Evaluating semantic zoom, linked displays, and brushing

Use this battery when changing Muriel's multiscale guidance or building an interface whose task depends on coordinated views. It tests whether the skill produces useful working behavior across domains. A passing reference demo is evidence for that demo; it does not establish that a future build, a different dataset, or Muriel's general capability passes.

Read the coordination contract for the design mechanisms, zooming interfaces for navigable spaces and return paths, and the adapters for host integration. The linked-atlas lab supplies one inspectable implementation. The independent fixtures below deliberately cover additional data shapes that the lab does not represent.

Four kinds of evidence

Keep these separate in the handoff. A failure in one cannot be repaired by accumulating passes in another.

Evidence	What to exercise	What a pass establishes
Model	Small independent data oracles, predicates, ID joins, transforms, event ordering	The declared state and aggregation semantics hold for the tested cases
Browser interaction	Actual pointer, keyboard, focus, wheel, and touch routes; rendered state read back from the page	The UI invokes those semantics through the tested input paths
Rendered inspection	Relevant sizes and tiers, selection/filter combinations, empty/loading/error states, text/mark contrast	The inspected outputs are legible and communicate the intended state
Human task observation	Representative tasks with unfamiliar users; a comparison when claiming improvement	What those participants understood or accomplished under the observed conditions

Calling a reducer from a browser console is model evidence. Synthetic pointer events can exercise handlers but do not establish hit testing, pointer capture, native scrolling, or physical touch acceptance. A screenshot does not establish keyboard access. Automated accessibility checks do not establish screen-reader comprehension. Record skipped and unavailable checks explicitly.

The repository includes focused checks for the linked-atlas implementation:

node --test tests/multiscale-coordination.test.mjs
node tests/multiscale-browser.mjs

The browser suite needs Playwright and its Chromium runtime in the test environment. PLAYWRIGHT_MODULE can point to an existing installation, while BASE_URL can supply an already-served lab. Otherwise the script starts and stops its own loopback server. Set SCREENSHOT_DIR to retain captures for a separate rendered inspection. Browser touch checks use Chromium input emulation; physical-device and assistive-technology acceptance remain separate. These checks exercise the reference lab, not all the independent briefs below. The browser suite also checks its rendered HTML and SVG text against the painted background at glyph locations, including SVG label backplates and halos, with an 8:1 floor. Planted low-contrast fills and changed backgrounds verify that the audit can fail. This audit is scoped to the lab's solid-color text; it does not establish chart-mark contrast, native popup-menu contrast, or general image-text accessibility.

Run the capability battery

Run the build briefs below in fresh sessions against a recorded skill revision. Give the agent the brief and fixture, with the required task and existing host constraints. Keep the expected answers available to the evaluator. Require a runnable artifact, its canonical-state description, and bounded evidence for what was verified. Preserve the resulting artifacts and event traces, including failed builds.

Evaluate the output, not the presence of vocabulary in its rationale. A response that says "semantic zoom" while enlarging the same marks fails the relevant task. A modest table with the required correspondence and recovery can pass a linking task without using a prescribed library or visual form.

For each brief report pass, fail, or not observed against its listed behaviors. Record the first reproducible counterexample and user consequence. Do not turn mixed evidence into an invented quality score. Repeat fresh runs when estimating reliability; report runs completed / runs attempted and the number passing each behavior. Fix the fixture before comparing skill revisions when a task or oracle proves ambiguous.

A. Time-series cohort investigation

Brief: Build a small investigation surface for field observations. Let a reader select a time window, compare its habitat composition, inspect individual records, and retain a chosen record while changing the viewport and habitat filter. An overview must retain temporal context. Use calendar day as the time unit; brushing selects a half-open interval [start, end).

ID	Day	Habitat	Value
r1	1	Coast	10
r2	2	Ridge	20
r3	3	Coast	30
r4	4	Coast	40
r5	5	Ridge	50
r6	6	Coast	missing
r7	7	Ridge	70
r8	8	Coast	80

Required observations:

• Brushing [2, 5) selects exactly r2, r3, r4. Day 5 is excluded. Reverse dragging produces the same set. Endpoint inclusion is visible in the range control or explanation.
• Filtering to Coast retains three selected IDs, reports two selected records in the current filter, and exposes five eligible records. The selected visible value mean is 35, computed from r3, r4; the selection denominator is not silently replaced by the total corpus size.
• The full Coast mean is 40 over four observed values, with one missing value separately accounted for. Missing does not become zero or disappear from the record count.
• Pin r2, then filter to Coast. The pin remains identifiable with its hidden status, or the interface explicitly asks to replace it. Hovering r4 and leaving restores the pin; it does not commit a new selection.
• Zooming the main chart changes the viewport and representation as needed, while the overview retains the full temporal domain and the three selected identities survive. Resetting the viewport does not clear the selection.
• An empty brush has a visible zero-result state distinct from "no selection." Clearing the selection restores the declared unselected behavior.

B. Document or image collection across scale

Brief: Build a collection browser for eight synthetic documents. At a broad scale show two collections and their composition; at a middle scale expose documents; close up, expose readable excerpts and metadata. Let a reader select documents, cross a tier boundary, inspect one, and return to the original place. Supply an equally capable linear route for keyboard use and small screens.

Use IDs d1–d4 in collection A and d5–d8 in collection B. Give d2 and d6 the same title, "Survey notes," but different excerpts. Give d4 a missing thumbnail. Initially select d2, d5, d8; pin d6. Filter out d8 in a later step. The contents are fixtures, not an asserted semantic clustering result.

Required observations:

• Overview groups expose selection membership: A has 1/4 selected, B has 2/4. They cannot appear wholly selected. After filtering d8, show B's visible 1/3 selected alongside the fact that one selected document is hidden.
• A tier change adds or replaces task-relevant representation: counts and summaries become recognizable document objects, then readable content. Enlarging thumbnails without adding meaningful content does not pass.
• Selecting or pinning either duplicate title identifies the correct ID and excerpt in every view. Array position and display title are not identities.
• Crossing the representation boundary back and forth preserves selected IDs, the pin, and the spatial anchor. If the focused DOM element is replaced, focus transfers to its semantic successor; it does not drop to the page body.
• Slow movement around a threshold does not flicker between tiers. Returning from detail restores enough location context to find the same neighborhood.
• The missing thumbnail has a usable identity and inspection route. Zooming does not trigger repeated unbounded image loads. Motion reduction preserves the state transition without requiring a flying camera.

Bounded ZUI navigation and return extension: Use the same collection as a two-dimensional space, with world bounds [0, 1600] × [0, 1000]. Place collection A near (300, 300) and B near (1200, 700), with the document positions stored in world coordinates. Start at a fitted overview; select d2, d5, d8 and pin d6. Record camera position, scale, focused entity, and navigation-history length at each action. Declare the allowed scale interval and camera-boundary policy before testing.

• Anchored zoom: Choose an off-center point that is away from world bounds. Zoom by a factor of two and back using pointer, keyboard, and any offered pinch route. The same world point remains at the same screen location within a stated rendering tolerance, such as one CSS pixel. Check both axes; a horizontal chart-domain change cannot satisfy this collection task. Boundary clamping is separately visible and tested, rather than treated as anchor preservation.
• Fit and Back: Pan to an asymmetric view, then inspect d2. Fit computes a camera that contains its declared target (document, collection, selection, or whole space). Back restores the recorded previous camera, including its offset and scale. They must produce distinct results in this fixture. Returning to the preceding location must not clear selection, change the pin, or replace a user's manually chosen camera with a newly fitted one.
• Scale limits and history: At minimum and maximum scale, repeat the outward zoom action. It is a true no-op: the camera, selection, preview, pin, focus, and history length remain unchanged. One continuous gesture creates the declared single navigation checkpoint, not one entry per input event. Back must not require traversing invisible no-op entries.
• Reflow and intentional position: Resize the host from a wide screen to a narrow one after an intentional pan. Preserve the declared world anchor or give a visible recovery action. Browser resize or delayed thumbnail arrival cannot silently refit the entire space. If the design offers a reflowing collection layout, record the world-to-entity correspondence before and after reflow and preserve the selected/pinned IDs; test reflow separately from camera navigation.
• Preview and pin: Crossing the close-up threshold may reveal d2's content or initiate a preview under an explicit policy. It cannot pin d2 merely because the camera approached it. d6 remains pinned; zooming away removes a transient preview and restores the pinned detail. An explicit inspect/pin action remains available through keyboard controls.
• Bounds and reachability: Pan toward all four corners and apply repeated large wheel or key inputs. The camera remains finite and obeys the declared bounded or elastic policy. Every document, including those near the edges, can still be reached and inspected at a readable scale. The overview and a keyboard recovery control can return from an empty region without requiring knowledge of the offscreen coordinates.
• Focus continuity: Focus d2, cross every representation boundary, then return. Its semantic successor keeps focus and a visible indicator even if DOM nodes, canvas layers, or thumbnail representations change. If d2 becomes unavailable, focus moves to an explained collection or navigation target. Virtualization cannot send focus to the page body or leave it on an invisible recycled object.

This extension tests navigation of a collection. It does not require a new faceting runtime. When applying it to Wunderkammer, evaluate any existing filters or collection membership behavior against their own declared task and data contract; do not equate that behavior with camera movement.

C. Heterogeneous linked data and non-bijective joins

Brief: Build a customer, order, and event investigation surface. Let a reader select a customer, inspect the related orders and events, and compare tags. The views have different units of analysis. Make unmatched data and count units visible. A shared x-axis is neither required nor meaningful.

Customers are c1, c2, c3. Orders are o1(c1, 10), o2(c1, 30), o3(c2, 20), and o4(no customer, 5). Events are e1(o1), e2(o1), e3(o2), e4(o3), and e5(unknown order). Tag A relates to o1, o2; tag B relates to o2, o3.

Required observations:

• Selecting c1 links to exactly two orders totaling 40 and three events. The second event on o1 does not count its order amount again.
• Selecting both tags yields three distinct orders, not four. Any weighted or multiplicity-based metric is separately named and justified.
• Selecting c3 yields no related orders and visibly empty target views. Clearing that selection restores all records; the two actions are distinct.
• o4 and e5 remain discoverable as unmatched records. A join cannot silently drop them while claiming to show the entire input dataset.
• The source entity and join direction are intelligible. Selecting event e1 may link to o1 and c1, but any expansion to sibling event e2 is explicitly defined and visually distinguishable from the original selected entity.
• Sorting, filtering, and aggregate expansion do not change identity mapping. The rendered count labels name customers, orders, or events as appropriate.

D. Geographic and temporal views with independent navigation

Brief: Build an observation map linked to a timeline and record list. Select observations in either view, inspect their correspondence, and inspect a dense map region without losing the global time context. The map has longitude and latitude; the timeline has dates. Declare how multiple selections combine.

Use five records: p1(-5, 0, day 1), p2(0, 0, day 2), p3(5, 0, day 3), p4(0, 5, day 4), p5(5, 5, day 5). Define the test rectangle as -1 <= longitude <= 1 and -1 <= latitude <= 6, which yields p2, p4. Define the time brush [2, 4), which yields p2, p3.

Required observations:

• Replace produces the most recent set; add produces p2, p3, p4; intersection produces p2; subtracting the time set from the map set produces p4. Offer only operators the interface declares, and test each offered operator.
• Map navigation changes the map viewport without changing timeline scale or turning pan into selection. The selected IDs remain shared.
• Pixel-to-data conversion works when the plot begins away from the document origin, the page is scrolled, and the graphic is CSS-scaled. Nested transforms and SVG viewBox letterboxing must be covered when the host uses them.
• Small-screen interaction has an explicit way to choose navigation versus brushing. Touch scrolling outside the plot still works; multiple pointers do not silently commit the first pointer's unfinished selection.

Adversarial interaction sequences

Run applicable cases against every built artifact, not only a purpose-built test harness. A feature absent from the product is not applicable, with the reason recorded; an offered but untested feature is not observed.

Sequence	Observable invariant
Start a replacement brush, move outside the plot, press Escape	The prior committed set survives; transient geometry disappears; keyboard focus remains usable
Start a brush, then pointercancel / loss of capture / window blur	No unfinished selection commits; capture and temporary listeners are released
Draw a nonzero rectangle that contains no marks, then clear	Explicit empty selection and absence of selection remain distinguishable
Click or tap without crossing the drag threshold	The documented click action occurs; no accidental empty brush replaces the selection
Start from {a,b}; add {b,c}; subtract {b}; undo if offered	Sets become {a,b,c}, then {a,c}; undo restores exactly the preceding committed state
Preview A while B is pinned; move to empty space or another view	Preview clears through the shared path and B remains pinned
Change a filter while records are selected offscreen	Selection identity and visibility counts agree; filtered or offscreen records are not confused with deleted records
Change data revision during a pending brush	The gesture is cancelled or recomputed under an explicit policy; IDs from incompatible revisions cannot be committed silently
Issue request A, then B; resolve B before A	Visible results stay on B; a stale A response cannot repaint them or restore its selection
Change viewport while a selection overlay exists	The overlay reprojects from its declared domain or IDs; stale screen coordinates do not change meaning
Resize or cross a semantic tier while a child control has focus	Focus has a visible, relevant destination and remains operable
Remove a selected or pinned record from the data	The missing identity is explained or cleared according to a visible policy; another row cannot inherit it

Async, deleted-data, and multi-pointer cases require host fixtures; a static reference lab cannot establish them. Do not retrofit a fake loading spinner to claim asynchronous verification.

Narrow routing counterexamples

These briefs test restraint. Use fresh sessions and inspect the runnable result or proposal; checking whether the agent mentions "brush" is not a valid test.

Brief	Expected decision and evidence
"Fix the spacing of the already specified zoom buttons; behavior is approved"	Follow the host's settled control model; no new selection or linking system
"Show the latest total and its units on a status page"	A readable direct display; no invented exploratory canvas or selection state
"Add a date filter to this appointment form"	A conventional bounded input is sufficient unless a concrete distribution-comparison task is supplied
"Make this static publication figure readable when printed"	Work through the figure channel and print constraints; no interaction required to interpret the export
"The gallery's image magnifier must show the original pixels"	Geometric magnification preserves the requested content; semantic substitution is not imposed
"Build three independent operational controls with unrelated data"	Preserve independent state; shared appearance alone does not warrant cross-filtering

For a paired positive test, add a genuine relationship task to the last brief: "Choose a deployment and inspect its errors, hosts, and traces." The response should then define correspondence and linked state. This tests whether routing responds to task meaning rather than surface nouns.

Human comparison when an outcome claim matters

The following is a study plan, not evidence that these mechanisms improve performance. Start with observed workflow failures and a relevant existing interface. Pick tasks that both conditions can complete, and preserve data, labels, available information, and answer criteria across conditions.

1. Define the claim and task before recruiting. For example: find which habitat changed within a time window, recover a previously inspected record, or explain why an aggregate includes certain members. Decide which task tests correspondence, navigation, selection, or learning. Do not claim a general cognitive benefit from one successful task.
2. Choose the comparison that addresses the mechanism. Compare the host's current interface with the proposed one. If separating zoom from linking matters, use a targeted variant or staged comparison. Keep the baseline capable and equally polished; otherwise the result cannot isolate the mechanism. A factorial study is useful only when the research question and participant budget support it.
3. Separate learning from practiced performance. Capture unassisted first use, provide the same amount of instruction, then use parallel datasets for practiced trials. Counterbalance condition and task order. Use different values with comparable structure to reduce answer memorization. A delayed retention task is needed for a retention claim.
4. Define denominators and failures. Report correct task answers / assigned tasks, participants completing each condition / recruited participants, recoveries / induced recoverable errors, and condition-specific missingness. Measure elapsed time from an explicit start to an explicit completion. Report incomplete tasks and timeouts alongside time for successful tasks; do not hide them in a faster-successes-only mean.
5. Observe the mechanism. Record chosen IDs, operations, tier crossings, resets, accidental selections, backtracks, and help requests alongside task correctness. Think-aloud can reveal confusion but changes task timing; use a separate session or acknowledge that limit. Logs alone do not reveal why a person acted.
6. Treat insight tasks carefully. Collect an answer and its cited data. Define valid findings and scoring examples beforehand; blind raters to condition where practical. Report agreement, unsupported findings, duplicates, and findings per participant within a fixed observation period. Verbosity or number of interactions is not an insight measure.
7. Recruit for the claimed population and input modes. Include relevant domain experience and access needs. Test keyboard, screen reader, or touch with people who use them before making claims about their usability. Use pilot variance and a smallest meaningful effect to plan a comparative sample; a convenient small pilot supports defect discovery, not a universal ranking.
8. Report uncertainty and scope. Analyze paired observations as paired when appropriate, account for repeated tasks per participant, and report effect estimates with uncertainty. Predeclare primary outcomes and exclusions when testing a performance claim. Preserve contradictions: improved speed with worse accuracy or harder first use is a tradeoff, not an overall win.

Evidence receipt

Attach a compact record to a completed evaluation:

Skill and artifact revisions:
Brief, dataset version, and expected answers:
Host browser, viewport, device/input mode:
Model cases passed / failed / not observed:
Real input cases passed / failed / not observed:
Rendered states inspected and captures:
Human participants, assigned tasks, completed tasks (if any):
First reproducible defect and its consequence:
Claim supported by this evidence:
Unverified boundaries and next check:

Do not mark the battery complete merely because the reference example passes. Keep a new failing build as a regression fixture when it exposes a transferable skill or implementation defect.