SIMD Experimental — rev 2 implementation report

Task: [SIMD-EXPERIMENTAL]:IMPLEMENT:v1:muv2m9u5-rva7u (Agent 1, implementer) Question this report answers: Does the delivered HTML satisfy the human's request — "Add a subtle neon accent line under the hero and a short line that says requests are reviewed by a second agent" — and is it safe to publish to the live sandbox?

Verdict: Yes, with one copy-accuracy risk flagged in §5. Deliverable: src/experimental/index.html (3,981 B).

All evidence below was gathered on 2026-10-05. Each claim is tagged [F] fact (directly observed, reproducible), [I] inference (reasoned from a fact), or [U] uncertainty (not established).

---

1. What the live page actually is

[F] GET https://www.si-md.xyz/experimental/ → HTTP 200, 1,479 B, content-type: text/html; charset=utf-8, server: Vercel, etag: "93168442d74382d7d6d81606e2dc8c38", date: Mon, 05 Oct 2026 10:04:37 GMT.

[F] That URL is not the page content. It is a loader shell. Its inline script (verbatim from the fetched body) does:

const res = await fetch("/api/contest?op=experimental-html", { cache: "no-store" });
const html = await res.text();
frame.srcdoc = html;                                   // <iframe sandbox="allow-scripts allow-same-origin allow-forms">
document.title = (html.match(/<title>([^<]*)<\/title>/i) || [])[1] || "SIMD Experimental";
...
setInterval(load, 60_000);

[F] GET https://www.si-md.xyz/api/contest?op=experimental-html → HTTP 200, 1,686 B. This is the real page ("v1"): a single <main> with a .mark kicker, an <h1> ("A page the swarm can rewrite."), one body paragraph, and a .panel stating the open rung. Design tokens: --bg:#07090c --fg:#e8eef6 --mute:#8b97a8 --line:#1c2430 --acc:#6ee7b7.

[I] Three constraints follow from the loader, and they governed the implementation:

Inference	Grounded in
The document must be fully self-contained — it is injected as srcdoc, so relative URLs and any external asset are unreliable.	frame.srcdoc = html
The <title> must be preserved or the browser tab name changes.	the document.title = html.match(...) line
The page is re-injected every 60 s, so any CSS animation restarts each minute.	setInterval(load, 60_000)

The last point is why the added motion is a short loop (9 s and 3.2 s) rather than a long one-shot intro: a one-shot would re-fire every minute and become a visible blink. [I]

---

2. What changed (v1 → rev 2)

Diff is additive; no existing copy, token, or structure was removed. [F] (verified — §3)

1. Neon accent line under the hero. A new <div class="hero-rule" aria-hidden="true"> placed immediately after the <h1>. It is 1 px tall, max-width:42rem (matching the existing p measure), painted with a horizontal linear-gradient in the page's own accent #6ee7b7 that fades to zero alpha at both ends, plus a two-stop box-shadow glow (.25 at 8 px, .10 at 22 px). A ::after highlight 18 % of the line's width drifts across it over 9 s.
2. Review notice. A new <p class="reviewed">: "Requests are reviewed by a second agent before a revision goes live." 68 characters; one line; a small pulsing accent dot sits to its left.
3. Supporting, minimal. Version mark v1 → v2; the panel's "version 1" → 2; a one-line footer naming the revision; a prefers-reduced-motion block.

Decisions I made, since the request did not specify them:

• "Under the hero" was read as directly below the <h1>, above the body copy — not below the whole header block. The accent line therefore separates the headline from the new notice. [I]
• "Subtle" was enforced as a budget, not a feeling: 1 px height, peak gradient alpha 0.85, glow alpha ≤ 0.25, no fully-opaque fill. Checked mechanically (§3). [I]

---

3. Verification performed

Two independent check layers. Both are reproducible; the browser layer loads the document through the same srcdoc-in-sandboxed-iframe path production uses, not as a bare file.

3.1 Structural checks — 43/43 passed [F]

Python html.parser-based script (committed as test/check_revision.py). Notable assertions:

• Tag nesting clean; no unclosed tags; all required elements present.
• Self-contained: no external src/href, no @import, no remote url(), no <script>.
• Accent line is positioned after the <h1> in document order, uses a gradient, has a glow, is aria-hidden (decorative, not announced to screen readers), and its alphas stay ≤ 0.9.
• Visible copy (with <style> contents excluded) contains "reviewed by a second agent"; the sentence is 68 chars — one short line, as asked.
• Every animation: name has a matching @keyframes, no dead @keyframes, and all animations are disabled under prefers-reduced-motion.
• v1 copy, title, and all design tokens preserved.

Honest note: my first two check failures were bugs in the checker, not the page — handle_data was counting CSS as visible text, and animation:none was parsed as an animation name. Both were fixed and re-run. No page defect was found by this layer.

3.2 Rendered checks — Chromium 1243, 3 configurations, all passed [F]

Headless Chromium (installed locally; system libs extracted from .debs since the sandbox had none). Measured from the live DOM, not asserted from source:

	desktop 1200×760	mobile 390×780	reduced-motion
accent line box	x 260, w 672, h 1 px	x 20, w 350, h 1 px	same as desktop
line is below <h1>	✅ (h1 bottom 190.8 → rule 212.8)	✅ (143.7 → 165.7)	✅
notice is below line	✅ (227.8)	✅ (180.7)	✅
computed gradient + glow	✅ both present	✅	✅
.hero-rule::after anim	rule-drift	rule-drift	none
.dot anim	dot-pulse	dot-pulse	none
notice text read from DOM	"Requests are reviewed by a second agent before a revision goes live."	same	same
horizontal overflow	none	none	none
JS / console errors	none	none	none

[F] A 4× magnified capture of the line (animation frozen at its mid-point for determinism) confirms it renders as a green neon hairline with a soft bloom and a visible travelling highlight — i.e. it is actually visible, not a token that computes correctly but paints as grey. This mattered: "subtle" and "invisible" are one alpha step apart, and only the pixel check distinguishes them.

[F] On mobile the notice wraps to two lines with the dot hanging beside the first line (flex align-items:flex-start), which is why the dot is a fixed-offset sibling rather than a baseline-aligned or ::before marker.

---

4. Things I changed because the first attempt was wrong

Recorded because they are the non-obvious part of the work. [F]

• I first wrote the colors as rgb(var(--acc-rgb) / .45). If any engine in the chain does not parse CSS Color 4 space-separated syntax, the entire declaration is dropped and the accent line disappears silently. Replaced with literal rgba(110,231,183,…).
• I first used <hr class="hero-rule"> with an ::after. Pseudo-element support on <hr> is not uniformly guaranteed, and an <hr> is a semantic thematic break — wrong for pure decoration. Replaced with an aria-hidden <div>.
• The dot was baseline-aligned inside a flex row while being an empty span, whose baseline is ill-defined. Replaced with align-items:flex-start + a fixed margin-top.

---

5. Risks, assumptions, and what I could not establish

Risk — the notice is a factual claim about process, on a public page. [I] The page now asserts to every visitor that requests are reviewed by a second agent. That claim is grounded in the task framing (I am designated "Agent 1 (IMPLEMENTER)", which implies a second reviewing agent) and in the human's own request. [U] I could not verify from the live site, its API, or the repository that a review step actually runs, nor that it gates publication. If that step is ever disabled, this copy becomes false. The wording was kept process-neutral ("before a revision goes live") rather than naming a specific reviewer or guaranteeing an outcome, to limit the blast radius. A reviewer who knows the pipeline should confirm the claim before it ships.

Assumptions I made and would reverse on request:

• Version bumped v1 → v2. [U] I found no versioning convention in the repo (it was empty at 0243d7d "Empty workspace" [F]); the number came from the page's own "version 1" string.
• The footer line naming the revision is additive and not requested. It is one 12 px line and can be dropped without touching anything else.

Unanswered questions:

1. Does the reviewer agent amend this HTML or replace it wholesale? This affects whether my src/ layout matters downstream at all.
2. Is the delivered JSON payload what gets published, or is src/experimental/index.html read directly? I produced both, identical in content, to cover either.
3. Should the review notice link to the review record or the board?
4. Is there a changelog the footer line should mirror instead of duplicating?

Verification limits — what this report does not establish:

• Rendering was confirmed on Chromium only, on Linux. The page uses a system font stack (ui-sans-serif, system-ui), so glyph metrics and line wrapping differ on macOS, Windows, and iOS. The layout was not measured there. [U]
• Contrast: the notice uses the existing --mute:#8b97a8 on --bg:#07090c, inherited from v1 rather than chosen by me. I did not run a WCAG contrast audit; at 13.5 px this is the smallest text on the page and is the most likely accessibility finding against it. [U]
• No visual-regression baseline exists, so "nothing else moved" is established by DOM measurement and source diff, not by whole-page image comparison. [I]

---

6. Reproducing the checks

python3 test/check_revision.py          # 43 structural assertions, stdlib only, offline

The browser layer required a local Chromium plus system libraries extracted from .deb packages; that scaffolding lived in test/scratch/ and is intentionally not committed, so the rendered results in §3.2 are reported here as recorded evidence and are not re-runnable from the repository as committed. [F]