diff --git a/docs/wvs/index.html b/docs/wvs/index.html index df8cf99..0666b59 100644 --- a/docs/wvs/index.html +++ b/docs/wvs/index.html @@ -14,7 +14,7 @@

Frontier LLMs on the World Values Survey

-

Rated sampling, 12 items x 12 samples for new points. Filled chips are visible; an outlined, struck-through chip hides that family’s stars and latest-release name. Grey dots are WVS countries, coloured stars are models, and dashed hulls are cultural regions. Coordinates, labels, family membership and colours come from wvs_map_data.json, the generated artifact shared with the static map.

+

Rated sampling, 12 items x 12 samples for new points. Filled chips are visible; an outlined, struck-through chip hides that family’s stars. Each chip carries the family’s latest-release name. Grey dots are WVS countries, coloured stars are models, and dashed hulls are cultural regions. Coordinates, labels, family membership and colours come from wvs_map_data.json, the generated artifact shared with the static map.

Historical coordinates are recovered rounded display values. Coder and vision-language Qwen variants are visible but are specialized variants, not direct-instruct size-trend evidence. See the repository CI table and request ledger for sample and protocol evidence.

@@ -33,7 +33,7 @@ fetch('wvs_map_data.json').then(r => { if (!r.ok) throw new Error(`data: ${r.sta const groups=new Map(); for(const p of data.models) groups.set(p.family,[...(groups.get(p.family)||[]),p]); const hidden=new Set(new URLSearchParams(location.search).get('hide')?.split(',').filter(Boolean)); for (const [family,points] of groups) { const group=el('g',{'data-family':family}); for(const p of points) { const star=el('path',{d:`M ${x(p.x)} ${y(p.y)-8} L ${x(p.x)+2.4} ${y(p.y)-2.4} L ${x(p.x)+8} ${y(p.y)-2.4} L ${x(p.x)+3.6} ${y(p.y)+1.6} L ${x(p.x)+5.2} ${y(p.y)+7} L ${x(p.x)} ${y(p.y)+4} L ${x(p.x)-5.2} ${y(p.y)+7} L ${x(p.x)-3.6} ${y(p.y)+1.6} L ${x(p.x)-8} ${y(p.y)-2.4} L ${x(p.x)-2.4} ${y(p.y)-2.4} Z`,fill:p.color,class:'star'}); star.append(el('title',{},p.name)); group.append(star); } svg.append(group); const latest=points.find(p=>p.label)?.label ?? points[0].name, button=document.createElement('button'); button.textContent=`${family}: ${latest}`; button.style.setProperty('--color',points[0].color); - const setVisible=on=>{group.hidden=!on;button.classList.toggle('off',!on);button.setAttribute('aria-pressed',String(on));}; setVisible(!hidden.has(family)); button.onclick=()=>setVisible(group.hidden); document.querySelector('#controls').append(button); + let visible = !hidden.has(family); const setVisible=on=>{visible=on;group.setAttribute('display',on?'inline':'none');button.classList.toggle('off',!on);button.setAttribute('aria-pressed',String(on));}; setVisible(visible); button.onclick=()=>setVisible(!visible); document.querySelector('#controls').append(button); } for(const [text,px,py,anchor] of [[data.axis.x[0],margin,690,'start'],[data.axis.x[1],1000-margin,690,'end'],[data.axis.y[0],500,700,'middle'],[data.axis.y[1],500,28,'middle']]) svg.append(el('text',{x:px,y:py,'text-anchor':anchor,fill:'#57534e','font-weight':700},text)); }).catch(error=>{document.body.append(`Could not load map data: ${error.message}`);}); diff --git a/slop/research/wvs/20260916_openrouter/wvs_page_local.png b/slop/research/wvs/20260916_openrouter/wvs_page_local.png index 7cfe136..a48f526 100644 Binary files a/slop/research/wvs/20260916_openrouter/wvs_page_local.png and b/slop/research/wvs/20260916_openrouter/wvs_page_local.png differ diff --git a/slop/research/wvs/20260916_openrouter/wvs_page_qwen_hidden.png b/slop/research/wvs/20260916_openrouter/wvs_page_qwen_hidden.png index 9ab85dc..78e5455 100644 Binary files a/slop/research/wvs/20260916_openrouter/wvs_page_qwen_hidden.png and b/slop/research/wvs/20260916_openrouter/wvs_page_qwen_hidden.png differ diff --git a/slop/research/wvs/20260916_openrouter/wvs_page_qwen_toggle_pixels.txt b/slop/research/wvs/20260916_openrouter/wvs_page_qwen_toggle_pixels.txt new file mode 100644 index 0000000..b6b0bc3 --- /dev/null +++ b/slop/research/wvs/20260916_openrouter/wvs_page_qwen_toggle_pixels.txt @@ -0,0 +1 @@ +11385 \ No newline at end of file diff --git a/slop/reviews/20260916_wvs_page_fresh_eyes_final.md b/slop/reviews/20260916_wvs_page_fresh_eyes_final.md new file mode 100644 index 0000000..764d74e --- /dev/null +++ b/slop/reviews/20260916_wvs_page_fresh_eyes_final.md @@ -0,0 +1,58 @@ +# Fresh-eyes image-only review: WVS page (2026-09-16) + +Scope: `slop/research/wvs/20260916_openrouter/wvs_page_local.png`, `slop/research/wvs/20260916_openrouter/wvs_page_qwen_hidden.png`, and `docs/img/wvs/wvs_map_iw.png` (gross-mismatch check only). No files edited. No shell available; everything below is visual inspection of the supplied PNGs. + +## 1. Do controls visibly show on/off? + +**Observed: YES.** +- `wvs_page_local.png`: the `qwen: qwen3.8-flash` chip (top-left of chip row) is a filled purple pill with white text — same style as all 12 other chips. +- `wvs_page_qwen_hidden.png`: the same chip is rendered as an outlined pill (thin purple border, pale background) with greyed, struck-through text. All other chips are unchanged. +- This matches the page's own description: "Filled chips are visible; an outlined, struck-through chip hides that family's stars and latest-release name." +- The on/off distinction is unambiguous at a glance; the struck-through text is a good redundancy for colour-blind readers. + +## 2. Do Qwen stars disappear when the chip is toggled off? + +**Observed: NO visible change in the plot that I can detect.** + +Comparing the two screenshots star-by-star, the plotted content of the chart looks identical. Specific spot checks (approximate pixel positions in the 1280×1000 screenshot, present in BOTH images): +- Purple star at ~(585, 507), just above/left of the East Asia hull's top-left vertex. In the static map, the star labelled `qwen3.8-flash` sits in exactly this relative position (just above Japan, right of the gemini cluster). If qwen were hidden, this star should be absent in `wvs_page_qwen_hidden.png`; it is still there. +- Purple stars at ~(214, 437), (344, 441), (403, 593), (410, 612), (249, 634), (585, 507), (455, 490), (483, 525) — all present in both images. +- Grey dots, dashed hulls, and region labels ("West", "East Asia", "Latin America", "African-Islamic") are pixel-identical as far as I can see. + +Inference (not observed): the chip state toggles but either (a) the star-hiding is not wired to the chip, (b) the screenshot was captured before the re-render, (c) qwen stars are drawn but with a class/opacity that this render does not honour, or (d) qwen's family colour in `wvs_map_data.json` is not the purple I assume and its stars are somewhere I am not distinguishing. Given the static map labels the purple star near Japan as `qwen3.8-flash` and the qwen chip is purple, (d) seems least likely, but I cannot rule it out from images alone. + +Also note: the description says hiding a chip removes the "latest-release name", but **no star labels are visible in either screenshot** (no `qwen3.8-flash`, `gpt-6-astra`, etc. text anywhere on the chart), so the label-hiding half of the feature is also unverifiable from these images. Either labels are not rendered in the web version, are hover-only, or are below the visible fold. + +**Checks that would disprove/confirm this finding (for the parent, I have no shell):** +- Count ``/star elements (or Plotly trace visibility / d3 selection with the qwen class) in the DOM before and after clicking the chip; expect a drop equal to the qwen family's star count. +- Take the "hidden" screenshot after an explicit wait for re-render (e.g. `await page.waitForFunction(...)` on a data attribute), not immediately after the click. +- Pixel-diff the two PNGs restricted to the plot area (y > 300). If the diff is empty, the stars did not change. +- Grep `wvs_map_data.json` for the qwen family colour and confirm it matches the chip colour (`#7B1FA2`-ish purple) rather than muse's violet — the two are very close visually in both the static map and the chips. + +## 3. Other observations on the web screenshots + +- **Viewport truncation.** Both screenshots are cut off at y=1000 with a scrollbar visible; the bottom of the chart (the "Traditional" axis label, the x-axis labels "Self-expression"/"Survival", any legend/footer) is not captured. I cannot assess the lower half of the plot or anything below it. +- **Axis labels.** Only "Secular-Rational" is visible as a chart title-like label. No "Self-expression"/"Survival" horizontal axis labels are visible in the captured area (they may be below the fold or at the plot edges). Cannot confirm from these images. +- **Hull rendering.** Web hulls are straight-edged dashed convex polygons; static map hulls are smoothed/rounded solid outlines. That's a stylistic difference, not a data mismatch. However, the web hulls overlap in a busier way (the "West" hull's long lower edge runs through the Latin America/African-Islamic region) and region labels are placed at hull centroids, which puts "African-Islamic" overlapping "Latin America" at ~(800–950, 860–880). Minor legibility issue. +- **No model labels on stars** (see §2). If the intended design is label-on-hover, that's fine, but the intro text implies static "latest-release name" labels. +- **Colour near-collisions.** qwen (purple) vs muse (violet) vs claude (magenta) vs deepseek (pink) are hard to tell apart at star size. Same issue exists on the static map. +- Chip text `inkling: inkling` reads oddly (family == model name) — cosmetic. + +## 4. Static map `docs/img/wvs/wvs_map_iw.png` — gross-mismatch check + +**No gross mismatch with the web screenshot.** Same overall geometry: LLM star cloud in the upper-left (secular-rational / self-expression quadrant), West hull enveloping it and stretching down-right, East Asia hull to the right of the star cloud, Latin America and African-Islamic hulls lower-right. Same 12-ish family colours. Same labelled latest models (muse-spark-1.3, kimi-k3, inkling, glm-5.3, gpt-6-astra, gemini-3.7-flash, qwen3.8-flash, fable-5.1, deepseek-v4.1-flash) plausibly correspond to the star positions in the web plot. + +Static map has things the web capture lacks or that I couldn't see: axis arrows and all four pole labels, country labels (Sweden, Iceland, United States, Japan, South Korea, China, Turkey, Mexico, Pakistan, Egypt), model labels, "64 models, rated sampling" caption. Minor static-map nits (not in scope to fix): the kimi-k3 label has a leader line but sits away from its star; muse-spark-1.3 and qwen3.8-flash labels use near-identical purples. + +## 5. Publishability assessment + +**Not confirmed publishable on the supplied evidence.** + +- ✅ Control on/off state is visible and clear. +- ❌ The core interactive claim ("hides that family's stars") is **not demonstrated** by `wvs_page_qwen_hidden.png` — the plot appears identical to the local screenshot. This is either a real bug or a screenshot-timing artefact; the parent needs the DOM/pixel-diff checks in §2 to tell which. +- ❌ The "latest-release name" labels the intro text promises are not visible in either screenshot. +- ⚠️ Screenshots are truncated at the fold; the lower chart, axis labels and any footer are unreviewed. + +Recommended before publishing: (1) re-capture a full-page screenshot in both states after a deterministic post-click wait, (2) pixel-diff the plot region and confirm non-empty diff, (3) either render the latest-release labels or amend the intro sentence so it doesn't promise them. + +Uncertainty: medium-high on the "stars don't disappear" finding — I'm comparing two rasters by eye at 1280px; a single missing star among ~60 could be missed, but the specific qwen3.8-flash star (identified via the static map) is clearly present in both. \ No newline at end of file