52af61b67a8d398c8d
Author: Steve Abrams <steve@designerwallcoverings.com>
Date:   Thu Sep 10 18:18:10 2026 -0700

    TK-11449: fold codex adversarial pass — downgrade 'dominant signal' to association, scope the money-exposure claim to the tested path, mark the Mayenne staleness mechanism INFERRED, add the appeal-then-disambiguator fallback
    
    Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
    Claude-Session: https://claude.ai/code/session_016yzBkdi8E67XZz1TJNmvca
---
 data/tk11449-pricemismatch-jsonld-test.json | 4 ++--
 data/tk11449-tobacco-REPORT.json            | 4 ++--
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/data/tk11449-pricemismatch-jsonld-test.json b/data/tk11449-pricemismatch-jsonld-test.json
index f6ec793..49a4f46 100644
--- a/data/tk11449-pricemismatch-jsonld-test.json
+++ b/data/tk11449-pricemismatch-jsonld-test.json
@@ -24,9 +24,9 @@
    "position1_variant": "Sold Per Roll - 53.54In Wide @ $1335.60 (AVAILABLE)",
    "tk11404_applicable": false,
    "why": "TK-11404 requires the SAMPLE at REST position 1 on a BARE url. Here position 1 is the $1335.60 roll and it is available, so selected_or_first_available_variant resolves to the roll. Live JSON-LD, microdata and data-vprice ALL say 1335.60. There is no $4.25 leak on this page.",
-   "classification": "FEED-SIDE STALE PRICE / UNIT MISMATCH",
+   "classification": "FEED-SIDE STALE PRICE / UNIT MISMATCH (INFERRED - I compared the CURRENT page markup against the CURRENT feed value; I did not observe the crawl moment Google actually used, so the staleness mechanism is inference, not proof. The MISMATCH itself is measured.)",
    "root_cause": "$63.53 == Kravet-family MAP computed from the mirror cost 42.35 (42.35 x 1.5 = 63.53 exactly) i.e. the PER-YARD MAP. Shopify now sells this as a per-ROLL variant at $1335.60. MC lastUpdateDate 2026-08-20T03:18Z predates Shopify updated_at 2026-09-10T20:59Z, and this offer is NOT among the 4,228 offers MC updated today - so the feed is frozen at the pre-conversion per-yard number. Same family as the channel stall (TK-11397), not an independent pricing bug.",
-   "money_exposure": "NONE in the under-charge direction. PROVEN: live JSON-LD 1335.60, microdata 1,335.60, and the storefront's own product meta carries variants[0].price = 133560 cents. A customer clicking the $63.53 ad is charged $1,335.60. It is NOT orderable at the advertised price. The defect is advertising-integrity + a Google disapproval, NOT a live money leak. NOT urgent in the way the brief feared.",
+   "money_exposure": "NO CHECKOUT UNDERCHARGE ON THE TESTED PATH (scoped claim, not an absolute 'no exposure'). PROVEN: live JSON-LD 1335.60, microdata 1,335.60, and the storefront's own product meta carries variants[0].price = 133560 cents. A customer clicking the $63.53 ad is charged $1,335.60. It is NOT orderable at the advertised price. The ad still displays a wrong price, which is a compliance/trust exposure; and I did not exclude other entry paths or cached states. So: no undercharge on the path tested, NOT 'no exposure'. NOT urgent in the way the brief feared.",
    "OPPOSITE_EXPOSURE_FLAG": "UNRESOLVED and worth its own owner: $1335.60 / $63.53 = 21.02, implying a 21-unit roll, but this product's own metafields say Unit of measure = YARD, Minimum order qty 1, Order increment 0.5, and Wallcover-Length(YD) = 8.76. 8.76 yd x $63.53 = $556.53, not $1,335.60. Either the length metafield is wrong or the roll price is ~2.4x over-stated. This is the OPPOSITE direction (overcharge) and was invisible while the story was framed as 'we advertise too cheap'. Hand to the MAP/pricing owner - out of scope for a GMC read-only pass."
   },
   {
diff --git a/data/tk11449-tobacco-REPORT.json b/data/tk11449-tobacco-REPORT.json
index 6df6691..38d9bd7 100644
--- a/data/tk11449-tobacco-REPORT.json
+++ b/data/tk11449-tobacco-REPORT.json
@@ -9,7 +9,7 @@
  "PROVED": [
   "The DW primary feed carries ONLY 9 attributes: availability, condition, googleProductCategory, identifierExists, imageLink, link, price, shipping, title. There is NO description, NO product_type and NO brand attribute at all. So 'description text' and 'product_type' were never candidate signals for ANY of the 16.",
   "googleProductCategory is ALREADY 'Home & Garden > Decor > Wallpaper' on flagged AND clean offers alike. Google already knows these are wallpaper and flags them anyway.",
-  "PIPA (9): 43 Pipa offers. Contingency: bare titles 9/23 flagged, suffixed ('... Wallcovering | DW Bespoke Studios') 0/20 flagged (Fisher p~0.002). 19 pairs share one source image; 7 are DISCORDANT (bare flagged / suffixed twin clean). Independently re-verified by SHA-256: 4 of those 7 pairs are BYTE-IDENTICAL images (Dig-700011/700015/700019/700031); 3 differ (700043/700051/700063). The two cohorts were created 22 minutes apart the same night (bare 01:08:48-01:09:38, suffixed 01:31:40-01:32:49 on 2026-08-20) and re-evaluated in the same 02:50-03:35 window, so crawl-epoch is NOT a plausible confound.",
+  "PIPA (9) [ASSOCIATION, not proven dominant cause - within this tested subset a bare title APPEARS NECESSARY (9/9 flagged are bare) and is clearly NOT SUFFICIENT (14/23 bare are clean)]: 43 Pipa offers. Contingency: bare titles 9/23 flagged, suffixed ('... Wallcovering | DW Bespoke Studios') 0/20 flagged (Fisher p~0.002). 19 pairs share one source image; 7 are DISCORDANT (bare flagged / suffixed twin clean). Independently re-verified by SHA-256: 4 of those 7 pairs are BYTE-IDENTICAL images (Dig-700011/700015/700019/700031); 3 differ (700043/700051/700063). The two cohorts were created 22 minutes apart the same night (bare 01:08:48-01:09:38, suffixed 01:31:40-01:32:49 on 2026-08-20) and re-evaluated in the same 02:50-03:35 window, so crawl-epoch is NOT a plausible confound.",
   "Feed title == Shopify product title, verbatim (verified on the Fentucci set via Shopify Admin API). The titles are not synthesised by any DW feed script for this datasource."
  ],
  "CORRECTIONS_to_the_prior_pass": [
@@ -55,7 +55,7 @@
    "Google's own per-offer classification rationale. The API returns only the generic 'Tobacco products & related equipment' description and no attribute pointer (itemLevelIssue.attribute is null on all 16), so there is NO machine-readable signal naming what it objected to.",
    "Whether the 6 would clear on re-review without any change. All 6 have lastUpdateDate == creationDate (2026-08-19) and have never been re-evaluated in 22 days."
   ],
-  "honest_conclusion": "The most parsimonious reading consistent with ALL the evidence is that these 6 are stochastic false positives from a policy classifier operating on a texturally ambiguous natural-fibre product with a thin feed record (9 attributes, no description, no brand) - i.e. there may be NO discriminating feature to find. I am stating that as the most likely reading, NOT as a proven cause. The correct remedy for them is an APPEAL, not a content change, because there is no identified content defect to change."
+  "honest_conclusion": "UNDETERMINED - likely classifier error. Deliberately NOT upgraded to 'proven stochastic': latent inputs I could not observe (crawl timing, render variance, account-level priors, unseen policy-model features) remain possible, and 'the feed carries only 9 attributes' does not prove Google used only those nine. The correct FIRST remedy is an APPEAL, because no content defect was identified. FALLBACK if the appeal fails: a cheap reversible title disambiguator via the live supplemental (Path C) - not first, because 462 sibling offers share the identical suffix-less template and are clean."
  },
  "TITLE_FIX_CANDIDATES": [
   {

← 719c07f TK-11449: read-only diagnosis — tobacco_policy_violation + t · back to Gmc Titlefix · TK-11449: refute the Mayenne overcharge claim (verification 7f2c1f9 →