Evidence for retail displays and complete pages

Bounded review, 4 October 2026. Three pending semantic choices. No production mutation or complete-system certification.

Recovered history

Read the original retail question pack, docs/atlas/_data/review/content-retail.json, and the complete current Content pillar. Read both 16 August source dictations: recommendation-carve-notes.md and brand-identity-retail-round-notes.md in docs/reference/2026-08/2026-08-16/.

The recorded October decisions already preserve whole pages, useful parts, photographs versus depicted work, overlapping discovery and endcaps versus other aisle displays. The earlier retail pack's claim that recognizing a whole page would reopen the model is therefore an unsound premise. The kiosk question concerns type-list membership, which the approved relationship paths do not settle. Its residual heading choice remains open; the preparation must not conflate those two questions. Exact type/placement choices were explicitly left for this review.

The original retail batch also contains eleven candidate names and three bundled redirections. This review splits the Shoppable question only. It does not approve the other candidates or convert a partial answer into completion.

Inspected external examples

• IAB/IAB Europe December 2024 final standards: read definitions and store-zone sections; rendered and visually inspected page 27, showing a screen within a kiosk and a screen in an aisle. This is a dated measurement reference, not a claim to the latest edition or a universal creative taxonomy. No video playback or interactive behavior was tested.
• Apple Shop iPhone 17: read the product-specific configuration and purchase links. No checkout was performed; no current price claim is needed for the recommendation.
• Anthropic Claude on Mars: read the complete returned article text. Its story structure does not become a product-detail composition because its URL includes “features.” Embedded media were not fully reviewed.
• A retail photo-booth supplier page and an exhibition index were considered as leads, but no installation-versus-booth recommendation relies on them. That example remains incomplete.

Selected system checks

Literal SQL and responses are preserved in docs/reference/2026-10/2026-10-04/retail-pages-live-evidence.json.

The schema probe covered selected content_item columns and the vocabulary/alias relations before naming them here. The exact-name registry probe returned only the shot_type value product_detail among the selected names. This does not prove equivalent content or vocabulary is absent throughout the database.

The two original Anthropic sample records still have content_type=feature_article, page_type=product_detail and content_format=image; their source URLs distinguish a page hero and a full screenshot. These are stored labels, not verified ground truth about the composition. A selected aggregation returned 20 undeleted records carrying page_type=product_detail, including null types and several other labels. No global retail-collection absence claim follows.

Local path and implications

Read the full backend/services/page_type_classifier.py. It separates structural page labels from routing and contains URL heuristics. Its current /features/<slug> branch returns feature; the stored sample values therefore cannot be justified simply by citing that current local branch.

Read _backfill_page_type in backend/services/post_processor.py: it applies a website-source guard, gives URL-derived values precedence and can propagate a value to a linked source record. Read the channel expansion in frontend/src/lib/content-view.ts: website results use page_type for subrows. These are concrete consumers for the later audit, not evidence that the deployed system equals this checkout.

Before repair, trace collection, parent/part identity, source publication, capture, classification writers, derived channel behavior, view/RPC exposure and every browse/tracker reader. Determine where a complete-page form and asset placement belong without duplicating conflicting facts. Verify that human corrections survive reanalysis. Do not infer a migration from naming approval.

For physical displays, inspect evidence and multiplicity for work → component, photograph → depicted work and appearance → location/time. For shopping, distinguish the interactive presentation from the media file and preserve historical observations when links change. Permissions, withdrawal and unknown targets remain part of the full contract audit.

Choice mapping

• R-01/display-and-location: display naming rule; exact subtype list remains open.
• R-04/type-or-purpose: complete Product Detail Page type versus broader page type with a purpose description. Whole-page membership remains approved.
• R-07/shoppable: interaction description versus whole-piece type; other R-07 rows remain open.

R-03 type-list membership remains unresolved in the heading review. Approved overlapping relationship paths do not approve multiple heading membership for a type. R-02 and R-05 approvals are carried forward, not sent again. The implementation state remains unverified.