What is included​

• One fixed-timestep, seeded Runner kernel shared by browser and Node adapters.
• A versioned JSONL worker protocol and Gymnasium package.
• A reproducible Stable-Baselines3 PPO example with local run artifacts.
• A lazy-loaded Runs workspace and loopback-only job service.
• Independent application, environment, protocol, manifest, and trajectory versions exposed through an accessible version panel.
• MIT licensing, DCO contribution sign-offs, and explicit open-core extension seams that do not implement billing or hosted infrastructure.

Scope and privacy​

This is a prerelease candidate, not a public release or deployment. Existing Dashboard, Arcade, AI Rescue, Chat, Prompt, Observability, and Journal behavior remains available. Training data, trajectories, and model checkpoints remain in ignored local storage and are not uploaded without a separate explicit action.

The conformance target is deterministic Runner transitions for the same environment version, seed, and actions. Learned PPO weights can vary across hardware and numerical stacks.

The implementation inventory maps the delivered source areas, contracts, security boundaries, validation, and deliberately excluded hosted or commercial capabilities.

Root cause​

The Game Boy shell is scaled independently from Phaser. Phaser's FIT mode read the already transformed parent bounds and wrote centering margins back onto the canvas, producing a visible gray band above the game and cropping the rendered width at some Size settings.

What changed​

• The Runner remains a 320x180 scene, but its rendered canvas is now top-centered, full-width, and uncropped inside the LCD. The unused lower remainder shares the game's dark background.
• Supporting secure browsers show POWER {percentage}%. Battery events update the percentage, level arcs, LED state, charging glyph, and accessible status without polling.
• Charging uses the EXEPERT lime accent. Discharging uses green above 50%, amber from 21-50%, and red at 20% or below; the percentage ensures status is never communicated by color alone.
• Unsupported, blocked, rejected, or invalid Battery API results keep the classic static POWER label and red LED. No percentage is guessed.
• Battery information is never persisted, logged, transmitted, or added to Arcade research telemetry, and listeners are removed when Arcade closes.

The Battery Status API is a progressive enhancement: browser support is limited, secure contexts are required, and Permissions Policy or privacy controls can prevent access. The Game Boy remains fully usable in every fallback state.

Verification​

Automated coverage validates percentage normalization, visual thresholds, live levelchange and chargingchange updates, listener cleanup, missing and rejected API fallbacks, intrinsic 320x180 canvas geometry, LCD containment, Runner and AI Rescue alignment, console resize behavior, and document overflow. The combined telemetry release passed app and docs typechecks, 257 unit tests with six expected skips, all 24 Arcade and responsive Playwright scenarios, and both production builds. Visual checks at desktop, 970x820, and 390x844 confirmed the top-aligned LCD, live charging readout, contained phone Fit behavior, and the absence of document-level horizontal overflow or browser-console errors.

Root cause​

Dashboard and Arcade previously shared the same implicit two-row compact grid. When Chat restored a longer conversation, its intrinsic content height expanded the lower row and compressed the brain and Game Boy into a shallow top strip. The brain canvas then appeared stretched, its status bar consumed most of the tile, and the Arcade viewport could not use the available application height.

What changed​

• At 961-1240px, Arcade uses explicit responsive rows: minmax(260px, 0.85fr) for Brain and minmax(0, 1.15fr) for Study/Chat. Brain and Study/Chat share the left column while Game Boy spans both rows in the right column.
• The tracks remain stable in Study and Chat modes, so restoring a long transcript cannot resize the brain, Chat, or console panels.
• Tablet Arcade Chat tabs use a compact 30px height. Existing phone touch targets and the Dashboard tablet layout are unchanged.
• Compact Affective Field telemetry is a single bar-free horizontal strip. It retains the dominant state, phase, intensity, valence, arousal, and toxicity; narrow tiles scroll locally while Firefox and WebKit scrollbars stay hidden.
• #affectField is keyboard-focusable and shows a focus-visible outline. Arcade menu navigation yields arrow keys while this strip is focused, letting keyboard users scroll telemetry without changing normal Game Boy controls.

No API, event, storage key, chat-state, DOM-order, or DOM-ID contract changed. Desktop widths above 1240px, the Dashboard tablet reflow, and the single-column Brain -> Study/Chat -> Tool order at 960px and below remain intact.

Verification​

Responsive Playwright coverage now restores a deliberately long conversation and verifies fixed Arcade tracks at 1240px and 970px, brain/canvas containment, the full-height console column, the 30px tabs, locally scrollable telemetry, keyboard scrolling, phone ordering, and the absence of document-level horizontal overflow. Existing Dashboard, Prompt, Observability, Journal, Arcade menu, Fit, resize, and narrow-panel behavior remain covered.

The release passed app and docs TypeScript checks, 252 unit tests with six expected skips, all 18 focused responsive and Arcade Playwright scenarios, the production app build, and the Docusaurus production build. Visual checks at 1240x820, 970x820, 390x844, and 390x667 confirmed the intended proportions, a 46px affect strip, readable transcript, and pinned composer.

Root cause​

The tablet Chat tile had only about 269px of internal height, while its model header, context pills, 149px Affective Field, quick prompts, and 73px composer required more than the tile could hold. Focusing the input then scrolled the clipped #leftChat container, hiding the header and leaving the message list only a few pixels tall.

What changed​

• At 961-1240px, active Chat receives a minimum 390px lower row. At 960px and below, active Chat uses clamp(420px, 72svh, 640px) without changing the Study panel height.
• The compact Affective Field is now a bar-free summary strip. It keeps the dominant state, phase, intensity, valence, arousal, and toxicity visible, wrapping the four values into a 2x2 grid on narrow tiles.
• The redundant context-pill row is hidden below 1241px; the selected model remains visible in the Chat header.
• Fixed Chat chrome no longer shrinks. The transcript owns the remaining space and scrolls independently, while the composer stays pinned to the bottom.
• Quick prompts remain visible for an empty chat, disappear after the first user turn, and return when New chat resets the conversation.
• Programmatic composer focus prevents browser scrolling, so #leftChat stays at scrollTop = 0 and its header remains visible.
• The docs package now declares its Docusaurus-compatible React type dependency directly, preventing the docs typecheck from resolving the app's React 18 types when both projects are installed in the same checkout.

The existing desktop Chat, message history key, model picker, Affective Field calculations, IDs, events, and public interfaces are unchanged.

Verification​

The responsive Playwright coverage exercises 1440px, 1240px, 970px, 960px, 390x844, and the short-height 390x667 case. It checks transcript allocation, telemetry height, prompt reset behavior, composer containment, brain-stage containment, parent scroll position, and horizontal overflow across Dashboard and Arcade while preserving Prompt, Observability, and Journal layouts.

The release passed the app and docs TypeScript checks, 252 unit tests with six expected skips, ten focused Chat/responsive Playwright scenarios, the production app build, and the Docusaurus production build. Visual checks at 970x820, 390x844, and 390x667 confirmed the compact telemetry, readable transcript, pinned composer, zero parent scroll, and absence of horizontal overflow.

Root cause​

The Cortical Response Index, onset, sustained activation, and duration values were already refreshed from neural telemetry every 100 ms. The nearby LOW/HIGH gradient was purely decorative, however, and the score occupied a separate row above a 78px spacer. Narrow right panels then stacked a large empty 16:9 preview above the useful metrics, making the card appear static and pushing live data below the fold.

The first compact treatment also placed the variable-length semantic band in the same flex row as the heading. Medium forced the heading onto a second line, adding 13.75px to the metric and score card. Separately, opening Chat let its intrinsic content expand the lower grid row from roughly 344px to 526px and compress the Brain row from roughly 352px to 170px. The canvas CSS height fell from 271px to 89px while its old 602x337 backing buffer and camera aspect stayed unchanged, visibly stretching the neural scene.

What changed​

• Cortical Response is now the first item in a 2x2 group with Duration, Onset Response, and Sustained Activation.
• A normalized live meter keeps the number, fill, accessible value, and Low, Medium, or High band synchronized. The boundaries are 0-39, 40-69, and 70-100.
• The headline score is 26px, reducing to 24px in compact panels, with a muted 12px /100 unit.
• The stimulus preview remains beside the metrics down to a 361px panel width. On narrower panels, metrics come first and an empty preview collapses to 64px; loaded media remains 16:9.
• The old potential row, static gradient, LOW/HIGH labels, and 78px content gap have been removed.
• The semantic band now occupies a fixed-width column beside the meter, the heading remains one line, and the score reserves three numeric characters. Low, Medium, and High therefore share identical geometry.
• Compact Dashboard now uses the same stable approximately 300px/406px rows as Arcade at 970x820. Short tablets scroll inside the workspace instead of squeezing the brain below 260px or clipping Chat.
• A ResizeObserver synchronizes the Three.js camera and DPR-aware backing buffer whenever the canvas container changes size. The window listener remains as a fallback.

No telemetry schema, public API, event, storage key, chat behavior, or existing Dashboard metric ID changed.

Verification​

Validation completed on August 2, 2026:

• pnpm typecheck and pnpm build passed.
• pnpm test passed 257 tests; the six skipped tests remain the existing opt-in live-integration cases.
• The responsive Playwright suite passed all nine cases, including score and meter synchronization, compact empty and loaded previews, metric containment, stable score geometry, synchronized canvas buffers, short-tablet scrolling, Dashboard ordering, and document-level overflow checks.
• The Arcade Playwright suite passed all 15 cases, confirming that the shared panel styling did not regress the Game Boy workbench.
• The documentation typecheck and production build passed.
• Visual checks at 970x820, 1186x810, 390x844, and 390x667 confirmed the intended hierarchy, stable band treatment, undistorted brain canvas, and absence of horizontal overflow or console errors.

What changed​

• css/app.css reorders the initial, Dashboard, and Arcade workspace at the existing 1240px and 960px breakpoints without changing the desktop grid.
• At 961–1240px, the brain and Tool/Metrics panels occupy the first row while the shared Study/Chat panel spans the full second row.
• At 960px and below, the primary flow is Brain, Study/Chat, then Tool/Metrics, keeping Chat reachable before the longer metrics stack.
• The compact tablet brain card can shrink below its desktop 560px minimum, preventing the WebGL canvas and status bar from being clipped by the tile. Existing phone height clamps still control the taller single-column stage.
• The responsive selectors include the unset first-paint state, preventing a layout jump before the workbench controller marks the Dashboard active.
• tests/e2e/responsive-workspace.spec.ts covers desktop preservation, the 1240px and 970px tablet layouts, the 960px and 390px phone layouts, Chat containment, brain-stage clipping, horizontal overflow, and the unchanged Prompt, Observability, and Journal workspaces.

User impact​

The brain visualizer now appears in the former top-left position at compact widths instead of being pushed into a clipped full-width row. Chat gains the full lower row on tablets and remains immediately below the brain on phones, which improves line length, input visibility, and access to the conversation.

Prompt Playground, Observability, Journals, Settings, the wide-screen layout, chat state, DOM identifiers, and runtime APIs are unchanged.

Verification​

Checked locally with:

pnpm typecheck
pnpm test
pnpm exec playwright test tests/e2e/responsive-workspace.spec.ts
pnpm build
pnpm --dir docs-site typecheck
pnpm --dir docs-site exec docusaurus build

The app suite completed with 252 passing unit tests and six expected skips. All four responsive Playwright scenarios passed, and visual smoke checks at 970px and 390px confirmed contained brain telemetry, the intended panel order, a full-width tablet Chat surface, and no document-level horizontal overflow.

What changed​

• docs-site/scripts/generate-activity-calendar.mjs generates static activity data from changelog markdown and committed Git stats.
• docs-site/src/components/ActivityCalendar renders the calendar, activity markers, day details, month navigation, and click-to-filter behavior.
• docs-site/src/theme/BlogListPage adds the tracker above /blog entries and filters visible changelog posts when a day is selected.
• docs-site/package.json runs activity generation before docs start and build commands, while docs-site/docusaurus.config.ts shows all posts on the blog index so filtering works across the full changelog.
• The calendar detail panel now uses container-aware layout rules to prevent hover flicker and to stack cleanly on narrower blog columns.
• css/app.css gives the Prompt list a thin neutral scrollbar, stable gutter, and removes the older duplicate prompt-list style block.

User impact​

• Developers can scan changelog and code activity by date from /blog.
• Hovering a day previews the related changelogs, commits, file count, and line activity.
• Clicking a day filters the changelog list to that date, and clearing the filter restores the full list.
• The Prompt list scrollbar feels quieter and no longer shows chunky browser scrollbar buttons.

Verification​

Checked locally with:

cd docs-site
pnpm activity:generate
pnpm typecheck
pnpm build

Browser checks confirmed the calendar renders on /blog, June 17 and June 18 filter to the expected changelog posts, hover previews no longer flicker, and desktop, tablet, and mobile widths have no horizontal overflow.

What changed​

• index.html removes the .brain-badge overlay from the brain stage and updates the initial status text with N 0, A 0, and SIG 0.
• css/app.css removes the desktop and mobile badge styles plus the live-dot styling that only existed for that overlay.
• src/brain-ui.ts no longer caches or updates #brainTelemetry, and the removed buildTelemetryLine() helper is gone.
• src/scene.ts adds neuron and axon totals to the status bar. The full status uses NEURONS and AXONS; the compact status uses N and A to keep the line readable on narrow canvases.

User impact​

• The brain canvas is visually cleaner and focuses on the 3D brain plus the existing affect label.
• Duplicate telemetry is gone. Neuron and axon counts now live with the rest of the stage readout in the bottom status bar.
• The right-panel telemetry tiles are unchanged, so dashboard metrics remain in the same place.

Verification​

Checked locally with:

pnpm typecheck
pnpm test
pnpm exec vite build

The built app was served locally and checked at desktop and mobile widths. The checks confirmed the badge DOM is gone, #brainTelemetry no longer exists, the affect overlay remains, the status bar includes neuron, axon, and signal counts, the right-panel metrics remain populated, and the WebGL canvas renders with visible pixel variation.

What changed​

• docs-site/src/css/custom.css removes hero gradients, normalizes hero title spacing, and simplifies the eyebrow badge.
• docs-site/src/pages/index.module.css tightens homepage feature cards with an 8px radius, quieter hover states, and more compact spacing.
• docs-site/docusaurus.config.ts keeps local Docusaurus search, adds page indexing, expands result limits and context, and excludes navbar, footer, breadcrumbs, pagination, table of contents, and doc footer chrome from the index.
• docs-site/package.json adds preview:search for the production-search workflow: build the docs site, then serve the generated static output.
• Current docs, versioned docs, and blog posts replace long dash separators with cleaner punctuation or explicit wording.

User impact​

• The docs homepage has a flat visual style with no gradient coloring.
• Search validation is less confusing: pnpm start remains for live editing, while pnpm serve or pnpm preview:search validates the generated search index.
• Searching for terms such as tele in the built site returns real local results instead of the build-index warning.
• The visible docs copy no longer uses em dash or en dash separators.

Verification​

Checked locally with:

cd docs-site
pnpm typecheck
pnpm build
rg -n '\x{2014}|\x{2013}' docs-site/docs docs-site/versioned_docs docs-site/blog docs-site/src/pages docs-site/docusaurus.config.ts
rg -n 'linear-gradient|radial-gradient' docs-site/src

The built site was served at http://127.0.0.1:3001/. Browser checks confirmed the homepage hero has no background image, the mobile layout fits at 390px, and searching for tele returns Telemetry results without the build-index warning.

What changed​

• src/affect/* adds the affect engine: shared types, the emotion palette, the deterministic local classifier, text hashing, affect-to-region mapping, and the browser client for saving affect annotations.
• src/brain-ui.ts classifies user input immediately, classifies assistant output after streaming, renders message affect chips, updates the new Affective Field, and triggers emotion-colored neural pulses. Chat error states can also become assistant_error affect events when the failed span is available.
• index.html adds the Affective Field readout inside the Chat tile and the canvas-stage affect overlay.
• css/app.css styles the Affective Field, message chips, stage overlay, and responsive narrow-panel behavior.
• rust/brain_sim/src/lib.rs extends the WASM signal pool with per-signal RGB color buffers. Stimuli can now pass a hex color, and signals propagate that color through neuron firing.
• src/brain.ts binds the WASM particle color buffer to a Three.js color geometry attribute so simultaneous signals can render different emotion colors in the same canvas.
• supabase/functions/chat-affect/index.ts adds an authenticated Edge Function that writes affective_state span annotations. It always stores the local CODE classifier output and can also store an LLM-refined annotation through 9Router.
• supabase/functions/run-chat/index.ts now includes trace_id, span_id, otel_trace_id, and otel_span_id on streaming error payloads when the error span was persisted successfully.

Affect model​

The local classifier scores:

• anger
• fear
• sadness
• joy
• disgust
• surprise
• curiosity
• calm
• distress
• neutral

Profanity is context-aware. For example, "fuck yeah" routes toward high-arousal joy, while directed hostile profanity routes toward anger, distress, and higher toxicity. The UI can look dramatic, but docs and metadata call the output synthetic affect telemetry.

Persistence​

chat-affect writes to existing span_annotations rather than adding a new table:

• name = "affective_state"
• annotator_kind = "CODE" for local analysis
• annotator_kind = "LLM" for optional refinement
• identifier = chat-turn:<turn_id>:phase:<phase>:source:<source>

Metadata stores score vectors, region mix, color palette, classifier version, selected chat model, affect model, phase, and text hash. Raw message text is not duplicated in affect metadata.

Verification​

Checked locally with:

pnpm run typecheck
pnpm test
pnpm build
pnpm --dir docs-site build

A browser smoke check opened Chat at http://127.0.0.1:9001/, verified the Affective Field on desktop and narrow viewports, confirmed no Chat overflow, and confirmed the WebGL brain canvas screenshot was nonblank.

What changed​

• src/ui/keyboard-shortcuts.ts adds a shared registry for visible keyboard commands and a guard that decides whether global scene shortcuts should run.
• src/events.ts now scopes the visualizer shortcuts so Space, A, and S only affect the 3D scene when focus is outside inputs, buttons, chat, search, Settings, diagnostics, prompt playground, and lil-gui controls.
• index.html adds a Keyboard Shortcuts & Commands card to Settings and changes the initial clip preview from a fake 0:06 duration to an explicit empty state.
• src/ui/workbench.ts renders the Settings command list from the shortcut registry so the visible docs match the active runtime behavior.
• src/ui/media-duration.ts centralizes duration labels: no stimulus shows :, images show Still, videos show their metadata duration, and unknown loaded media shows Loaded.
• src/brain-ui.ts now resets cleared media to an empty state, waits for video metadata before showing a duration badge, and removes the clip-thumb badge when no stimulus is loaded.
• css/app.css adds the empty clip thumbnail style and responsive Settings shortcut rows with compact kbd chips, scope labels, and wrapped descriptions.
• src/__tests__/keyboard-shortcuts.test.ts and src/__tests__/media-duration.test.ts cover the shortcut guard and duration label behavior.

User impact​

• Pressing A, S, or Space while typing in Chat or interacting with Settings no longer toggles visualizer helpers by accident.
• Settings now lists the app's current global and contextual commands, including Search, Diagnostics, Chat send, model picker navigation, and the Floor Platform visualizer command.
• Fresh app load and cleared clips show Duration : and a muted No clip thumbnail instead of 0:06.
• Uploaded images show Still instead of a fake video duration; uploaded videos show their real metadata duration once available.

Verification​

Checked locally with:

pnpm typecheck
pnpm exec vitest run src/__tests__/keyboard-shortcuts.test.ts src/__tests__/media-duration.test.ts
pnpm build

The browser smoke test opened http://localhost:9001/, confirmed the empty clip state, confirmed the Settings list rendered nine command rows, and verified A, S, and Space type normally inside the Chat input.

What changed​

• index.html replaced the native #brainModelSelect with an app-owned model picker: #brainModelTrigger, #brainModelMenu, #brainModelSearch, and #brainModelList. It also added the data-view="observability" right-panel view and the new #observabilitySlot mount target.
• src/brain-ui.ts now renders the model list as a custom listbox while keeping the existing chatModels, selectedChatModel, and exepert.brainChat.model.v1 persistence flow. The menu supports provider grouping, search, variant badges, Escape, ArrowUp/ArrowDown, Enter, outside-click dismissal, and disabled state while streaming.
• src/ui/observability.ts now prefers #observabilitySlot, falls back to the old #obsSlot only for older shells, and uses exepert.obsCollapsed.v2 so the activity-page panel defaults to expanded.
• css/app.css adds the dark model popover, selected/active row states, responsive badge behavior, Observability workbench view rules, and the wider Trace Ingestion page layout.

User impact​

• Opening the Chat model picker no longer shows a white operating-system dropdown in the dark UI.
• Codex and Claude models from 9Router are searchable and grouped by provider.
• Chat keeps its full vertical space because Trace Ingestion is no longer mounted under the chat tile.
• The Observability activity icon now opens a focused Trace Ingestion page with the existing import, sample, trace viewer, live, project, Supabase status, and summary controls.

Verification​

Checked locally with:

pnpm typecheck
pnpm build

The browser smoke test confirmed that the menu listed 25 models, filtering worked for Claude models, keyboard selection worked, a chat request used the selected cx/gpt-5.4 model, and #obs-panel mounted inside #observabilitySlot rather than the Chat area.

What changed​

• Added scripts/vercel-build.sh as the Vercel build command. It prepares Rust, installs the wasm32-unknown-unknown target and wasm-pack, runs the Vite production build, builds docs-site/, and merges the rendered docs into dist/.
• Added scripts/copy-docs-build.mjs to copy docs-site/build into dist/ while skipping the docs site's root index.html. The app stays mounted at /, and Docusaurus pages are available under routes like /docs/getting-started/introduction/.
• Replaced hardcoded localhost:3000 documentation links with environment-aware links. Local app origins use http://localhost:3000; production uses the app's deployed origin unless VITE_DOCS_URL is explicitly set.
• Added EXEPERT favicon assets for the app and docs site: public/favicon.png, public/favicon-32.png, public/favicon-192.png, public/apple-touch-icon.png, and docs-site/static/img/favicon.png.

Deployment note​

Set Vercel's Build Command to:

bash scripts/vercel-build.sh

Leave the Output Directory as dist. Do not set VITE_DOCS_URL for the normal Vercel deployment unless the docs are intentionally hosted somewhere else.

Decluttered chat header​

The chat tile header (.brain-chat-header) previously showed a "Neural Brain" title plus a live-telemetry status pill (pulse dot + "Live telemetry" / "Waiting for signal"). Both were removed; the header now holds only the close (×) button, right-aligned.

• index.html: dropped .brain-chat-title and .brain-chat-status-inline (which contained #brainChatStatus). The dialog now names itself with aria-label="Neural Brain chat" instead of the dangling aria-labelledby="brainChatTitle" (no such element ever existed).
• src/brain-ui.ts: removed brainChatStatus from the cached-element list and the per-tick els.brainChatStatus.textContent = … write, so nothing references the deleted node.
• css/app.css: .brain-chat-header is now justify-content: flex-end; removed the unused .brain-chat-title, .brain-chat-status-inline, .pulse-dot, and @keyframes pulseDot rules.

Collapsible Trace Ingestion tile​

The observability panel (#obs-panel) gained a chevron toggle in its header (#obsCollapse) that collapses the tile to just its header, hiding .obs-body. Because the slot is flex: 0 0 auto and #leftChat is flex: 1, collapsing the panel hands that vertical space straight to the chat.

• Defaults to collapsed so the chat gets maximum real estate out of the box.
• The choice is persisted across reloads in localStorage under exepert.obsCollapsed, wrapped in try/catch for private-mode / SSR safety.
• aria-expanded + aria-controls="obsBody" keep the toggle accessible; the chevron rotates 180° when expanded.

Help link now opens the rendered guide​

The panel's ? help link pointed at Plans/guide-trace-ingestion.md, which Vite resolved to a raw, unrendered .md file on the app origin (a blank page). It now opens the rendered Docusaurus guide page at /docs/guides/trace-ingestion.

On local app origins, docs links default to http://localhost:3000, which matches the separate Docusaurus dev server. In production, leave VITE_DOCS_URL unset: the Vercel build bundles the rendered docs into the app's dist/ folder, so the link resolves on the deployed app origin under /docs.

See the UI guide for the chat tile contract and the trace-ingestion tile notes.

Symptom​

With the Chat tab active, the DOM collapsed into the chat header:

• div.brain-chat-header > div#brainMessages: message column ~30px wide instead of full width.
• div.brain-chat-header > div.brain-chat-input: input overflowing the panel (333px wide inside a 329px panel).

#leftChat is a flex-direction: column tile and its five regions are meant to be direct siblings. Because they were nested inside the header row instead, they inherited display: flex; align-items: center and collapsed.

Root cause​

index.html had a duplicated opening <div class="brain-chat-header"> tag with only one matching </div>:

<div class="brain-chat-header">
  <div class="brain-chat-header">
  <strong class="brain-chat-title">Neural Brain</strong>
  …status · close…
</div>
<!-- chat-context, brainMessages, quick-prompts, brain-chat-input
     were now all trapped INSIDE the header -->

The unbalanced tag pushed .chat-context, #brainMessages, .quick-prompts, and .brain-chat-input inside the header, so the column layout never applied.

Fix​

Removed the duplicate opening tag so the header self-closes and the five regions are siblings again:

<div class="brain-chat-header">
  <strong class="brain-chat-title">Neural Brain</strong>
  <span class="brain-chat-status-inline">…</span>
  <button class="brain-chat-close" id="closeBrainChat">&times;</button>
</div>
<div class="chat-context">…</div>
<div class="brain-chat-messages" id="brainMessages">…</div>
<div class="quick-prompts">…</div>
<div class="brain-chat-input">…</div>

No CSS changed: the stylesheet (.brain-chat-messages { flex: 1 }, the column container, the input row) was already correct. See the UI guide for the sibling contract that this tile depends on.

What's documented​

This first pass covers the project as it actually runs today:

• Getting started: prerequisites (Node, pnpm, Rust + wasm-pack), installation, the dev loop on port 9001, and production builds.
• Architecture: the two-layer design where a Rust World compiled to WebAssembly owns the state and the Three.js front end reads it through zero-copy buffers.
• Simulation core: neuron/axon generation, the struct-of-arrays signal pool, the four cortical regions, stimulus injection, and the telemetry contract.
• Frontend modules, the diagnostics logger, a UI guide, and development notes (testing, CI, the legacy reference pages).

A note on the rendering path​

EXEPERT currently renders with a classic WebGL2 WebGLRenderer and GLSL ShaderMaterial shaders, driven by a CPU simulation step (step_cpu). Earlier design docs in the Plans/ folder describe a WebGPU + TSL compute path that was built and then reverted. The documentation describes the shipped code; treat those plans as history.

Browse the Introduction to get started.