Skip to content

Latest commit

 

History

History
495 lines (444 loc) · 33.3 KB

ui-design-reference.md

File metadata and controls

495 lines (444 loc) · 33.3 KB

Signal Deck design language

Signal Deck is the shared design standard for the dashboard. This reference governs the appearance of ROCm Readiness and Upstream Health, including their detail views. It supersedes the visual styling in the MVP1 prototype and shell plan. Use ROCm Readiness as the dashboard identity and keep source and repository attribution visible where it belongs.

Palette

Role Dark theme Light theme
Canvas Charcoal #1A1A1B Light gray #F3F4F5
Raised surface Graphite #27282B Light gray #E9EAEC
Primary text Soft white #DEDFE1 Graphite #27282B
Secondary text Gray #B2B4B7 Gray #55575B
Accent Blue #2899FF Blue #0051C6
Text on accent Black #000000 White #FFFFFF

Signal Deck takes its charcoal and graphite from the ROCm documentation header and its blues from the ROCm documentation theme. Use the accent for navigation, focus, and selected controls. Reuse the role tokens in the shared stylesheet. Keep the established palette in both themes. A palette change belongs in an explicit revision of this standard and its shared tokens.

Keep status colors separate from the accent. Use labels or distinct symbols with color for pass, fail, skipped, and unavailable states. Keep missing qualification evidence neutral. Qualification, test outcomes, and hardware validation each retain their own meaning. Use a 38% status-color tint in dark mode and 28% in light mode for pass, warning, and failure table cells so build outcomes are easy to scan against each theme's surfaces. Keep skipped and unavailable cells at a quieter 16% neutral tint, and retain the existing status foregrounds in both themes.

Layout and interaction

Phones and portrait tablets

  • At 640 CSS pixels and below, use compact phone headers: brand/engine context, navigation and theme controls, then the view title and a labeled filter icon. Keep sorting inside the expandable filters so results begin near the top. Expanded phone filters use one aligned column, full-width inputs and selects, paired option buttons, and a sort-direction button beside its select.
  • Performance measurements and qualification packages use separated square records on phones. Keep ROCm version, channel and build identity above their records; model names are prominent and may wrap. Preserve source metric units and explicit missing values. Machine specifications belong in evidence. At 540 CSS pixels and below, place full Tier 0/Tier 1 labels above their check strips, with spacing and dividers between tiers. Package evidence controls show a right/down chevron reflecting their expanded state. A card's evidence chevron also reflects open check details and collapses those details on activation. The selected check has a visible outline. At 540 CSS pixels and below, the shared evidence surface stacks each field label above its value and omits the separate Close button; its disclosure control or Escape collapses it. Performance cards reuse the same evidence surface and disclosure affordance.
  • Phone evidence expands in document flow, preserving the shared structured label/value format. Measurement cards use “Evidence & configuration”. Check and CI matrix evidence appears immediately below its owning card or scrolling matrix, with the selected check/job/date identified. Opening evidence requires activation; hover and focus do not open it. Closing returns focus to its control.
  • At 540 CSS pixels and below, overview calendars use full UTC date rows and platform columns; qualification trends use horizontal stacked bars in dated rows, newest first, with the same shared count scale. Upstream Overview uses date rows and category columns. CI Activity uses expandable date cards with horizontal bars, category selection, and run evidence inside the selected card. Project metrics use square summary tiles, four columns where space allows and two on phones. Qualification date rows expand their packages and evidence inside the selected row. CI jobs open individual dated outcome tables with their metadata and selected evidence. At wider widths, retain scrollable calendars and CI matrices. Do not let any table force page-wide horizontal scrolling.
  • Mobile Package qualification shows ROCm once above its groups. Stable builds have prominent build headings; nightly groups have prominent GPU-family headings and subordinate build metadata. Put the check legend beneath each build/family group's platform cards, with clear spacing between groups.
  • Collapse the navigation rail below 1280 CSS pixels, regardless of height or orientation. The header Navigation button reveals the same workspace/view hierarchy above the content. Use nested lists with modest indentation: workspace → Overview → Build Qualification → Performance, and Upstream health → Overview → CI Activity. Keep touch rows at least 44 CSS pixels high. Selection closes navigation, including selecting the current page. Escape closes it and returns focus to the Navigation button. At 1280 CSS pixels and above, retain the sidebar. Keep tables above the independent 640 CSS-pixel phone breakpoint.
  • Check light/dark themes at 320/390/430 phone widths, 768/834/1024 portrait tablet widths, and 1024/1194/1366 landscape tablet widths. Check keyboard, touch, text wrapping, URL-restored evidence and resizing as well as visual fit.

Shared layout conventions

Link arrows come from CSS with a subdued opacity. Do not append arrow glyphs to link text or add duplicate arrow elements. External links use the shared external indicator; tooltip navigation uses the shared internal indicator.

Ingestion warnings use square neutral surfaces, a restrained warning-color leading border, 1rem padding and at least 1rem separation from subsequent content in every view. Keep the heading prominent and retain all attempt, success and error details.

  • Share navigation, typography, spacing, and control styles across views. Use headings and spacing to show hierarchy.
  • Use the selected ROCm version and engine to scope the qualification history. Show one ROCm version at a time and default to the configured stable anchor. Show only the selected release’s track in the trend and identify the track and ROCm version for each package.
  • Use independent platform toggles with a filled and accessible pressed state. Keep at least one platform selected, and keep unavailable platforms visible and disabled. Determine platform availability from the version and engine before applying search or date paging. Keep filter selections in the URL. Pair platform labels with the Windows mark or Linux Tux in filters and package headings. Selected filters use a subtle hover tint while retaining their selected fill and text contrast.
  • Group version and platform controls together, with search at the trailing edge when space permits. Let the controls wrap at narrow widths. Use the workspace rail for navigation between dashboard views. Place ROCm version selection with the view filters. Show short version lists as visible choices. Keep the current title in the shared header and the theme control at its trailing edge. Label the selected date and platform above the package history.
  • Put workspace navigation at the top of the desktop side rail and dataset provenance at the bottom. At narrow widths, keep navigation above the content and provenance below it. Nest view links under their workspace and use the shared filled selection style for the current view. Nest Performance under Build Qualification. Keep keyboard focus visually distinct from the selected state.
  • Make the ROCm version the primary heading and place engine names above the timeline surface. In Upstream Health, group project metrics separately from the dated activity table. Place the CI legend beneath its own calendar.
  • Lay out related cards side by side when space permits and stack them on narrow screens. Keep timeline dates and build links readable at either width.
  • Preserve System, Light, and Dark preferences. Check text and controls against their actual backgrounds in both themes. Use at least 4.5:1 contrast for normal text and 3:1 for meaningful controls, icons, and focus indicators.
  • Verify visible keyboard focus, touch targets, and tooltip access by keyboard and touch. Respect reduced-motion preferences.

Visual hierarchy

  • Keep the technical character in the layout. Use compact sections, aligned data, and monospace for instrument labels, engine names, and comparable numbers. Use the shared sans-serif typography for prose and navigation.
  • Separate related areas with space or a subtle change in surface tone. Use rules within data tables where they help track rows or date groups. Keep panel fills close to the canvas and reserve decorative treatment for the shared page header.
  • Give parent sections a larger heading and more space above and below it. Place a short label beside a nested technical identifier when its role would otherwise be unclear. In package qualification, keep the build heading above its Nightly packages, with GPU family before the OS columns. For Stable shared archives, use a single table with Build as the first column.
  • Keep header controls quiet. Use a borderless theme icon with a visible focus state and make the active destination clear.
  • Keep version and project headings prominent. Give the engine label enough weight to scan above the timeline panel. Show Latest Stable for the configured anchor and Nightly for versions newer than it.
  • Keep the rail focused on dashboard navigation and dataset provenance. Treat additional destinations or selectors as separate design decisions.

Evidence and status

  • Place each calendar’s legend directly below its table so it explains the marks used there. Give calendar tables a visible heading above their shared, low-contrast surface. Use the shared heading style for the title, with the track, platform, or date window in smaller secondary text beside it. Keep headings and table labels available to assistive tools.

  • End the default ROCm Readiness calendars on the current UTC date, even when there are no builds that day. Mark today's date with stronger neutral type. Build Qualification uses the same current 14 UTC dates. Date links narrow that window; historical paging controls are omitted. Keep GPU-family names and their performance controls together on one line in a sufficiently wide column.

  • Group consecutive UTC dates under a spanning month name. Show only the day number in each column and carry a subtle divider through the month boundary. Apply this pattern to Readiness Overview and Upstream CI calendars. Preserve the full UTC date in the accessible labels.

  • Lead Build Qualification with daily stacked outcome bars for the selected release track. Put day numbers below the bars and month groups below the days. Use separate labeled Windows and Linux charts with the same package-count scale. Show publishing by OS beside each build.

  • Group packages by release workflow run, then GPU family when reported, then platform. List compiled Stable GPU targets by OS in the build-number popup. Show the runtime-validated target separately. Overview popups use a month/day/year and platform title. Keep the shared label-left/value-right layout, showing the latest build's qualification, job result, publication and produced timestamp. Include daily package and matched measurement counts without enumerating GPU families. Use one View qualification details link for the date/platform group. Keep channel, revision and internal build identity in detailed evidence. Date-only tooltip labels use full month/day/year without UTC; produced timestamps include date, time and UTC. Published archive filenames link to the corresponding release asset. Overview links preserve the selected package and its UTC date, version, track, engine, and platform. Use short platform or GPU package labels and compact tinted check cells with distinct glyphs. Use plain tier labels. Show each check’s pinned evidence in a tooltip anchored to its outcome block. Use Floating UI for anchored tooltips and interactive package popups, with an arrow to the trigger, viewport-aware positioning, keyboard and touch access, Escape and outside dismissal, and a visible close control for interactive panels. Table evidence popups open to the right and flip left when space requires it; keep them beside the row so readers can scan vertically between triggers. Keep the tooltip open while the pointer or keyboard focus is within its trigger or panel. Size package popups to their content within the viewport. Show visible platform packages in one compact column. Use top and bottom spacing, a shared heading typeface, and clearly weighted subsection headings to establish hierarchy. Separate platform sections with stronger horizontal rules and subsections with lighter horizontal rules. Avoid vertical heading rules and borders around nested navigation. Show package metadata, the pinned producer summary, and provenance links in the package popup. Keep full model names in performance tables. Tooltip links inherit the surrounding text color with a faint dotted underline and a small navigation arrow. Strengthen the underline on hover or keyboard focus. Performance glyphs open their own graph popup, separate from build metadata. Show horizontal Prefill TPS (pp_tps) and Decode TPS (tg_tps) bars on independent shared scales, retaining full model labels and values inside the graph tracks. Label both as higher is better, in tokens per second. Do not substitute TTFT or generic mean throughput when either throughput metric is absent. Use the latest observation per model and OS. Platform bars expose workload, hardware and timestamp context and link to that exact observation. Windows/Linux toggles can show either platform or paired bars. Gray out unmeasured platforms with an Unavailable hover label. Keep VRAM out of this preview. On mobile, stack models and metric groups vertically with full-width horizontal graph tracks. Keep Windows and Linux bars together under each metric, identify each bar with its OS icon, and retain values and independent metric scales. Separate the performance trigger from adjacent evidence controls with spacing only. Performance controls must not have a left selection stripe or vertical divider. Retain their disclosure chevron, reflecting the expanded state, accessible name, and keyboard focus indicator. Performance and Evidence & configuration disclosures have independent expanded states. Opening or closing either must not change the other's chevron or panel; the evidence control may still reflect and collapse qualification check details. Show the same matched-performance glyph on qualification build/family labels, and selected build headings. Only verified package matches establish the indicator. Pair OS packages only within the same qualification run, source, release track and GPU family. Show full model names, wrapping on narrow screens, and a visible scroll cue when more content is below the panel. List the visible platform packages and report links in the GPU-family panel, or under the platform heading for shared archives. Size check cells from the longest tier in the package and keep that width across its tier rows. Existing check links open the matching check tooltip for the exact package.

  • Compare benchmark results only within one compatible workload and hardware cohort using verified package matches.

  • Preserve the Readiness calendar's operating-system rows, month and day headers, and links to individual builds. In Upstream Health, keep metrics, coverage notes, and daily CI activity in distinct groups. Use the shared header for the dashboard title and lead with the CI summary in the body.

  • Use tinted status cells with distinct glyphs for table evidence. Show the latest build per engine, track, platform, and UTC date in Overview. Keep individual packages available in Build Qualification. Reserve stacked bars for trends. In Upstream Health, the shared legend defines rate bands and distinguishes requested jobs that never ran from skipped jobs and days without collected evidence.

  • Let wide tables scroll inside their own labeled, keyboard-focusable regions. Keep project and build drill-downs accessible by keyboard and touch.

Reviewing UI changes

Review the affected view in both themes at narrow and wide widths. Compare it with the other dashboard for shared navigation, spacing, and status conventions. Check table semantics, keyboard focus, touch targets, and evidence links where the change affects them. Update this reference when intentionally changing a shared convention.

Upstream Health mirrors the Readiness workspace navigation with nested Overview and CI Activity pages and the shared active-page treatment. Overview presents rolled-up project metrics and 14-day activity calendars in flat sections with primary-text headings, spacing and subtle rules. Use one project per row at all widths, with a prominent Projects section heading. Keep the summary compact; coverage explanations belong with the project evidence rather than above Projects. Do not use rounded project containers or generic View details controls. Project headings are neutral primary-text links to that project's CI Activity page; do not repeat navigation beneath the calendar. Keep repository/fork information and collection notices in CI Activity. Overview keeps aligned project metrics and clickable daily activity, retaining unknown/incomplete evidence semantics. CI Activity reuses Build Qualification's grouped filter controls and selected states: Project, Platform, Stack, and Target. Stack and Target are independent inclusive toggle groups: all included values use the filled selected state, without an Everything AMD control. Retain at least one value per group and keep both groups in the URL; legacy lens links remain supported. Disable platforms without configured jobs; retain at least one available platform. Preserve project/filter context when switching sidebar pages. Overview day links open the exact UTC date and category in CI Activity; retain that selection in the URL through reload and browser history. Keep legacy project-only URLs working. List source repositories directly in the compact project summary, without a repository disclosure or breadcrumb. Show incomplete collection as quiet project metadata while retaining collection failures as alerts. Keep Overview as the rate-band summary. CI Activity uses the same month groups, typography, table rules and scroll behavior, but uses Qualification trend’s stacked outcome bars for each category, with a common run-count scale. Show the total through bar height, with pass/failure segments proportional to their counts. Keep every day's zero baseline aligned, including mixed outcomes and empty days; filter changes must not shift individual bars below zero. Tooltip wrappers around bars use block layout so outcome segments cannot affect inline baseline alignment. Put exact counts and totals in the tooltip, keeping the chart free of per-day numeric labels. Opening or closing a day tooltip must not add in-flow content or shift the chart. Never hide an individual failure inside a pass-rate band. Open day evidence in the shared interactive anchored popup with job logs, commits and pull-request links; preserve selected UTC dates in the URL and browser history. Keep missing evidence distinct from zero failures and label requested-but-unrun, skipped and undecided outcomes. Place daily activity before project-wide coverage and findings, with job-log, commit and pull-request evidence available in run drill-downs. Keep wide tables independently keyboard-scrollable.

Performance keeps the compact model dropdown with the shared filter buttons. Show measurements in one table, including filtered and legacy model URLs, with TPS, VRAM, TTFT, OS, and context. Keep quantization in the evidence popup. Model names open anchored evidence popups instead of a second detail page. Label UTC once and keep timestamps in the evidence. Display context workload labels as 4K/32K without changing source identities or inferring missing numeric context. Keep workload details in the evidence popup.

Performance mirrors Package Qualification’s structure: stable uses a single table with Build as its first grouped column; nightly uses the shared release headings above separate tables with GPU family as the first grouped column. Reuse the qualification matrix, release, scrolling, header, and cell styles directly.

Keep all measured builds visible by default; select the latest observation within each build/configuration rather than across builds. Keep test details in evidence popups instead of a separate Recorded tests section.

Build Qualification and Performance place Find build and Reset filters together in the filter toolbar. Reset clears view filters and package/day/model drill-down constraints while preserving the ROCm version and current qualification window. Model-name links from performance previews open all workloads for the selected model/package; exact metric links must show the observation’s workload selected.

Use compact metric headers with shared definition tooltips for units. Show Context for the configured window capacity. Sort measurement rows within their build and GPU-family groups. Rank builds by a selected metric only within one fully known compatible cohort; do not rank different models or workloads as faster builds. Use removable scope controls consistently in qualification and performance filters.

In CI Activity, place repositories beside visible project-wide findings below the CI jobs table, stacking on narrow screens. List individual CI jobs in an always visible table with Job, an icon-only Platform column and 14 dated outcome columns matching Daily activity. Put Stack, Target and the pass rate for those 14 days in the job-name tooltip; retain accessible platform names. Job-name tooltips identify the job and its evidence; pass rate covers decided results in the displayed 14-day window. Keep job metadata and per-day run evidence available in accessible anchored tooltips. Use a matching calendar surface for a ranked failing-step table with exact failure counts and affected-job counts. Give each step equal row height so rare failures remain readable beside dominant failures; do not encode these counts as relative bar lengths. Label its collection window separately from the 14-day activity chart. Truncate long job and step labels using ellipses and expose full labels on hover, keyboard and touch.

Keep the CI jobs heading directly above its table without helper paragraphs. Use compact, evenly aligned rows with full-cell day triggers and no inline baseline gaps. Keep quiet-job details in the job-name tooltip. Show incomplete collection as a subdued status with a warning symbol beneath the project title, without a bullet or a support disclaimer there.

Project findings use neutral, low-contrast notices with a restrained alert symbol; reserve failure colors for specific outcome evidence. Run tooltips use the shared structured header and body, with a date, outcome counts and grouped run/evidence links. Keep the selected day fixed: no previous/next controls inside tooltips. Follow Decision 0001: Shared evidence tooltip format. Qualification check tooltips establish the default: compact headers, a consistent label column, left-aligned values and fine dividers beneath each detail row. Use shared TooltipContent formatting instead of page-specific alignment rules. Each CI run has a small job/time heading and explicit labeled result, log, failed step, commit and pull-request fields. Model performance explicitly uses the custom body format for comparison graphs while retaining the same tooltip shell. Use the shared popup close control, with a close button on standalone saved-day fallbacks. Match qualification evidence links with inherited text color, dotted underlines and a small external navigation mark.

Shared outcome icons and legends

Use the shared outcome registry, BuildOutcomeIcon, and OutcomeLegend for new evidence tables and charts. Do not introduce local glyph maps or handwritten legend markup. This applies to every project, including empty and partially collected projects. Render status symbols as the shared inline SVGs: a centered 24 × 24 viewBox, 2.3-unit rounded strokes and a consistent 1rem display box. Circular outcomes share an 8-unit radius. Inherit the status color; keep each icon decorative with its accessible name supplied by the owning cell, control or legend.

Outcome Glyph Meaning
Pass ✓ Recorded results passed
Warn ! Recorded warning; explain its meaning in the evidence
Fail × At least one recorded failure
Skipped ↷ The job/check was explicitly skipped; collected evidence, not absence
Not applicable ∕ The check does not apply
No observation — No observation or qualification verdict was collected for this scope
Requested, no runs □ A requested job never ran; not a failed execution
Nonbinary result ○ Collected activity has no binary verdict; consult evidence, not necessarily pending
Running ◷ An observed run has not completed
Cancelled ⊘ The source explicitly recorded cancellation

Use the same compact wrapping legend layout beneath each related table or chart. Cell legends show the same tinted block and glyph as their cells; stacked-bar legends use matching outcome swatches. Keep skipped, unreported, pending and requested-but-unrun evidence neutral. The neutral em dash replaces both the former readiness “No data” text and upstream question mark. Skips use a distinct turn arrow; never reuse the not-applicable slash for running jobs. Give icons accessible labels through the cell, control or tooltip, and keep the legend labels visible.

Domain labels may clarify what an outcome means, without changing its glyph or color role. Package qualification uses the established check labels. Coverage uses “Observed” and “Not observed / unreported,” with unknown evidence remaining neutral. Upstream Overview retains explicitly labeled pass-rate bands; these are aggregates, not individual job verdicts. In the CI jobs table, any recorded failure renders Fail; mixed running/other aggregates use the nonbinary symbol and explicit “Running / other recorded result” wording. Individual evidence distinguishes running, cancellation and other source conclusions. No project gets its own icon set, legend layout or tooltip format.

Place a domain-specific legend directly below every overview calendar. Readiness labels refer to qualification; upstream bands refer to decided-run pass rates. No decided tests is unavailable, never 0% passing. Do not add impossible graph states. No-observation controls support hover, focus and touch through shared evidence surfaces. Include UTC date, track/job/platform and snapshot freshness; keep the basic observation concise. Observed state remains separate from assessment classification (expected/normal, explained abnormal, unexplained). Absence does not prove a daily schedule, retirement, hardware outage or hidden workflow condition.

Only exact-scope persisted assessments supply causes and investigation URLs. Collection incompleteness must not be relabeled producer inactivity. Team-notified wording requires persisted GitHub sent; acceptance does not imply reading or acknowledgment. Show unavailable notification data honestly.

Use the declarative heading Failing steps on main for the failure ranking.

Latest CI job evidence belongs in the existing job tooltip, using the shared detail rows: latest recorded outcome, UTC completion, exact job-log link and observed runner. Keep the historical pass rate separately labeled with its 14-day window. Missing completion evidence is unavailable, not failed. Label runner usage “Observed on devlab” with the collection window; never infer current availability from a historical runner name. Show completion and observed runner in individual run evidence when recorded. These fields do not change upstream-support verdicts or daily history. Overview adds a quiet, accessible DevLab observed indicator beside the project badge; its tooltip identifies jobs and the collection window. Preserve the flat project layout.

Coverage labels distinguish no configured jobs from incomplete collection. Do not show absence counts for incomplete collection. Name activity metrics by their recorded condition and collection window, rather than using the present-tense “tests not running”.

Overview pages keep the calendars prominent: omit publication-summary lines beneath readiness calendars and per-project aggregate metric lines on Upstream Overview. Put incident notification status and its investigation link inside the matching status element's hover/focus/touch evidence; do not add a page-level incident panel. Keep healthy tooltips and navigation unchanged. Findings show direct cause text without a confidence prefix. Use one Finding sentence and an Investigation row in the shared grid only for relevant abnormal results, without generic fallback copy or repeated footnotes. Missing titles use YYYY-MM-DD · Platform. Calendar markers have no link-like underline; preserve accessible names and focus. CI Activity retains detailed project metrics. Upstream status labels use plain text and a small square status marker, without pill backgrounds. DevLab detail triggers have no dotted underline; retain hover details and visible keyboard focus.

Upstream project headings group compact lowercase status labels directly after the title: upstream, partial, unobserved, incomplete, or unconfigured, plus an independent devlab label when runner evidence exists. Use subtle tinted fills and borders with 3px corners, without status squares or dotted underlines. Keep full explanations in shared hover/focus/touch tooltips. This supersedes the plain-text status treatment.

Overview and CI Activity project headers align the compact status-label group to the right, allowing it to wrap at narrow widths. Include the same project-wide DevLab indicator in both headers, independent of selected job filters. Shared Tooltip defaults to styled evidence. Select tooltipMode="system" with an explicit plain-text systemText only for a required native tooltip or fallback. Choose exactly one renderer; do not add native title attributes inside styled triggers. Keep keyboard focus, hover and touch details in styled mode.

Project status labels and explanations are pipeline-owned presentation data. Render status_labels text, tooltip and semantic tone as supplied; do not derive labels or wording from evidence in React. Activity classification and metric copy also come from activity_assessment; the UI only applies visible-job filters. Older snapshots without these optional fields display no invented labels.

In CI Activity, place the project status-label group beneath the project title, left-aligned. Overview retains its right-aligned labels.

CI Activity summary metrics use a prominent percentage with a smaller inline sample size and the caption “Collected results passed”. Show the failed-check merge count with an inline PR unit and “Merged with failed checks”; avoid color jargon. The project collection notice communicates incomplete coverage. The desktop shell page heading aligns with the main content inset.

Section descriptions in CI Activity sit below their section titles, left-aligned. Align the desktop app identity with the workspace rail text inset; align the page heading with content.

CI Activity displays both inactivity metrics as “No runs observed” and “Requests not started”. Use zero when the published assessment omits a count, including incomplete collection. This is a display fallback, not verified absence; retain the project-level incomplete-collection indicator. Tooltip descriptions come from the collector’s optional metric_descriptions fields, including window scope and the zero-fallback limitation.

The Build Qualification Performance filter uses the compact label “Measured” for builds with performance data.

ROCm version filter buttons explain track scope using the dataset's stable anchor and numeric version ordering, matching the readiness overview. Do not hard-code version classifications; older datasets without an anchor report unavailable classification. The Measured filter explains matching measurement availability. Upstream Project, Stack, and Target legends open shared styled explanations with repository and configured filter values from the dataset. Platform filters omit help tooltips and consistently pair OS names with shared platform icons.

Filter help stays in anchored popups on phones so it does not replace filter actions or become page evidence. Other evidence keeps the mobile disclosure pattern.

When workspace navigation is collapsed, place its Navigation button at the leading edge of the header, before the ROCm Readiness identity. Keep the theme control on the right and the current page heading below on phones.

The top bar places a plain-language dashboard description beneath the current page title, replacing the numbered context label. Keep it visible on phones and allow natural wrapping beside the filter control. Collection notices keep the warning symbol in its own column and wrap their text without a separator bullet.