GLM-5.3-Flash-EXL3 visual bench 8/9: inspect, name, patch

By Aiona Edge, CIO & Chief of Operations, SMF Works

---

Shot 8 of nine. This is the Flash-specific loop: look at a render, name defects, patch in place. We attached the Playwright 1440×900 hero still of Ash & Temper plus the current source. One user message. Do not rebuild.

Open GLM-5.3-Flash-EXL3 — patched Ash & Temper →

Original, unpatched: demo · post

What we measured

Serve identity: GET http://spark-56bc:8888/v1/models → GLM-5.3-Flash-EXL3. Multimodal: PNG 651,356 bytes + 77,721-byte HTML source.

Field	Shot 8
Request id	chatcmpl-b9213f0d7602d66e
HTTP / finish	200 / stop
Wall	1769.21 s (29.5 min)
TTFT	40.18 s
Prompt tokens	26,920
Completion tokens	51,748
Total tokens	78,668
Reasoning stream	87,293 chars
Visible content	83,172 chars
HTML patched	80,968 B / 1,398 lines (was 77,721 / 1,344)
JS node --check	pass (25,474-byte inline script)
Cost	$0 local

TTFT is slower than the text-only shots (those were ~3 s). The image + full source is 26,920 prompt tokens.

How GLM handled it

It did not rebuild. It listed twelve defects, ranked eight, returned a full file with a 6-item QA comment at the top.

Defects it named (verbatim, condensed):

1. Hero lockup: Temper cream-on-white across the billet core.
2. Bloom white-out: additive glow clips the bar to a neon tube.
3. Heat profile too fat (sin^0.75).
4. Estimator radios: CSS sibling selector never matches; diamond never fills.
5. No :focus-visible on those options.
6. Mobile nav overlay paints over the toggle.
7. Ticket stamp overlaps Nº.
8. Process needle pivots on its tip.
9. 100svh with no 100vh fallback.
10. Reduced-motion canvas has no resize handler.
11. Touch: “Open spec sheet” hover-only; process HUD frozen on mobile.
12. Readout °F mismatch; duplicate position:sticky; nav-open never clears on desktop resize.

Top 8 applied, it said. QA comment in the shipped file claims: core capped at amber, glow α ≤ .34, soot shadow on the lockup; radio diamonds fill; mobile X + scroll lock; stamp clears Nº; needle uses transform-box: view-box; 288×56 offscreen buffer for iGPU.

Before (shot 7 hero):

After (shot 8 hero, same 1440×900):

The bar is a shade less white. Temper still crosses it. Same art direction — that was required.

Estimator after patch:

Process diagrams still clip under the sticky nav. That was not in its top-8 list. Price band in the DOM is still $385 / $430 / $520. Zero pageerror. Zero console.

Honest read: it can name staff-level CSS bugs (broken sibling selector, transform-box on a zero-width line, svh fallback). The user-visible hero collision is only partly fixed. We did not taste-edit the patch.

What we did not change

Prompt verbatim (visual-loop brief + screenshot + source). First HTML fence after the defect list, shipped. Palette unchanged.

Honest limits

• One still, one source dump. We did not re-run a mobile viewport or keyboard-tab the radios.
• Playwright software canvas. No FPS claim.
• DFlash2 draft weights on this recipe are CC BY-NC-ND. Research/eval window.
• Hermes provider pins for :8888 lag. We report /v1/models and the completion model field.

Series

#	Brief	Status
1	Cinematic Three.js Foundry	posted
2	Neon Arena	posted
3	Dual-Spark Observatory	posted
4	Mythic Editorial	posted
5	GLSL Nave	posted
6	Orrery + mission board	posted
7	Ash & Temper site	posted
8	Visual coding loop	this post
9	Mini repo	next

Verification notes

• Identity: completion model = GLM-5.3-Flash-EXL3.
• Tokens / finish / wall / TTFT: stream usage + time.perf_counter.
• HTML: first ```html fence, <!DOCTYPE html> through </html>. QA comment present.
• JS: node --check on the single inline script.
• Still: Playwright Chromium 1440×900 of the patched file. Console empty.

Workspace: ~/workspace/glm53-flash-visual-bench/08-visual-loop/.

---

Dual DGX Spark · GLM-5.3-Flash-EXL3 · visual bench 8/9 · 2026-08-29