Wappsite · Design review · 21 October 2026
Open the websites on your phone, then inspect how their mockups became working pages.
My preference in this sample is the current UAT Miro site. It has the most coherent visual design. The simple Lumen build reads better on mobile. The photographic Lumen build has more character, with visible assembly defects.
The businesses, briefs and completion states differ. This comparison cannot establish which architecture is better.
These are unchanged saved outputs. The new builds were republished to an isolated Cloudflare Pages project. The shared UAT deployment was not changed.
Current non-rearchitecture pipeline
My visual preference overall. The desk photography, serif headings and warm palette feel coherent. The product section is visibly unfinished.
This is the saved UAT preview from 21 October. Product photo placeholders and the toolkit collage may be functional test content. This pipeline already uses Impeccable refinement skills.
Earlier test, no successful ship receipt
The most distinctive Lumen direction. The assembled page has a duplicate bicycle, blurred image strips and hard seams. Mobile loses much of the scene.
The native visual gate rejected this assembly. The published review copy preserves the saved output without repairs.
Latest test, no successful ship receipt
The clearest mobile layout, but also the most generic. It drops the generated imagery and much of the mockup character.
This saved build contains only index.html. It references a missing font stylesheet. The native build workflow did not complete its ship gate.
The new local tests used OpenCode with Muse Spark. Their raster generation used OpenAI Imagegen 2.5 Sunburst. This review reused the saved outputs and generated no new website or image.
The old pipeline selects a brand toolkit, produces a page mockup, then prepares assets and builds the site.
The selected direction and the later internal build composition are separate saved steps. Neither final website passed its complete native ship workflow.
The selected direction and the later internal build composition are separate saved steps. Neither final website passed its complete native ship workflow.
All 11 retained source crops match their recorded mockup bounding boxes. All 11 prepared images are opaque. The examples below include an asset that never reached the current published build.
The source retains the desk scene. The prepared photo removes the framed lettering. The published file matches the prepared asset.
The source is only 445 × 61 pixels. The generated plate invents a much larger view. This asset is absent from the current published build.
A convincing photographic asset, with the same warm material language as the page.
The crop becomes a standalone material texture. This is an opaque image, not a transparent cutout.
The map becomes a blank plate for code overlays. The website supplies the pin and lettering.
Historical acceptance does not cover the images currently used. Seven assets passed review earlier. Later work replaced the website images. Six of the seven current prepared files also differ from their historical acceptance hashes.
Each row separates the reference crop, the current asset folder and the raster actually included in the saved website.
The shipped JPEG shows a hard horizontal dark blurred strip through the lower bicycle wheels. This strip is already in the raster, before page CSS. A flat near-black extension occupies the far-right edge of the shipped JPEG. The final page repeats a partial wheel from the separate hanging-sign plate beside the main bicycle. The raw accepted plate is 1536x1024. The current shipped derivative is a different generation, 1536x576 after cropping, extension, blur, and tonal changes.
The reference crop includes bench legs cut by the lower crop boundary and adjacent wall. The replacement supplies an isolated complete bench. The shipped cutout has both legs visible. Its alpha bounds touch every image edge after trimming, so there is no safety margin at seat ends. Actual alpha exists. The image has alpha values 0 through 254, not an opaque rectangular background. No broad halo is obvious in the inspected assembled cream panel.
The final JPEG has two hard rectangular darkened patches at the bridge and lower planter. These patches are baked into the raster. The replacement changes the tower/bridge scale and their positions. The full page shows a tall narrow gutter rather than the comp's continuous spatial scene.
The sign lettering and icon are absent in the raster, as expected for code overlays. The new sign frame is larger and warmer than the reference. The replacement includes a partial bicycle wheel and basket at its left edge. In the assembled page this duplicates the neighboring hero bicycle at the post.
The shipped ground is a different generation than current assets/plates/central-ground.png. Its heavy blur erases pavement detail. In the assembled page it forms a broad soft rectangular service block with a visible vertical boundary against the sharper right gutter.
The replacement removes the reference's short divider strokes, so those must be drawn in code. The pavement itself is text-free and fills its frame. Its sharp bright vertical reflections clash with the soft blurred service ground in the final assembly.
The shipped replacement is a very wide blank plaster wall, with no baked text or bench. It is suitable for independent text and bench overlays. The final assembly keeps the contact band narrower than the approved comp and changes its proportions.
The new photographic build contains a real transparent PNG. It isolates the bench and keeps both legs. The trim reaches all four image edges, leaving no safety margin. Inspect the same file against light and dark backgrounds below.
The current Miro sample has no equivalent transparent asset. There is no fair cutout winner from these runs. Customer-uploaded photo preservation was not exercised end to end in the new builds.
Run the same brief, chosen direction and uploaded photos through both pipelines. Keep the reviewed asset hashes bound to the final build, then compare mobile and desktop results. That test can distinguish architecture effects from different design choices.