Added 1,560 further exact-GTIN community identity records from a fresh Open Products Facts bulk export. The 3,000-record open tier remains searchable and downloadable, with no manufacturer model, current-market, offer, price, score, rank or purchase recommendation inferred from those observations.
Extended the flexible exact-model review approach across all eight categories. Power banks, air fryers, cordless vacuums, robot vacuums and air purifiers now publish bounded qualitative cohorts with one clear order and visible specialist labels. Portable air conditioners and dehumidifiers show useful reviewed profiles but withhold an overall order because their comparable evidence is too narrow. Every primary review card now shows an exact-product image. The 20-product scores remain secondary specification tables. Current offers join only after the quality result and each buy action selects the lowest fresh tracked exact-product price. The same release corrected air-purifier CADR caveat wording that had leaked into 80 products in unrelated categories; scores and legacy ranks did not change.
Replaced the battery-heavy Soundcore list as the primary earbud result with a bounded eight-model, cross-brand comparison based on 24 hands-on review documents. A current trusted-site sense check moved Technics above Sennheiser, raised AirPods into the 2-4 range and routed Soundcore P41i to a travel and battery specialist result. Written verdicts, use-case conditions and ten missing candidates remain visible. The legacy 20-model specification table is still inspectable below. Current tracked prices are joined after the quality order and cannot change it.
Added one category-generic evidence-report contract across all 160 ranked products. Eight category profiles now state the primary real-world question, critical topics, evidence gate and required test conditions. Existing hands-on synthesis is included only where exact-model sources pass the gate; other products say not collected or evidence insufficient. Reports distinguish 14 category-aligned earbud syntheses from two legacy contextual pilots. They omit numeric LLM review scores and cannot alter ranks, offers or purchase-advice status. Wireless-earbud reports now put sound first and label the current rank as a Soundcore-only specification shortlist.
Registered review coverage for all 20 products and collected 49 usable hands-on sources. Fourteen products meet the three-source, two-publisher-group gate. Each eligible product now receives three holistic analyses plus an independent judge. Product pages lead with a verdict band, strengths, risks and use-case fit while keeping the retailer action in the main product summary. Six products remain explicitly evidence-insufficient and receive no synthetic verdict or estimate.
Compared three LLM analysis methods across nine exact-model reviews for wireless earbuds, a robot vacuum and a dehumidifier. The judgement-led prompt retained the strongest source coverage and paragraph-level grounding. Product pages and API v1 now publish its strengths, weaknesses, conflicts, use cases, unknowns, trust judgements and uncertain review estimate for three products. The estimate is experimental and cannot alter specification score, rank or purchase-advice eligibility.
Removed one inherited independent-performance flag from Devola DVAC09CW because the retained record has no qualifying independent performance source. Ranked membership, scores and ranks are unchanged. Purchase advice remains blocked across the catalogue.
Added 1,440 exact-GTIN community identity records alongside the existing 160 ranked UK evidence profiles. The wider records are selected from the global Open Products Facts bulk export with checksum, identity, category and concentration gates. They contain no manufacturer model, current market, retailer SKU, offer, price or stock. Scores and ranks are null, and both ranked-comparison and purchase-advice eligibility remain false. A cursor-paginated API, searchable catalogue and open ODbL download keep the larger dataset usable while API v1 remains intact.
Product cards, pages, search results, comparisons, readiness rows, JSON and MCP fetches now share one current-offer selector. The action links to the lowest fresh price among tracked exact-product retailers and reports how many current retailers were compared. It does not claim whole-market coverage or alter category rank.
Added a read-only MCP endpoint for exact-product search, source-linked evidence fetches, same-category comparisons and recommendation-readiness checks. Search and fetch now cover both publication tiers, while comparison and readiness reject identity-only records. Added privacy-bounded daily counters for page views, attributable retailer exits and MCP tool calls, with human, user-directed agent and crawler activity reported separately.
Added clickable product images, freshness-gated checked prices, availability, visible product-detail actions and attributable retailer links to all 24 homepage ranking previews. Category retailer actions now retain their click context. Product-page hero images no longer create an unlabelled retailer exit; the source remains available through an explicit source link.
Expanded recommendation readiness from the leading 40 to all 160 ranked products, producing 1,240 normalized input receipts. Receipt provenance now separates source access from field observation and marks 480 configured specification inputs as not re-extracted. Homepage and category pages now lead with dense rankings, sourced inputs, sensitivity ranges and visible evidence gaps. Every profile remains blocked from purchase advice.
Hardened structured-data selection around registered model, SKU and GTIN identities, excluded an unresolved Anker A134 family-page collision, and refreshed category scores from exact source offers. Six catalogue identities were corrected or replaced. The leading portable-air-conditioner and wireless-earbud records changed. Cordless-vacuum, robot-vacuum and wireless-earbud methods advanced to version 1.1 so missing binary fields receive neutral half-credit instead of silently becoming false.
Added complete 20-product candidate records for all eight categories and deeper profiles for the leading 40 products. The new records expose field-level normalized input receipts, tie-aware method sensitivity, buying-constraint coverage, explicit retailer slots and missing human review. Receipt hashes cover normalized decision inputs rather than whole webpages. Every profile remains blocked from purchase advice.
Expanded the live catalogue to 160 exact products across eight methods, adding 20 source-gated air purifiers. Added exact-model search, scoped outcomes, a two-product claim-level trace pilot, explicit purchase-advice blockers and a new evidence-system guide. Retailer actions now say “Check retailer listing” and remain visually secondary to evidence inspection.
Added a 140-model evidence-coverage audit and a dated offer snapshot. Both reports publish their method, limitations and downloadable model-level CSV rows. Search Console and Bing verification are complete. At publication, fresh in-stock offers used explicit retailer actions on product pages, ranking cards, top picks and comparison tables.
Added operator and funding disclosures, provenance-bearing comparison facts, semantic ordered rankings, distinct retailer and evidence actions, a 48-hour offer-claim gate, llms.txt, OpenAPI documentation and IndexNow support. API v1 remains backward-compatible.
Expanded to seven category-specific groups of 20 exact UK products. Every category has its own method, ranks 1–20, source-linked images, offer observations and attributable redirects.
Product pages, APIs and comparison tables now state that scores are meaningful only within a category. The shared platform contract covers identity, provenance, freshness, owner pools and commercial independence.
Exact identity, traceable inputs, complete scoring fields, a checked offer and either resolved conflicts or an explicit conservative conflict policy remain hard gates. Tolerated conflicts stay visible and block purchase advice. Missing independent testing now lowers evidence quality and must be disclosed instead of making the entire catalogue unusable.