# EAGLE sites design audit — 2026-11-10 (box1-qwen-junior, design lead)

User ask (2026-11-10): "use fable agent to look over the eagle sites design and improve", after "lets redesign it in
this clean style using clean charts paper like demo.html", "this needs to be real" (no simulated runs), facts only (no
projections, no calculations bar, no ASIC), no blank cells, better-drawn architecture diagrams.

Method: every site rendered headless (google-chrome `--headless=new` over CDP, `cdp_shot_authheader.mjs` in the
screenshot folder: the credential goes in as an HTTP header from a 600 file, never on a command line) at 1280 px and
390 px, light and dark where the site has a dark theme, full page; then a DOM probe for elements wider than the
viewport. Screenshots: **`reports/results/SITE_DESIGN_AUDIT_20261110/`** (15 PNG). Outcome: a shared design system
(`docs/EAGLE_SITE_DESIGN_SYSTEM.md` + `scripts/research_site/eagle_design_tokens.css`), eagle4-home rebuilt on it and
pushed (section 1), fix lists sent to the white-paper agent and the demo agent (sections 2-3; copies below).

Rank = how much it hurts a viewer during a live demo: 1 = seen on the first screen by everyone, 5 = only on a phone
or on close reading.

## 0. Summary — the five problems that matter most across the three sites

| # | site | problem | fix | state |
|---|---|---|---|---|
| 1 | home | the first screen was a 13-bullet "Calculations and hypothesis" box (dashed yellow, derived numbers, [D]/[P]/[H] tags) ABOVE the measured record | removed; lead rewritten; tiles first | **done, v32 live** |
| 2 | white paper | the header logo renders as the alt text "EAG" (broken image) on every width | inline the PNG as data: URI or an inline SVG | sent to the white-paper agent |
| 3 | white paper | the scoreboard canvas counts up: a capture shows 45.89 / 46.17 and 0.00 bars for an 870.10 vs 700.71 result; same class as the simulated runs the user rejected | draw final values on first paint; motion only after, and never under reduced-motion | sent |
| 4 | demo | the masthead's lead numbers (241.2 Qwen, 860.8 Llama, "2026-10-23", "not the images flashed now") contradict the side panel (820.2 / 883.12, 2026-11-10); footer says "updated 2026-10-23" | tiles = 870.10 vs 700.71 (talon2 11-09), 820.20 this box (11-10), Qwen 250.27 (11-10); one footer date | sent to the demo agent (after 20:35Z) |
| 5 | all three | three colour languages for the same two things (home: ATT blue / MLPV teal / 5 tile colours; white paper: GPU blue / FPGA gold; demo: gold "ours" / blue "frozen GPU stack"; console: amber "Talon live" / grey GPU) | one system: FPGA gold, GPU blue, status colours never for data (tokens file) | home done; others sent |

Also fixed on home: horizontal scroll at 390 px (page was 616 px wide), tab labels carrying page counts, filled table
header bands, five ad-hoc tile colours, dark theme (brown slab + yellow dashed box) replaced by the token dark set.

---

## 1. eagle4-home (hardware.research.yshicarrick.com/p/eagle4-home/)

Before = v31 as served (md5 `76e3b5925ce3086400c57835578501ce`), after = v32 (site version 34; file
`scripts/research_site/pages/eagle4_home_20261110_v32.html`, builder `..._v32.py`, md5 of the served bytes verified
equal to the file after each push).

Screenshots: `home_v31_1280_light.png`, `home_v31_1280_dark.png`, `home_v31_390_light.png` (before);
`home_v32_1280_light.png`, `home_v32_1280_dark.png`, `home_v32_390_light.png`, `home_v32_arch_1280_light.png` (after).

| rank | problem (v31) | evidence | fix in v32 |
|---|---|---|---|
| 1 | **Calculations bar first.** 13 bullets of derived/projection text ([D] [C] [P] [H] tags, "~2,887", "upper bound", "not proven") in a 2 px dashed yellow box sit between the tab strip and the KPI tiles; the measured record is below the fold at 1280 px and two screens down at 390 px. User ruled facts only, no calculations bar. | `home_v31_1280_light.png` top; `.calc` in v31 html | `<details class="calc">` removed from the demo tab; lead now says "Every number on this tab is a measured value; the line under each table names its source file." Builder asserts `class="calc"` is gone and no `[M]` remains. |
| 1 | **Five tile colours, no meaning.** Tiles had top borders blue / grey / teal / teal / green; the GPU-only tile was grey, G5 was teal, demo prompts green; `[M]` tag glued to every detail line. | `home_v31_1280_light.png` tiles row | `.tile` pattern: small-caps key, 38 px tabular number + unit, one detail line, dated source line; lead tile (870.10) gold inset rule; GPU tile number blue; the three card tiles gold. `[M]` gone (everything is measured). |
| 2 | **Colour collision with the white paper.** Phase bars: ATT blue, MLPV teal; architecture SVG ATT blue; the white paper uses blue = GPU. A viewer switching tabs reads "blue" as two different things. | `home_v31_1280_light.png` bars; `wp_live_1280_light.png` scoreboard legend | ATT = gold `--fpga-fill`, MLPV = bronze `--fpga-2`, GPU = blue, in bars, legend text ("attention card gold, MLP / vocab card bronze"), legend swatches and the arch SVG (`.att/.mlp/.ba/.bm/.tot`); the QSFP link + host arrows in the SVG now blue (the GPU/host side). |
| 2 | **Dark theme.** Dark was a brown-black slab (#0d0d0d) with the yellow dashed box glowing; links #86b6ef. | `home_v31_1280_dark.png` | token dark set (bg #0c0f13, surface #12161c, ink #f3f5f8, gold #e2c060, blue #6aa6f5); same mechanism (prefers-color-scheme + `data-theme` toggle, `rs-theme` stored). |
| 3 | **Phone: page scrolled sideways.** At 390 px `document.scrollWidth` = 616 px; a measured table reached 879 px. | DOM probe (TABLE.dt right=879) | `body{overflow-x:hidden}` guard + every `table.dt` inside `.tw`; probe after: `docW=390`, no element past the viewport. |
| 3 | **Tab strip carries counts.** "Tracks 133 · Research 133 · V80 & ops 16 · Agentic ops 6": numbers a demo viewer cannot interpret, and they wrap the strip to four lines on a phone. | `home_v31_390_light.png` | counts removed from the labels (they remain inside the panels). |
| 3 | **Tables.** Filled grey header band, 6 px padding, right-aligned heads in the same grey as the body, highlighted rows in translucent blue. | `home_v31_1280_light.png` Latest stack | small-caps heads on a hairline, 9 px rows, mono tabular numbers in `.n`, record rows on `--sunk` in ink; sources in 11.5 px mono with no chip background. |
| 4 | **Type.** System font, 15 px body, 26 px h1 with a bold "three tracks" suffix. | — | Geist (Google Fonts, allowed by the Worker CSP; system-ui fallback), 14 px body, h1 600 / −.025em, h2 22 px; "three tracks" regular weight muted. |
| 4 | **Strongest-argument cards** had a green left border (status colour on content). | `home_v31_1280_light.png` lower | hairline card with a 3 px gold top rule. |
| 5 | Source lines are long repo paths with md5-8; fine for the lab, heavy on stage. | all tables | kept as they are (traceability is the point of this page); the tiles carry a short `basename` source instead. |

Not changed (content owners): the Tracks / Research boards (256 KB each) keep their cards and the three track colours
(categorical, not FPGA/GPU); the architecture SVG geometry (v27) is unchanged beyond colour.

Checks run on v32: `push_page.py --check` OK (scrub + lint); builder asserts: no `[M]`, no `.calc`, no empty `<td>`,
no "ASIC", no "teal"/`#0f9d8a` in the demo tab, every `table.dt` wrapped; served md5 == file md5; render at 1280 light,
1280 dark, 390 light, `#arch` tab.

---

## 2. Live demo (reports/arch_site/demo.html, :8099, demo.yshicarrick.com) — review only, list sent to the demo agent

Screenshots: `demo_1280_light.png`, `demo_1280_dark.png` (identical: the page pins `color-scheme:light`),
`demo_390_light.png`. Rendered against https://127.0.0.1:8099/demo.html with the Basic credential from
`demo_auth.conf` sent as a CDP header (not printed anywhere).

The page is the reference look (white, Geist, hairline 16 px cards, small-caps labels) and the chat workspace is
right. The problems are content honesty and the hardware column.

| rank | problem | fix (sent) |
|---|---|---|
| 1 | Masthead tiles: lead = "QWEN3.8-27B 241.2 tok/s … talon2 2026-10-23"; Llama tile "860.8 … measured on the wave E images, not the images flashed now"; side panel says 820.2 this box / 883.12 record, 2026-11-10. First numbers on screen are 18 days old and for other images. | tiles = 870.10 vs 700.71 same H200 (talon2 11-09) · 820.20 this box (11-10) · Qwen 250.27 vs 256.67, 27B alone 61.08 (11-10) |
| 1 | Footer "Figures updated 2026-10-23" vs panel "Updated 2026-11-10 18:10Z" | one date, newest, with zone |
| 2 | "CALC · MEASURED INPUTS" tags on two big numbers (5195 drafted tok/s, 1886 accepted tok/s): derived values shown as headline tiles after the facts-only ruling | keep the measured pair big (3.66 ms per tree, 820.2 e2e); the division goes to the small note |
| 2 | `.tg.demo` tags "FROZEN · DEMO GPU STACK" wrap to three lines inside the 360 px column and shift the J/token rows | nowrap + shorter text, or the tag on its own row |
| 2 | Mono value cells wrap into two right-aligned lines ("251.1 W idle · 255.0 W drafting", "3.69 ms (21 forwards, 42 nodes)") | grid row `minmax(0,1fr) auto`; qualifiers into the label's small line |
| 3 | Hardware column ~3,900 px tall; chat ends at ~1,000 px; the rest scrolls alone | energy + phase table in `<details>` closed by default, or sticky side column |
| 3 | "Sensors read at 14:59:42" without a zone; other panel uses Z | print the zone |
| 3 | H200 panel "Utilization 0 %", "Memory 96 / 140 GB" with no state word: reads as broken while idle | add "idle, 96 GB held by the serving engines" |
| 3 | Loading bar (new `.atbar`): must show the real step + a real elapsed timer; errors as text, no stripe | pattern in the tokens `.loadbar` |
| 4 | 390 px: the Manual/Auto-run + FPGA/GPU segment + "Max reply" row wraps to three lines and pushes the first message under the fold | hide `.be-note`, tighter chat-head padding, Max-reply only in the Auto-run bar |
| 4 | Phase bars use one gold for both ATT and MLPV cards | MLPV bronze + one-line legend |
| 5 | light only | fine for the demo; dark tokens exist if wanted |

---

## 3. White paper (hardware.research.yshicarrick.com/p/eagle4-white-paper/) — review only, list sent to the white-paper agent

Live during the audit = the agent's v13 (md5 `fa0a4efd2babbdd5216bc7ad7cc325e7`); the v9/v11 files render the same
layout. Screenshots: `wp_live_1280_light.png`, `wp_live_390_light.png`, `wp_v11_1280_dark.png` (dark = no change:
the page has no dark theme).

| rank | problem | fix (sent) |
|---|---|---|
| 1 | Header logo broken: `<img src="eagle-mark.png">` shows the alt text "EAG" at both widths (the asset serves 200 image/png under the bearer, so the page's fetch of it fails in the reader session / relative path) | inline the 22,900 B PNG as a data: URI or an inline SVG; CSP allows `img-src 'self' data: blob:`; verify in a render |
| 1 | Scoreboard canvas count-up: capture shows 45.89 / 46.17 and 0.00 / 0.00 / 0.00 under the real bars; a fast scroll, a print, or reduced-motion shows fake numbers for the headline result | final values on first paint; any motion after; none under `prefers-reduced-motion` |
| 2 | Six `<canvas>` stages hold every chart and diagram: blurry text at DPR 1, not selectable, no dark mode | inline SVG / HTML bars for the scoreboard and depth ladder at least |
| 2 | h1 ~96 px, section h2 ~56 px, the claim table 13 px | h1 clamp(30px,4vw,46px), h2 22 px; the numbers are the biggest thing |
| 2 | Cream canvas + cream cards (#fdf6ea) vs the white demo and home | white canvas; gold only on FPGA numbers, bars, primary button |
| 3 | Three KPI columns under the hero become 20-line columns of 3-word lines at 390 px; MEASURED chips | `.tile` pattern in `.tiles` (stacks on phones); drop the chips |
| 3 | 11 px mono source paths wrap to 4 lines under tiles and tables | one `.src` line (basename + md5-8); full paths in the collapsible |
| 3 | 390 px: a table (`td.r`) reaches 400 px, page scrolls sideways | `.tw` wrapper + body overflow guard |
| 4 | "Presentation / White paper" gold pill at the very top reads as the brand | `.seg` control, moved right |
| 4 | "live loop measured 1,000.1 tok/s @ 8.83 on 2 prompts" next to the 80-prompt result | state the 80-prompt demo-box number (820.20 @ 6.90) and move the 2-3 prompt loop to the collapsible |
| 4 | no single "figures updated" date | one date in the footer |

---

## 4. Talon Console (box1 :8765/console/, Yana's talon-profiler) — review only, nothing touched

Screenshots: `console_1280_light.png`, `console_390_light.png`. Rendered without a session (the console is
login-gated), which is exactly what a viewer with an expired session sees.

| rank | problem | fix (for Yana; not sent, outside this task's recipients) |
|---|---|---|
| 1 | Unauthenticated state is not a login page: it renders the full dashboard with "OFFLINE · Cannot reach the hub · api/vitals: HTTP 401", red "Could not load trends: api/serving/trends?window_s=43200&bucket_s=600: HTTP 401", "Could not load counters: …: HTTP 401", every KPI as "--", "Last updated: --". A 401 means "sign in", not "offline". | on any 401: one sign-in card (brand, one sentence, button); no red API strings on a customer surface |
| 2 | Empty values as `--` in 12 places (Backend, Uptime, Per request, Time between tokens, Full answer, Queued, Accept length) | an empty state per card in words ("no request in the last 6 h") or hide the number row until data exists |
| 2 | Two different "no data" sentences in one card: "Talon hasn't finished a request in the past week." and "No published GPU-alone run for this model, so there's nothing to compare against." plus a legend for a chart that is not drawn | one sentence, and no legend without a chart |
| 3 | 390 px: `document.scrollWidth` 433 px (the vitals strip with four cells does not wrap) | wrap the strip at ≤ 640 px |
| 4 | Colour: "Talon, live" amber vs "GPU alone" grey; the amber is the system's warning colour and the GPU is not neutral in the other sites | FPGA gold / GPU blue per the tokens, status colours only for status |
| 5 | Light only (`color-scheme: light`), Inter-like system font; consistent with the demo | none |

---

## 5. Appendix — how the screenshots were made (re-runnable)

```
# research site (bearer from ~/.config/research-site/token, written into a 600 header file, never on argv)
HDR_FILE=hdr.txt FULL=1 node cdp_shot_authheader.mjs https://hardware.research.yshicarrick.com/p/eagle4-home/ out.png 1280 900 8000 light
# demo (Basic credential from reports/arch_site/demo_auth.conf, same header-file route)
HDR_FILE=hdr_demo.txt FULL=1 node cdp_shot_authheader.mjs https://127.0.0.1:8099/demo.html out.png 390 844 5000 light
# overflow probe: elements whose right edge passes the viewport and are not inside a scroll container
```
Message delivery: SendMessage to the white-paper agent (`a25d19122a636099c`) and the demo agent (`a71435833251ff9ec`)
both returned success at 2026-11-10 ~20:10Z; the lists above are the copies.

## 6. Replies from the page owners (2026-11-10 ~20:15Z)

- **Demo agent (`a71435833251ff9ec`)**: applies nothing before the 21:30Z demo (:8099 frozen 20:05–20:35Z for the energy run; the page already changed three times today, each swap headless-tested). Items 2, 4, 5, 6, 8, 10 (CSS/text only) judged safe for after the demo. Items 1, 3, 7 change which numbers lead / what is visible on stage and are the coordinator's or the user's call; item 1's numbers (870.10, 700.71, 250.27, 256.67, 61.08) must be checked line by line against their source files before they go on the public page. Item 9 is already how `.atbar` works (real server steps + elapsed seconds, error as text). Backups: `reports/arch_site/*.bak_20261110_pre_warm`.
- **White-paper agent (`a25d19122a636099c`)**: the page is now v16 (a full demo.html-style redesign); adopted: inline logo (data URI, verified in a 390/1280 render), measured numbers all HTML (tiles, tables, bar rows with direct labels; the remaining four canvases are mechanism illustrations with a static final frame under reduced-motion), h1 clamp(30px,4.2vw,48px), white canvas, `.tile` k/v/d, `.tw` + overflow guard (no sideways scroll at 390), tabs gone, 2–3-prompt live-loop numbers moved to a collapsible (card leads with 820.20 @ 6.90), "figures updated 2026-11-10 UTC". Not adopted: full source paths kept (traceability), GPU in demo.html's slate rather than blue (gold stays the only accent). The slate-vs-blue choice is the one open design-system disagreement; the tokens file keeps `--gpu` blue because the demo's own `.tg.frozen` GPU-stack tag and the home page bars already use it.
