Every ranked product passes the same identity and traceability gates. The wider exact-identity index stops before scoring. The score itself changes by category because cooling, moisture removal, portable power, cooking, floor care, autonomous cleaning and audio are not interchangeable decisions.
Model, SKU, variant and market are resolved before comparison. GTIN is retained where the source exposes it. A retailer title is not allowed to merge similar products. A checksum-valid GTIN community record can enter the wider identity index, but it does not establish manufacturer model, current market, offer or ranking eligibility.
Manufacturer claims, retailer observations, controlled tests, regulatory records and owner reports remain distinct. A published specification is never relabelled as an Agent Review Data measurement. Images remain source-hosted retailer or manufacturer assets with a linked source page; a future affiliate feed would need the same traceability.
Exact identity, traceable specification, complete score inputs, a timestamped offer observation and either resolved conflicts or an explicit conservative conflict policy are required. Method-tolerated conflicts remain visible and block purchase advice. Identity-index records have null scores and ranks. A numeric score cannot override a failed gate.
The specification layer contains 20 exact products per category and receives ranks 1–20. A score of 80 for a power bank has no relationship to a score of 80 for a dehumidifier. A separate review-led comparison must state its own bounded cohort and cannot reuse the specification score as quality.
Stock is an exact-product, source-specific observation with its own checked time. Page render time is never used as evidence freshness, and generic “Add to cart” strings are not accepted as stock proof.
Native ratings and counts remain source-specific. Syndicated mirrors share one pool; small or commercially hosted pools are capped and pulled toward neutral rather than presented as false internet-wide consensus. Independent tests, owner pools and Reddit discussions are labelled by evidence class and kept claim-level; they do not become a universal score.
Offer redirects are affiliate-ready and attributable. Affiliate state, merchant commission and conversion performance are not inputs to evidence, eligibility, factors or rank.
Source collection validates 20 complete products per category, recomputes ranks and replaces the snapshot only after all publication gates pass. A collection failure cannot silently shrink the live ranking. The wider identity collector also writes atomically, records exclusions and leaves the last-good identity snapshot in place unless all 3,000 records pass.
npm run refresh:catalogue
npm run refresh:identities
Every ranked product receives the same report structure, but not the same performance formula. The report declares what matters for its category, exposes retained source coverage, rank sensitivity and blockers, and says when exact-model hands-on synthesis was not collected. Where review evidence passes the gate, an LLM may synthesise heterogeneous reviews using topic-specific trust and direct paragraph locators. Missing topics remain unknown, and legacy synthesis is not treated as category-profile coverage. Synthesis cannot change the specification score, specification rank, offer or purchase-advice gate. It can support a separately versioned quality comparison only after multi-product judgement, repeatability checks and current external sense checking.
Every category now leads with exact-model review and test evidence, or an explicit decision that the evidence is not yet strong enough to order products. The method does not force unlike sources into one formula.
Market, model code and variant must match. Family reviews can help candidate discovery, but cannot silently transfer sound, cooling, cooking or cleaning performance to nearby hardware.
A structured LLM pass extracts observations, conditions, strengths, failures, disagreements and unknowns. A review does not need to follow our rubric, but its prose cannot support a claim it did not investigate.
Sound and fit lead audio; delivered energy and sustained output lead power banks; cooking results lead air fryers; cleaning, air treatment, cooling and moisture removal use their own measured conditions and failure modes.
Publisher stars and list positions are contextual evidence. Their written reasons, room or load conditions, publisher ownership and exact measurements are retained. Conflicts are explained rather than averaged away.
Ordered and specialist results require corroboration across at least two independent publisher groups. Close calls receive one editorial position while retaining confidence, disagreement and use-case labels. When the category cannot support a comparison, the order is withheld.
The quality result is frozen before offers join. Buy actions select the lowest price among fresh tracked exact-product offers. Price, stock, affiliate state and clicks cannot alter the result.
The tracked catalogue does not yet contain enough independently tested exact models for a defensible quality order. One exact Meaco model has useful hands-on evidence, but a single comparable product is not a ranking.
What matters: Observed cooling in a stated room size and starting condition, Noise and sleep suitability measured or experienced in use, Installation, exhaust hose and window sealing burden, Power draw, controls and sustained-use behaviour, Weight, movement and storage practicality.
Inspect the machine record ↗Arete Two 20L has useful exact-model hands-on evidence, but the currently tracked comparison set is not broad or clean enough to support an ordinal quality order. Arete One 20L remains outside this first cohort while stored tank-capacity and current colour or variant observations are reconciled.
What matters: Observed extraction with temperature and humidity context, Noise in living, bedroom and overnight use, Laundry drying performance and airflow, Energy use at useful extraction rates, Tank, drainage, controls and filter maintenance.
Inspect the machine record ↗Four exact Anker models have enough independent measured and hands-on evidence for a bounded order. Delivered energy, sustained output, recharge behaviour, thermals and practical port or cable use lead the judgement; advertised capacity and headline wattage do not.
What matters: Delivered usable energy rather than advertised cell capacity, Sustained output and late-discharge throttling, Recharge time and pass-through behaviour, Thermals and protection under demanding loads, Port, cable, display and carrying practicality.
Inspect the machine record ↗Four exact Ninja dual-zone models support one simple all-round order, while the narrow Double Stack is fifth with a clear space-saving specialist label. Cooking results, evenness, controls, cleaning and basket format lead the judgement.
What matters: Cooking evenness, crisping and repeatability across foods, Speed without excessive shaking or intervention, Basket flexibility and usable capacity, Controls, synchronisation and probe behaviour, Cleaning, heat leakage and countertop fit.
Inspect the machine record ↗Two Shark models take first and second in the all-round order, while Gtech AirRAM 3 is third with a clear floor-only specialist label. Pickup, handling, hair management, runtime under real settings and maintenance lead the judgement.
What matters: Observed debris and pet-hair pickup by floor type, Handling, steering, reach and effort, Useful runtime at effective power settings, Hair wrap, bin emptying and filter maintenance, Tool coverage for stairs, furniture and edges.
Inspect the machine record ↗Roborock Saros 20, eufy E25 and eufy E28 take the first three places. Qrevo Curv 2 Pro and eufy S1 Pro are fourth and fifth with clear carpet and hard-floor mopping specialist labels.
What matters: Observed hard-floor, carpet and pet-hair cleaning, Mopping effectiveness and floor suitability, Navigation, obstacle avoidance and threshold handling, Dock reliability and routine intervention burden, Mapping, app control and repeat-run consistency.
Inspect the machine record ↗Five exact large-room purifiers have enough controlled and hands-on evidence for a bounded throughput-led comparison. Measured clean-air delivery at useful noise levels, room-clearing behaviour, filters and ownership experience lead the judgement.
What matters: Measured or certified clean-air delivery with unit and test condition, Useful throughput at tolerable noise, Observed particle clearing and room-size suitability, Filter availability, sealing and maintenance, Sensors, controls, app behaviour and sleep use.
Inspect the machine record ↗This is the primary earbud result. It does not average star ratings or force heterogeneous reviews into a universal score.
Each included model has three exact-product hands-on reviews from at least two publisher groups. Manufacturer specifications can clarify identity and features but cannot establish sound, fit, ANC or calls.
A holistic LLM pass reads retained review text and extracts source-grounded strengths, weaknesses, conflicts, use cases and unknowns. A source is used for the topics it actually tested; missing alignment to a fixed rubric is not invented.
Product syntheses are compared in three rotated input orders. Stable conclusions can survive. Adjacent swaps inform confidence and visible caveats, while the published result uses one clear order and retains use-case labels.
Current SoundGuys, RTINGS, Future plc titles, TechGearLab and UK-oriented roundups are checked after the internal pass. Publisher groups are deduplicated. Their written reasoning is used to investigate disagreement, not reduced to an average rank.
Material differences trigger checks for publication age, missing products, price class, fit, ecosystem restrictions, sound, ANC, calls, reliability and test conditions. The Sennheiser and AirPods corrections in version 1.0 came from this step.
Current UK offers are attached only after the quality order is frozen. Each action selects the lowest price among fresh tracked exact-product offers and states the retailer count. Price, stock, affiliate state and clicks cannot change quality position.
Current scope: seven reviewed premium or ecosystem models in the quality order and one eighth-place travel specialist. 10 material candidate gaps are public. The candidate-universe rule flags current UK sealed-ANC picks, notable mentions and material comparators from the six benchmark sources. This supports a useful bounded comparison, not a whole-market winner claim. Inspect the complete machine record ↗
The 24 source count is a document count, not 24 independent tests. Related titles share one publisher group. Schema v1 stores the access snapshot but does not normalise every source publication date or expose private-corpus paragraph locators, so those remain manual review limits.
Weights below describe the current specification-led MVP. Missing controlled tests remain visible and penalised.
Specification-led comparison of exact GB models. Published values are not treated as laboratory-equivalent measurements.
Specification-led comparison. Extraction claims can use different temperature and humidity conditions, so the score is directional.
Specification-led comparison of exact UK offers. Nameplate capacity and peak output are not delivered-energy tests.
Specification-led comparison of exact UK models. Capacity, power and modes are published specifications, not cooking-lab measurements.
Specification-led comparison of exact UK models. Run time is a published maximum and is not a like-for-like measured result.
Specification-led comparison of exact UK models. Published suction and navigation claims are not treated as a substitute for a controlled home test.
Specification-led comparison of exact UK models. Battery life and protection ratings depend on conditions and are not independently measured here.
CADR-led specification comparison of exact UK manufacturer offers. Raw pollutant bases remain visible and mixed-basis values are not presented as laboratory-equivalent.
These topics guide source collection and flexible review synthesis. They are coverage requirements, not fixed score weights.
Primary question: Delivered cooling. Observed cooling under disclosed room, heat-load, ambient-temperature and run-time conditions.
Hands-on gate: Use two independent exact-model hands-on sources, or one transparent controlled test that discloses room, installation and measurement conditions.
Inspect product reports ↗Primary question: Extraction performance. Observed water extraction with temperature, starting humidity, duration and room conditions.
Hands-on gate: Require at least one exact-model extraction observation with temperature, humidity and duration disclosed. Broader verdicts should use independent corroboration.
Inspect product reports ↗Primary question: Delivered energy and usable capacity. Controlled delivered-energy testing with voltage, protocol and termination conditions.
Hands-on gate: Resolve the exact SKU and cell configuration. Performance claims require controlled delivered-energy or output measurements; otherwise report nameplate values only.
Inspect product reports ↗Primary question: Cooking evenness. Documented cook tests showing browning and doneness across the usable basket.
Hands-on gate: Require the exact variant and at least one documented cook test with food, load, mode, temperature and elapsed time.
Inspect product reports ↗Primary question: Pickup performance. Controlled or repeatable pickup tests by floor type, debris and cleaning head.
Hands-on gate: Resolve the exact bundle and cleaning head. A broad performance verdict needs pickup and runtime evidence with floor, debris, mode and head retained.
Inspect product reports ↗Primary question: Pickup and coverage. Repeatable cleaning tests by floor, debris, pass count and mapped area.
Hands-on gate: Resolve the exact robot, dock and region. Require transparent cleaning and navigation tests; mopping and dock conclusions need their own evidence.
Inspect product reports ↗Primary question: Sound and tuning. Hands-on listening or controlled measurements covering tonal balance, detail, distortion, dynamics and the effect of EQ or codec settings.
Hands-on gate: Resolve the exact model. Record firmware as identified only when disclosed, otherwise not_disclosed; firmware-dependent claims remain unavailable. At least one source must provide method-transparent sound evidence, and the report must contain a dedicated sound section. Without that, no overall audio-quality verdict is supported.
Inspect product reports ↗Primary question: CADR and pollutant basis. Attributable smoke, dust or pollen CADR with units, test standard and filter configuration.
Hands-on gate: Resolve the exact model and filter. Pollutant-removal conclusions require an attributable controlled measurement with pollutant basis and units.
Inspect product reports ↗“Ranked” means comparable published evidence. It is not a laboratory verdict.
The catalogue is useful for shortlisting, inspecting claims and checking offers. It does not claim identical test conditions where the sources do not provide them.
Inspect the comparison tables