Bring the portfolio to a clear, reviewable finish

The job the site has to do

This portfolio should make it easy to understand who I am, what I build, and why the work is worth trusting. It should introduce the BREAD stack through real products, show GRAIN as the design system behind the interface, and use Pantry as the app that brings the layers together. The notes, talks, teaching work, and badges are evidence of how I think, what I can explain, and what I have helped other people learn.

The finish line is a connected story across the pages. A visitor should be able to move from the person, to the work, to the proof without having to know the repository first.

Scope

This plan closes the portfolio experience and its presentation. It includes the introductions to BATCH and GRAIN, but it does not change their internal implementations. That deeper design-system work has its own later review. Existing page-level plans remain the source of implementation details; this plan owns the final path across them.

The finish line now also includes a portfolio-wide audit of browser-local AI. Builder generation is one workload, while DESK handles grounded questions, summaries, fuzzy navigation, and selected site actions. The audit will distinguish tasks that should stay deterministic from those that need the model, measure each model-backed task against visitor outcomes, and tune only where repeatable evidence supports a change. It will also produce a reusable standard in standards/LOCAL-AI-STANDARD.md so the same method can be applied to the owner's other projects. Browser responsiveness at desktop and narrow phone widths is the device criterion; physical handset testing is not required.

Open work: portfolio-wide local AI

• [x] Inventory every visitor-facing browser-local model path, its profile, prompt/context inputs, deterministic guards, model output, review step, offline behavior, and first-use download cost.
• [x] Build a task-specific evaluation matrix for DESK grounded answers, summaries, fuzzy navigation, any model-composed action, and Builder composition and editing. Keep deterministic controls in the matrix so the whole visitor flow is checked, not just raw model text.
• [x] Run repeatable live WebGPU baselines and inspect the captured raw answers and visible outcomes. Record model ID, runtime/profile, prompt or decoding changes, warm/cold state, latency, and failures.
• [x] Tune only the failing layer: retrieval and bounded context for grounding, task-specific prompts and decoding for model work, or deterministic code and review gates where generation is the wrong tool. Re-run the same cases and holdouts before keeping a change.
• [x] Verify first-use and warm-use behavior, WebGPU-unavailable handling, and desktop/390px browser presentation for the DESK and Builder surfaces. A fresh 0.5B profile displayed 66% download progress at 20 seconds and completed the first grounded answer in 29 seconds. The warm matrix then passed 51/51. The existing no-WebGPU E2E suite and 390px Builder/GRAIN browser checks passed. Do not require physical-phone testing.
• [x] Write the reusable method and evidence-backed lessons in standards/LOCAL-AI-STANDARD.md, then link it from the standards index and developer docs. Keep implementation-specific facts in this plan or the owning feature plan.

Initial local-AI evidence, October 14

The live WebGPU baseline exercised eleven DESK tasks on Qwen2.5-0.5B. Seven passed the current checks and four failed: “Who is TJ?” omitted teaching; a natural capability question missed the capabilities response and did not mention page summarization; “Show me something interactive here” fell into a long, irrelevant answer; and “what do you know about me?” failed to repeat the saved visitor fact. The “summarize this page” check passed but accepted generic text that named no subject, so the audit's own grader was too weak to count that result as useful. These are visible in the local baseline report at .cache/desk-audit/report-portfolio-local-ai-baseline-2026-10-14.json; the report is local because it includes visitor-visible content and is not part of the published evidence set.

The first code review found a source-of-truth mismatch: src/ai/facts.md did not carry the current About profile's one year of college teaching or the 15-to-25 person leadership range. The shared desk also had no deterministic path for reading back its own marked memory lines. Those are being fixed at the data and routing boundaries. The natural capability wording now uses the existing live capability catalog, and the interactive request maps to the existing model-driven showcase instead of asking the small chat model to invent a description. Summary grading now requires an actual subject from the feed. The first strict recheck exposed two further failures. The shared 0.5B answered a capability question with a page-specific capability list, which is appropriate, but its old grader demanded the literal word GRAIN. The model also invented industries in its Notes summary. The capability check now grades the actual actions the page offers. The Notes feed summary now reads the featured card's authored title and summary directly, a deterministic operation that avoids paraphrase drift. The exact "Who is TJ?" and "How do you work under the hood?" questions now read the matching section of the facts file, so the portfolio's current leadership and teaching facts and the live model/runtime description do not depend on small-model recall. These actions still use the model for open-ended questions, synthesis that needs it, and the Builder's generation/editing paths.

The complete live WebGPU matrix then passed 51/51 scenarios on October 14. It includes grounded DESK questions, the live capability catalog, the cross-page demonstration, summaries, deterministic routing and page actions, visitor memory, 30 Builder edit/composition cases, negative content constraints, and a guided tour. The report records model profile, raw model output, timings, page outcome, and grading: .cache/desk-audit/report-portfolio-local-ai-final2-2026-10-14.json. It is local-only because it contains visitor-visible text. A prior seven-task A/B using Qwen2.5-1.5B passed five tasks, with generation times from 14 to 27 seconds (18 seconds median); it still missed runtime and summary quality. The 0.5B shared-profile baseline took 5 to 18 seconds across comparable tasks. The larger model improved the bio but did not earn its additional download and latency for the whole DESK, so the shared DESK remains on 0.5B. The Builder retains its existing conservative 1.5B device gate.

The standard index and developer docs now link to LOCAL-AI-STANDARD.md. Its rendered page was inspected at desktop and 390px on October 14 with no horizontal overflow. The 390px GRAIN and Builder screens also have no horizontal overflow, and the Builder's live edit and generation model results were inspected at desktop. The affected browser suite passed 87/87, including the offline desk and narrow-phone Builder workbench. The Builder's real-model matrix is desktop Chromium on this Mac, so its phone behavior remains a responsive-browser check rather than a claim about mobile GPU performance. Cold-start evidence is local at .cache/desk-audit/report-portfolio-local-ai-cold-start-2026-10-14.json.

Current progress (2026-10-13)

The project index now gives GRAIN a direct introduction beside PANTRY instead of leaving the design system visible only as one layer inside the BREAD story. The index names TJ's role across the featured products, stack, design system, app and teaching platform. This audit found that its feature cards linked to the portfolio stories but hid the public repositories one click deeper. Source links now sit beside the story links for Greenroom, BREAD, GRAIN, PANTRY, and the course platform; its interactive demo remains a separate action. The classroom card also points to the teaching page, where visitors can inspect the courses, assessment approach, and badge criteria. Browser checks cover ownership copy, source and demo links, teaching handoff, wide-screen alignment, and narrow-screen stacking. The project index now provides both an explanation and a direct way to inspect the code, while the wider audit of each project's outcomes, evidence and next steps remains open. All five repository links and the interactive demo returned HTTP 200 on October 13. The cross-page claim review also found that the GRAIN showcase described its server SSE path as the deployed static transport. It now distinguishes the browser client door used on the static site from the server door shown by bun run dev:stub, with a browser assertion for both. The PANTRY install example now uses pantry init --kit, matching the public quickstart and the page's claim that the kit installs project instructions; the page check guards the flag. Its long install command also overflowed the desktop content pane, so the first-run steps now wrap in a responsive layout that was reviewed at desktop and phone widths and checked for overflow. The repeated page footer now uses one plain, human-accountable AI statement without the robot emoji or an unsupported claim about every commit. Builder exports retain GRAIN's own generated attribution, covered by the export checks; the two credits describe different kinds of authorship. The /plans board now keeps all four status columns readable: narrow desktop users can scroll the final column into view, and long path chips wrap so phone layouts do not gain sideways page overflow.

The October 13 follow-up connected the homepage's leadership and course-scale claims directly to the course platform story, its synthetic-data demo, teaching page, and leadership profile. The scale wording was checked against the public course-platform README, which says the platform runs in production across several live university courses and covers hundreds of student repositories per term. The course story now repeats that scope and links to the note explaining what the platform makes possible. Greenroom now links directly to its public engineering case study, beside the run demo and source. Desktop and 390px checks cover those evidence paths; the full case-story audit still needs each product's stated result, proof, and next step reviewed together.

The October 13 outcome-and-proof review checked the five Projects entries against their public stories and the available local repositories. Greenroom's case study and source explain the local app and deployable team edition; its visitor-facing run is explicitly seeded practice data, not a customer result. The BREAD request trace demonstrates the stack serving a real portfolio note, and the source plus layer pages explain the reusable architecture; that is a concrete integration proof, not an adoption or performance outcome. GRAIN's shared-control demo and action contract demonstrate the interaction model, while PANTRY's public quickstart and install-to-first-use path describe an installation tested in a separate project. The classroom story and public architecture document a production workflow across several courses and hundreds of repositories per term, but its public interactive demo is synthetic and cannot independently verify those operational counts. The four local repositories checked for this pass were clean on main and tracking origin/main; the course-platform repository is not present in the local repo collection. The current pages generally pair a clear problem or design choice with an inspectable artifact and a next link. The remaining proof gap is therefore narrower than missing project presentation: visitors cannot independently inspect customer usage for Greenroom, adoption of BREAD/GRAIN/PANTRY, or private classroom operation. Keep the existing privacy boundary. If stronger outcome proof becomes available, prefer a privacy-safe aggregate, a reproducible install/run transcript, or a clearly attributed customer or teaching case study; do not imply the seeded and synthetic demos prove production outcomes.

The About profile now places the course platform's production scale beside its own supporting link, so the 150–300 students-per-term claim leads directly to the platform story, architecture, and privacy-safe demo. The change passed the focused browser check and all 21 About browser cases. Its first full hosted CI run caught an older assertion that counted exactly three links across all role proof blocks; that selector had become too broad when a fourth, page-level evidence link was added. The selector now names the three role-specific links. Commit 83ae03e passed typecheck, unit tests, lint baseline, and the complete hosted E2E suite (425 tests); the Pages build and deployment passed, and the live /about route contains the new course-scale evidence. Its serial local full-suite run also passed all 425 tests. The parallel local run had four time-sensitive failures in desk animations and the long plan-detail crawl; all passed in serial mode. This increment is released, while the project-by-project outcome and evidence audit and Builder reliability work remain open below.

The October 13 role-introduction pass reshapes the About profile around three connected principles: as a teacher, treat students as adults and automate mechanical grading to return time to teaching; as an engineer, hold the architecture and quality bar while keeping direction legible to the next person; as a manager, grow people first and deliver goals through them. Each line links to its detailed role section. The homepage now opens with the same three principles, while the résumé summary uses that first-person framing and links to the full profile. The themes come from existing About role sections, the teaching and origin-story notes; no new biographical claims were added. Focused desktop and phone checks cover the three statements, role anchors, and résumé/homepage paths.

The first full route sweep covered 145 public routes at desktop and phone widths (290 checks). It found no route-level navigation failures, missing or duplicate page headings, or horizontal page overflow outside /catalog. The catalog's mobile grid had been allowed to size itself from wide component examples, its fixed-position live specimens could float over other entries, and its empty console-expand control had no accessible name. The portfolio now clamps that generated page to the phone viewport, confines each live specimen to its card, and names the control at the host integration boundary and supplies the catalog's missing search description, leaving GRAIN and BATCH internals untouched. Browser checks cover the phone width, full-page catalog navigation, and description. A refreshed headless audit then returned all 145 routes successfully, with canonical, Open Graph, structured-data, and description metadata present. The first report still counted generated catalog specimen titles as page headings. Those titles now keep their visual style without using heading elements, and that report found exactly one page heading on all 145 routes. The page-family review remains open.

The first desktop and phone review of the main visitor journey also found a stale fact in the talks section: the “Build the Floor” deck contains 33 slides, while the talks index said 27 and its note attachment said 32. Both public counts now match the deck, and a browser check compares every talk index count with its slide sections and checks the linked note's count.

The MILL introduction now links directly to a rendered note as well as its documentation. A visitor can inspect the content engine's output from the project page without first navigating the repository or reading the implementation guide. The landing-page browser check covers that evidence link.

The PROOF introduction now pairs the source markdown for a completed site-builder plan with the board's rendered plan detail. Its existing manual-refresh limitation remains visible, and a browser check follows the source-to-board links and verifies the completed status.

The developer-docs pass found pre-split package instructions that no longer matched the published repositories. The docs index and the BATCH, GRAIN, MILL, and PROOF setup guides now describe the npm packages, the portfolio's Pages workflow, and the difference between each library and the running app. Current PROOF package metadata still pins BATCH to a public GitHub commit, so the guide names that extra network dependency rather than claiming every install comes from npm. Phone-width browser checks cover the updated docs index and the BATCH and GRAIN package guides. The remaining docs sets still need their own content and instruction review.

The CRUMB introduction now leads with a direct link into the live portfolio tour. A browser check starts that tour from the project page and confirms the first guided step appears.

The PANTRY introduction now includes the documented install, scaffold, check, and serve commands next to its verified-install claim. It also states that the board reads the host project's plans in place.

On October 13, a fresh sitemap audit covered 148 canonical routes: all returned HTTP 200, and every route had one page heading, a description, canonical URL, Open Graph metadata, and structured data. The new Version One deck route and workshop attachment were checked at desktop and phone widths, and the full deck was reviewed for student repositories and records before its route was added to the canonical-page checks. The /batch introduction now reports that the linked audit covers 148 routes. The static export and dead-link verification pass with the workshop deck included.

The October 13 source-by-source documentation review found two capability overstatements: the GRAIN quick start said every surface was operable, and the BATCH static-export guide implied the client door exposed the same full vocabulary. They now describe registered surfaces and supported service-free actions. Browser assertions cover both statements. The three MILL guides have now been checked against the installed 0.4.0 package source and its current plan: the collection fields, route enumeration, renderer port, addressable surfaces, and deferred AI outputs match. The BATCH, GRAIN, and PROOF guides were then checked against the installed package sources. That pass found one remaining mismatch: the PROOF quick start said proof init writes a contract section into the host's CLAUDE.md. The package only creates files under plans/ and prints the manual wiring step. The guide now says that directly, and a browser assertion guards the boundary. The current setup, operator, and interaction instructions match the checked-out BATCH 0.2.1, GRAIN 0.1.23, MILL 0.4.0, and PROOF 0.1.4 packages. Deeper BATCH and GRAIN implementation review remains deferred as agreed.

The October 13 refresh of the 23-screen desktop gallery found two additional presentation issues: the Talks introduction described every deck as an interactive web page, despite the new event PDF, and the Standards index repeated the AI-REPO-STANDARD filename and summary in one awkwardly wrapped title. The Talks copy now distinguishes interactive web decks from the event PDF, and the standard has a shorter display title with its purpose in the summary. Focused browser checks cover both. The gallery now contains 35 captures: the original project and interaction views, six newly included desktop routes (Projects, Résumé, Teaching, Badges, Plans, and Version One), and first-screen phone captures for those six routes. Phone review confirms the content and primary controls remain legible. The PDF viewer's fallback is expected in headless Chromium; its direct link now has a browser check for a successful application/pdf response with a valid PDF signature.

The cross-page read of BREAD, BATCH, GRAIN, MILL, PROOF, CRUMB, and PANTRY confirms they now tell one consistent story: the BREAD page explains the dependency direction and the portfolio request path; BATCH, GRAIN, and MILL each state their own layer and link to adjacent layers; PROOF and CRUMB are shown at the same level above GRAIN and MILL; and PANTRY is clearly an installable app that mounts the stack rather than another layer. Each introduction links back to BREAD, documentation, and source or a live demonstration where applicable. Existing route and phone-layout checks cover the responsive presentation; deeper GRAIN and BATCH implementation work remains deferred.

The October 13 source check found one precision gap in that story: PANTRY's core docs app composes BATCH, GRAIN, MILL, and PROOF, while its optional toursDir mounts CRUMB's review-tour routes and client for projects being reviewed. The BREAD and PANTRY introductions now say so explicitly. Focused browser checks cover both explanations alongside their existing phone and install paths.

The project-story pass brought inspectable evidence onto the Projects cards: a real note request for BREAD, the working shared-control demo and action contract for GRAIN, and PANTRY's composition and first-use routes. The BREAD card also distinguishes the four packages I built from MILL, which the site uses. Four GRAIN implementation guides were checked against the installed package source for component locations, the closed RenderOp set and dispatcher lifecycle, token names, and the manifest contract. PANTRY's composition and install guides were checked against its docs and source so the optional preview-tour support stays separate from the four-layer core. The cross-page read across BATCH, GRAIN, MILL, PROOF, CRUMB, PANTRY, Greenroom, and the course platform confirmed that each project story leads to source, documentation, a working demo, a measured result, or a rendered example. One PANTRY link returned readers to the page top instead of the first-use commands; it now lands on that section directly, with the project browser check covering the target.

The badge certificate template pass found that “Copy share link” silently failed when clipboard access was missing or denied. It now tries the browser clipboard, then a copy-command fallback, and shows a selectable manual link if neither works. The phone-width browser check covers the normal copy and the unavailable-clipboard state. I reviewed the rendered certificate with personal fields masked; recipient records stay out of the audit notes. The same silent no-clipboard exit existed in event share blocks. They now try the same copy-command fallback and tell the reader to select the visible post text if copying still fails; the event-page browser check covers both paths.

The October 11 Builder audit scored 2/8 in each of two live 0.5B runs and 6/8 in one 1.5B run. The larger model handled both width changes, both explicit-ID moves, and refusal to rewrite existing copy, but failed both removal scenarios and added unsupported flourishes to its draft. Its first scenario took 40.5 seconds from a fresh cache, including model loading and generation; a browser profile occupied 1.6 GB after the run, which is not an exact download measurement. The Builder now selects 1.5B only on capable devices; the shared desk remains on 0.5B. A focused October 13 check applied both deletion requests after the review panel disclosed and corrected the model's wrong target for the second card. The same run's draft omitted Utrecht, so the reviewed workflow scored 2/3. That result measures target reconciliation, not model accuracy or reliable writing. Phone hardware and repeated grounded-copy measurements remain open. Copy on blocks already in the canvas cannot yet be edited.

The Builder now follows the part of Puck's AI approach that fits this site: a model chooses from GRAIN's code-owned blocks and can draft copy only for registered text fields on new blocks. The portfolio now sends that local request a JSON schema that bounds the block count, names, span, and copy lengths; the existing validator still filters each block's fields before rendering. Model quality remains an open gap. A local WebGPU run on October 12 scored 6/8 with the 1.5B profile, but the passing cases varied across runs. After the timeout increased to 60 seconds, a fresh-server run scored 5/8: both deletion cases failed, and the draft omitted Utrecht and added unsupported phrases. A warm repeat included the city but still added marketing language. A stricter prompt trial did not resolve the problem and was reverted. One earlier relative deletion timed out at 45 seconds; in the new run the first request took 50 seconds and the target guard refused the model's wrong card. The current in-app browser still has no WebGPU, so the live model remains untested in that browser. The local Playwright Chromium run supplies the live-model evidence. Existing copy cannot be revised because GRAIN does not yet expose a bounded text-write operation. The Builder names these limits in its visible guidance; reliable AI page writing remains unfinished.

On October 13, a fresh local 1.5B draft-only repeat again failed to retain Utrecht and added claims that the brief did not support. The draft review now shows the visitor's source brief beside the editable generated copy and asks the visitor to compare names, locations, and claims before adding it. A phone-width browser check confirms the brief remains readable without page overflow. This improves the review evidence but does not make generation grounded; repeated draft quality remains open in builder-ai-depth.md.

The owner confirmed that Builder generation must stay local. The page now says capable devices download about 875 MB of model files for the Builder, while other devices use about 350 MB; files are cached and prompt generation stays on the device. Shared example links open the predictable code-owned preview until a visitor presses Build it. Explicit counts now produce the requested number of blocks, card-specific layout language only places the cards side by side, and repeated cards receive distinct fallback copy. Unit and browser checks cover the four-block example, the two half-width cards, the distinct copy, and the local-generation notice. The narrower preview was also reviewed at 390 pixels with no page overflow.

The latest October 11 CI run on cf0af7a passed typecheck, all 705 unit tests, and 410 browser cases; it skipped one calendar case because its fixture has no single-photo event. Static export and export verification pass. The export contains 1,175 pages and 1,132 data routes, and every exported internal link resolves. The figure-loader assertions now wait for dynamic imports, so a late request no longer looks like a missing feature. A run with four workers produced seven failures under parallel load; the serial run is the reliable gate for this suite. The export needs PUBLIC_ORIGIN set to the deployed URL to avoid localhost in generated endpoints; the Pages workflow sets it. The lint gate now reports 4,491 flags, 79 below its previous high-water mark; the ratchet baseline has been lowered to that count. The link linter finds no dead relative links across 56 rendered files, and every served diagram has a committed SVG. The October 13 route audit returns all 148 canonical pages without HTTP or document findings. The October 10 BATCH and MILL architecture edits were rendered at desktop and 390 pixels; both pages have one page heading and no horizontal overflow. A refreshed 23-screen desktop gallery was available for review on October 11; it has since grown to 35 desktop and phone captures. The screenshot helper isolates browser storage for each capture and waits for the welcome greeting to finish. The GRAIN guide's package-boundary section now reflects the separate BATCH repository and the GRAIN monorepo that publishes GRAIN, MILL, PROOF, and CRUMB; package-doc navigation and phone checks pass. The Builder help now reports both October 11 local-model profiles and their measured limits, with the full comparison recorded in builder-ai-depth.md. The visual route review covers the 146-route desktop and phone captures and the refreshed 23-screen gallery; the two newer routes also pass focused desktop and phone checks. The October 13 source review of the BATCH, GRAIN, MILL, and PROOF instructions found and corrected the stale PROOF init claim described above. The remaining product work is Builder reliability. A production transfer check closed the shared-shell delivery decision: fresh Chromium contexts on October 11 received 136,617 JavaScript bytes on / and /notes (123,417 encoded response bytes), and 159,108 bytes on /grain/builder (143,208 encoded response bytes), before any model download. GitHub Pages serves the scripts with gzip. Keep the shared shell: its site-wide desk, navigation, catalog, and interaction controls are available on every page, while page-specific modules remain deferred. The previous local audit values were uncompressed response sizes, so its report now labels them as local build-size checks rather than visitor transfer estimates. The catalog's GRAIN sidebar specimen chat fields now receive an accessible name and placeholder at the portfolio composition boundary, with a browser check covering both rendered fields. The October 10 route-specific island pass deferred the image viewer and stylesheet, interactive figure setup, and reading progress on pages that do not use them. The local report measures 30–268 KB of uncompressed JavaScript responses per route; the calendar, long notes, talks, and Builder load their additional features only where needed. Production browser captures show the ordinary page's compressed transfer and the Builder's additional route modules; the shell decision is closed unless field measurements show a meaningful user-facing cost.

The October 12 documentation follow-up found that the MILL introduction still called it a CMS even though its package plan had retired that label, and that the GRAIN quick-start described every click as a server POST despite GRAIN's static client door. Both descriptions now match the package contracts, with browser assertions. The BATCH route guide's shell command also pointed at a portfolio-named directory instead of the default pages/ root; its example now works from the repository root and is checked in the docs browser suite. The full guide-by-guide review remains open. The next architecture review found that BATCH's bun run export wording attributed the portfolio-owned command to the package. The guide now distinguishes BATCH's exportSite() library from the host's command, with a browser assertion for both sides of that boundary. The PROOF setup guide also called chrome the only host input even though plansDir is required; it now separates the presentation seam from the data directory, with browser coverage for the distinction. The baseline ratchet also recorded 79 fewer flags than its previous high-water mark, without increasing any category.

The follow-up BATCH route-guide review checked makePageServer and createSitemap in the package against the portfolio's pageRoutes() exporter. The guide had implied every page was exported without exceptions. It now names the review-only and operable exclusions, the generated routes that need explicit entries, and the separate sitemap/export entries required for routes outside the page tree. Its phone-safe docs browser test verifies those boundaries. The remaining BATCH guides and the GRAIN, MILL, and PROOF guides still need their full source-by-source review. The same comparison found that the BATCH introduction implied the package generated robots.txt; its API exposes the route list and sitemap XML, while the host writes robots.txt. The introduction and architecture guide now state that boundary, and the browser suite checks the visible setup copy. The export guide also understated the route sources and described assets as verbatim despite base-path rewriting; it now names the portfolio's content, deck, plan, and generated route sets and explains CSS and JavaScript rewriting for subpath deploys. The older architecture summary now also distinguishes rewritten CSS and JavaScript from byte-for-byte binary copies. The docs browser suite checks these statements.

The GRAIN component guide still named the package's pre-workspace grain/components path and mixed the GRAIN monorepo's package checks with the portfolio's local preview server. It now names packages/grain/components, gives the package-scoped check and test commands, and explains that the portfolio catalog displays the package version its own config resolves. The documented workspace commands passed, and the docs browser test covers the path and preview boundary.

The GRAIN tutorial's copied contract excerpt had fallen behind the package registry: it omitted the check and block surface kinds plus the check.set and three block-edit verbs. The excerpt now matches the current closed unions, and the docs browser suite checks those entries against the rendered guide.

The GRAIN action guide still opened with task:42 and task-list, even though task is not a registered surface kind. Its examples now use addresses present in the portfolio, and label the console as push-only. The docs browser test checks the examples and confirms their code block fits at 390 pixels.

Page recommendations in progress

The October 12 cold Builder run used a new Chromium profile and received all 30 weight shards without cached responses. Chromium reported 875,219,838 encoded transfer bytes for the larger model's weights, tokenizer, configuration, index, and WebGPU library, plus 1,661,197 bytes for WebLLM's JavaScript runtime. The public download notice now rounds the model assets to 875 MB, replacing the earlier 1.1 GB estimate. The draft was visible by the 60.1-second observation and retained the requested facts, but added unsupported claims about artisanal bread and a staple product. This closes the desktop download measurement, while phone hardware and grounded writing remain open. The capture method and result are recorded in audit/builder-cold-2026-10-12.md.

The MILL source review also found inconsistent rendering and dependency descriptions. Its project card, setup guide, and architecture guide now distinguish GRAIN's package dependency from BATCH's optional host composition. The architecture explains semantic HTML output, includes the implemented link-attribute and heading-surface options, and no longer calls addressable headings wholly planned.

The October 12 operable-surface checklist review found an incomplete action definition: its example omitted the required description, payload schema, and behavior hints. The example now includes all three and explains how the manifest uses them. Its verification checklist also distinguishes the server POST/SSE path from the client door used by the default portfolio development command and static export, including the limit on server-dependent capabilities.

This pass passed 83 focused browser cases across documentation, Builder, and project presentation, four model-profile unit tests, typechecking, and link lint. The lint gate reports 4,477 flags against its 4,491 baseline. Desktop and phone screenshots were reviewed for the updated MILL introduction and operable-surface guide, alongside the Builder download disclosure.

The audit covers all 148 public routes and their shared page families at desktop and phone widths. The table records each page's recommendation and its current state. Page presentation and evidence paths have been reviewed, but some production and adoption claims remain owner-reported because the underlying records are private or unavailable. Builder reliability and stronger privacy-safe outcome proof remain open below.

Page recommendations and current state

Work

• [x] Reconcile the portfolio's purpose and the visitor journey with the owner on 2026-10-02.
• [x] Reconcile the shipped builder and credential work with the board, keeping ongoing AI and human-lane work visible under its actual status.
• [x] Archive the empty welcome-plan template so it no longer reads as unfinished portfolio work.
• [x] Close the completed builder v1 plan and move its remaining model-quality work into builder-ai-depth.md.
• [x] Audit the entry pages and project stories together. The home page establishes my role and course-platform scale; Projects names my ownership of the stack and teaching platform, then links to the detail pages. BREAD, BATCH, GRAIN, MILL, PROOF, CRUMB, PANTRY, Greenroom, and the classroom page explain their purpose and design constraints, with a live use, measured result, public source, documentation, or interactive demo to inspect. The portfolio and projects pages were reviewed at desktop and phone widths; e2e/projects-layout.e2e.ts covers the key evidence links and responsive group layout. This pass closed the missing source paths from the Projects cards. The source audit also confirmed and corrected the BREAD/PANTRY distinction between PANTRY's four core layers and its optional CRUMB review-tour mount. Focused checks now guard this relationship on both introductions. Projects now links directly to the BREAD note request trace, GRAIN's shared-control demo and action contract, and PANTRY's composition guide and install story. The BREAD card distinguishes the four packages TJ built from MILL, which the site uses. Desktop and phone browser checks cover these links and card layout. The cross-page read confirms that every project's stated result leads to inspectable evidence and a clear next step, and the PANTRY install example jumps straight to its first-use commands.
• [x] Review the BREAD, BATCH, GRAIN, MILL, PROOF, CRUMB, and PANTRY introductions as one explanation of the stack. Their dependency direction, the request path through the portfolio, PANTRY's role as an app, and cross-links to adjacent layers are consistent; existing route and phone-layout checks cover their presentation. Deeper GRAIN and BATCH implementation work stays deferred.
• [x] Review the repeated page-footer copy and GRAIN attribution as one system. Keep the authorship claim honest and consistent across the portfolio and the pages visitors can export. The portfolio footer now uses one tested statement; exported pages retain GRAIN's generated framework attribution and are covered by the builder export checks.
• [x] Keep Builder generation local and the composition shape code-owned. The local model chooses registered block names and their order through JSON mode. Code ignores model-written copy, builds visible text from phrases in the visitor's brief, and leaves unsupported fields blank. The visitor can edit and review that copy before adding it. Repeated warm WebGPU audits cover six briefs, field placement, required phrases, and prohibited wording. The 1.5B and 3B trials still produced unsupported raw prose, so the Builder does not use it.
• [x] Review Notes, Talks, Teaching, Calendar, and Badges as evidence of engineering, communication, teaching, and the path between them. Notes point into projects and teaching, talks link to their written methods, Teaching explains its issuing authority and links to the platform, calendar, and criteria, Calendar entries point to related proof where available, and badge pages link back to their issuer and criteria. The talk counts, teaching evidence, calendar links, badge privacy, and responsive routes have focused browser coverage.
• [x] Complete the Builder's measured AI reliability work in builder-ai-depth.md, then make it a convincing demonstration of the system the portfolio introduces. The current design keeps the model local and uses it to choose registered blocks and order. Code assembles visible copy from the brief, and review catches omissions before the visitor approves the page. Six briefs passed two warm 1.5B runs for required phrases, field placement, and prohibited wording. The held-out edit audit now covers twelve phrasings across removal, span, and move, with 34 reviewed outcomes across warm WebGPU runs. A new routing audit covers copy-field edits, insertions, field moves, possessives, terse commands, missing ordinals, and new-page briefs that mention definite titles and blocks. The 3B comparison did not improve the raw answers, so the Builder ignores model prose instead of displaying it. The bounded visitor workflow is measured; a 58-scenario live portfolio run passed all checks, including 35 Builder edit, refusal, and composition outcomes and 23 DESK and navigation outcomes. The complete raw and visible-result capture is here. This closes the defined workflow audit, not open-ended language understanding: raw model prose is unsupported and intentionally excluded from the visitor-facing page. Responsive browser checks satisfy device acceptance, and physical-phone performance is outside this goal. See the two-run draft report, original edit paraphrases, extended edit paraphrases, delete, nudge, and width holdouts, the target-guard baseline, the six-run guard recheck, and erase and reposition baseline, erase and reposition recheck, and 3B comparison, and the expanded composition matrix.
• [x] Keep every /plans status column readable at desktop and phone widths. The narrow desktop keeps horizontal scrolling inside the board and brings the blocked column fully into view; phone layouts stack the columns and wrap long path chips. e2e/plans-layout.e2e.ts covers both.
• [x] Finish the route-by-route visual and interaction review at desktop and narrow widths. Check navigation, page spacing, headings, accessible names, dead ends, and important interaction paths. On October 13, all 148 canonical routes were captured at desktop and phone widths in fresh browser contexts. The desktop pass also captured full-page views. Every route returned 200, had a title and one page heading, and fit within the viewport; the screenshot review found no new layout issues. The PDF routes show the direct-open fallback in headless Chromium. The existing gallery retains its 36 selected captures, including phone first-screen views for seven key routes. Focused phone checks cover the Builder, project stories, documentation, résumé, plans, catalog, and navigation. The serial full release run passed all 421 browser cases. The October 13 hosted CI run on commit d2df50a passed all 425 browser cases in 4.0 minutes, alongside typecheck, 712 unit tests, and the lint baseline. Link lint found no dead relative links in 56 rendered files. The local performance report measures 30–268 KB of uncompressed JavaScript responses per route; production captures measured 136,617 JavaScript bytes on ordinary pages and 159,108 bytes on the Builder before model download. The shared shell stays available across routes, while page-specific features load only where needed. The local report labels its values as build-size checks rather than production transfer sizes.
• [x] Commit the finished portfolio on main, deploy it, and verify the published pages after the remaining Builder reliability work is complete. The October 13 project-story increment d2df50a is on main and deployed: hosted CI passed all 425 browser cases, typecheck, 712 unit tests, and the lint baseline; the Pages build and deploy passed; and the live Projects and PANTRY pages include the direct evidence paths. The October 13 Builder and Badges increment 8fcdbc5 is also on main and deployed. Hosted CI passed all 432 browser cases, typecheck, 724 unit tests, and the lint baseline. The Pages build and deploy succeeded, and a browser check of the published Builder and Badges pages confirmed the updated review disclosure, subject descriptions, section cards, and phone-width fit. The full finish-line release remains open because the Builder still needs broader brief and device reliability evidence and a cold-start measurement on an actual phone. Stronger privacy-safe outcome proof for several project claims is still unavailable. The October 13 outcome-and-proof review records which project artifacts are reproducible and which production/adoption claims remain owner-reported or private. The October 14 negative-anchor fix passes 726 unit tests, all 73 Builder browser cases, link lint, and the lint baseline. Commit 1b8b5a1 is on main and deployed. Hosted CI passed all 432 browser cases, 726 unit tests, typecheck, and the lint baseline. The Pages build and deploy passed, and the published Builder route and its bundled code were checked after deployment. The guard's targeted phrase-and-year case passed one live WebGPU run. This closes that specific regression; broader writing reliability and a cold-start measurement on actual phone hardware remain open. A follow-up audit added the common instructions “avoid,” “exclude,” and “leave out,” plus a positive phrase after the prohibited wording. The unit and live WebGPU checks pass; this follow- up was released in commit 76e0c9d. Hosted CI passed all 431 browser cases, 727 unit tests, typecheck, and the lint baseline. The Pages build and deployment passed, and the published Builder module contains the new instruction forms. Both live negative-instruction checks passed. These targeted checks do not establish broad writing reliability or general negation understanding; the actual-phone cold start remains open. Documentation commit e780d66 records this release; its CI and Pages deployment passed, including 432 browser cases and 727 unit tests. A third live WebGPU brief combines three prohibited details with a required positive intro phrase; the guarded result passed. The actual-phone and broader raw-writing gaps remain open. Follow-up commit ee8c36d keeps sentence-initial “exclude,” “leave out,” and “without” out of the positive-anchor set. Its hosted CI passed all 432 browser cases, 728 unit tests, typecheck, and the lint baseline; Pages deployment passed. The live Builder module was checked after deployment, and its third negative-instruction brief passed. The wider Builder reliability and actual-phone gaps remain open. A further audit found that copy grounding could reuse a forbidden source phrase even after anchor filtering. The grounding step now applies the same negative-clause filter and stops a candidate span before a later prohibition. A regression confirms that “Hearth Bakery” survives when a raw copy span continues across the sentence boundary into “Do not use the phrase ‘award-winning bakery.’” A separate check preserves a positive phrase after “and include.” Two live WebGPU rechecks rejected raw “organic” suggestions and retained the requested intro; a third recheck after the span fix also passed. Commit af142b7 is published on main; production serves the sentence composer and revised review disclosure. Local verification passed, as did hosted typecheck, unit tests, lint count, and the Pages build and deployment. Hosted end-to-end passed all 432 browser cases on commit f170865. Broad raw-writing reliability and the physical-phone cold start remain open. A follow-up then addressed a visible copy-quality defect: some safe drafts were rendered as phrase lists. In six warm WebGPU scenarios, source reconciliation retained the requested facts; a separate exact-phrase check confirmed a complete intro sentence and excluded the forbidden claims. Explicit intro descriptions, exact intro phrases, and welcome requests now form short sentences from the supplied details. Local verification passed 736 unit tests, 74 Builder browser cases, typecheck, link lint, and the lint baseline. The review now distinguishes verbatim wording from a sentence composed from details in the brief and retains its warning that the checks cannot verify claims. The work is documented in the Builder sentence audit and was released in commit af142b7. Production serves the new sentence composer and revised review disclosure. Pages build and deployment passed; hosted end-to-end passed all 432 browser cases on commit f170865. Raw model output still invents claims or misplaces wording, and the cold start on an actual phone remains unmeasured. A six-scenario source-quote prompt trial still produced unsupported raw details; all guarded drafts passed after reconciliation. The trial did not establish a raw-writing improvement, so the temporary prompt rule was removed. See the prompt experiment record. A separate six-brief zero-temperature run still produced unsupported details and proposed a prohibited “organic” claim in the raw answer. Reconciliation removed it from the reviewed draft; temperature 0.2 was restored. See the temperature experiment record. The evidence update is on main as 69f5bf3. Hosted CI passed typecheck, 736 unit tests, the lint baseline, and all 432 browser cases. The Pages build and deployment succeeded, and the published Builder and plan routes both return HTTP 200. These release checks do not close the raw-writing or physical-phone gaps. The next hosted run on 0f5b7e4 reported one flaky first attempt: the scripted Builder test submitted an edit before its fake model had initialized and received the page's honest offline response. The test helper now waits for that model before sending model-backed edits. The refusal case passed three local repetitions, and all 73 Builder-canvas cases passed locally. Hosted CI on 8d39c1a passed all 432 browser cases with no retry, 736 unit tests, typecheck, and the lint baseline. The Pages deployment succeeded. This stabilizes the test evidence; raw-writing reliability and the physical-phone cold start remain open. A fresh full 36-scenario local WebGPU run then passed 32/36 overall, with all 13 Builder edits, refusals, and drafts passing. Four failures belonged to other desk scenarios. The raw Builder answers still added unsupported claims or blocks in five of six drafts; source reconciliation removed them before review. A stricter live grader now requires the complete mixed-prohibition intro sentence in its named field; three warm repeats passed 3/3 with the same visitor-facing result. The raw model still suggested unsupported phrases, which source reconciliation removed. These results strengthen evidence for this guarded Builder path without proving broad raw model quality. The physical-phone cold-start measurement remains open. See the full Builder audit and the three sentence rechecks. The authorized local Qwen2.5-3B comparison also passed the six guarded drafts, but its raw answers still included unsupported copy or malformed block names and warm responses took 17 to 76 seconds. The 1.5B profile remains shipped; this model-size comparison is closed without a promotion. Its reports are here and here. Raw model writing remains unreliable, but the Builder discards that prose and builds visible copy from the visitor's brief. The open Builder checks are broader natural-language coverage and a physical-phone cold start. A fresh two-run audit of the six 1.5B Builder briefs passed all 12 visitor-facing drafts, including field placement, required phrases, and prohibited wording. Raw answers continued to invent service history, staff experience, sourcing, opening claims, and unrequested blocks; source reconciliation removed or corrected them before review. This confirms repeatability of the guarded desktop experience, not reliability of the model's raw writing. See the two-run report. A 320px browser recheck found and fixed a wrapped Build it label and a long reliability note that pushed the workbench below the first screen. The short local-model and download note stays visible; detailed evidence is now collapsed with a link to the Builder plan. The focused browser regression passed, and the updated screenshot shows the workbench and example above the fold with no horizontal overflow. The cold-start evidence still needs an actual phone.

Builder paraphrase check, October 14

I added guarded recovery for “get rid of the final card” and “let the callout span the whole row.” Three independent warm WebGPU runs passed both outcomes in the reviewed flow. In all six cases, Qwen2.5-1.5B returned no usable action; the Builder disclosed its request-based proposal and waited for approval. The 38 focused reasoner tests, all 740 project unit tests, and all 75 Builder browser cases passed, including the approval flow at a 390px viewport. Typecheck, link lint, and the lint ratchet passed at the existing baseline. The combined live capture records model output beside each resulting canvas. This closes only these two tested phrasings; broader paraphrase reliability and the cold start on physical phone hardware remain open. The model's generated prose is still not used for visible copy, so these results make no claim that it can write reliable page copy. A fresh device check found no attached Android handset, no phone or tablet in the USB inventory, and no available simctl for an iOS simulator.

I extended the recovery check with three more phrasings: “take out the callout,” “let the callout fill the whole row,” and “shift the callout up.” Three separate warm WebGPU runs passed all nine reviewed outcomes. The model supplied no usable action for all six removal and width requests; it handled the move itself in two of three runs. The new capture records the raw answers. The 40 reasoner tests, 742 project unit tests, and 75 Builder browser cases passed, as did typecheck, link lint, and the lint ratchet. Broader natural-language reliability and the physical-phone cold start remain open.

The next held-out requests were “discard the final card,” “nudge the callout down one step,” and “stretch the callout to full width.” Across three warm WebGPU repeats, Qwen2.5-1.5B declined the first two each time; the guarded recovery proposed the explicit edit and waited for approval. The model handled the width request directly. Negative and ambiguous requests stayed refused, and the canvas remained unchanged before approval. The raw answers and resulting canvases record the distinction between model output and the reviewed visitor flow. The reasoner tests passed 42 cases, the full unit suite passed 744, and all 76 Builder browser cases passed. Typecheck, link lint, and the lint gate passed. The change is on main as c69fc17, and Pages deployed it; hosted typecheck, unit tests, and lint checks passed. Hosted end-to-end passed all 435 browser cases. Natural-language coverage beyond these measured phrasings and an actual-phone cold-start check remain open.

The next holdout checked “delete the intro” and “shrink the callout to a third width.” Across three warm WebGPU runs, the model declined deletion each time and handled the third-width request directly each time. The first width trial uncovered a target-guard bug: “third” was treated as the ordinal of the callout, so a valid model proposal was rejected. The resolver now treats ordinals as block references only when they qualify a block name, and recognizes “shrink” and “reduce” for explicit width values. The six-run capture records raw answers and resulting canvases. The regression test checks that both proposed edits wait for approval and then reach the intended block at a 390px viewport. Local verification passed 43 reasoner tests, 745 unit tests, and all 77 Builder browser cases; typecheck, link lint, and the lint gate also passed. Commit 30656f1 then passed hosted typecheck, unit tests, lint checks, and all 435 browser cases. Its Pages deployment succeeded, and the published finish-line plan and Builder routes return HTTP 200. A production browser smoke at 390 by 844 showed the Builder heading, composer, and canvas, with a 390-pixel document width and no page errors. The baseline failure and six-run recheck preserve the failed target check and the corrected model outcomes.

Current release and visual review, October 14

The latest functional Builder changes are on main at 47d1f55. Hosted CI passed typecheck, all 818 unit tests, the lint count, and all 438 browser cases. The Pages build and deployment succeeded. The published Builder and Plans routes return HTTP 200. Locally, all 79 Builder browser cases passed, and link lint found no dead relative links in 56 rendered files. A prompt audit closed copy-field routing holes across possessive, block-first, insertion, movement, and terse phrasing, while preserving new-page briefs that mention a title before their blocks. The paired browser test confirms that copy edits do not call the model or change the canvas. A local Qwen3-1.7B comparison passed five guarded Builder scenarios, but one relative deletion had the wrong raw target and succeeded only after the Builder's correction path; action responses took about 19 to 54 seconds. This does not justify replacing the shipped 1.5B profile. The screenshot gallery includes a 390 by 844 pixel Builder capture; all 36 portfolio captures were refreshed and the phone-width view was inspected. A 320-pixel browser case checks that the composer and workbench do not overflow horizontally. The production route was inspected at 390 by 844; the heading, composer, and canvas were visible, the document fit the viewport, and the browser logged no page errors. These are browser checks, not a physical phone or its model cold start. This Mac has no attached handset, and its iOS simulator is unavailable.

The October 14 routing follow-up in 86bc4d9 closes direct copy-field gaps found during the local model comparison. Requests to remove or write titles, move images, or resize text route to the current page's edit guard. New-page briefs that mention a title, menu, or callout remain compositions. The full unit suite passed 824 tests, all 79 Builder browser cases passed, and hosted CI passed all 438 browser cases, typecheck, unit tests, and the lint baseline. Pages built and deployed successfully, and the Builder and Plans routes returned HTTP 200. Qwen3 remains a local comparison; the shipped model profile is unchanged. Physical-phone cold-start testing and broad language coverage remain open, so the finish-line goal stays in progress.

More bounded edit phrasings, October 14

The recovery grammar missed “eliminate,” “clear out,” and “take away” for removal, “expand” and “narrow” for explicit widths, and “push” and “slide” for one-step moves. It now recognizes those words only when the request names one live block and one supported operation. Negative requests, ambiguous targets, and widths without an explicit supported value still do not recover into an edit. The full unit suite passed 827 tests across 45 files, including the 48 reasoner tests. All 80 Builder browser cases passed, and TypeScript checking passed. The browser cases verified that each recovered edit stays unchanged until the visitor approves it. This measures guarded request recovery, not the model's ability to understand those phrasings. A live preview attempt stayed at “Reading the page…” and returned no model result, so it did not add live evidence. Broader language coverage remains open. The run report records the test results and their limits. The cold-start measurement still requires a physical phone.

Negated Builder edit safety, October 14

A safety probe found that “I can't remove the callout” and “I shouldn't remove the callout” could become deletion proposals after punctuation normalization split their contractions. A second probe found that “Should I remove the callout?” also became a deletion proposal. Follow-up probes found the same problem with “Would it be better to remove the callout?” and “Is it better to move the callout up?” Later questions such as “What happens if I remove the callout?” and “How would it look if I make the callout full width?” also became edit proposals. “Can I remove the callout?” and “Would it help to move the callout up?” exposed permission and advice variants. Recovery now refuses common negative modal forms and keeps advice, informational, permission, and hypothetical questions out of the edit proposal path while preserving polite commands. The focused reasoner suite passed 51 tests, the full unit suite passed 830 tests, and all 81 Builder browser cases passed. At 390 by 844, one browser check confirmed that three negated requests and sixteen questions create no proposal. Another returned valid remove, width, and move commands for six question-shaped prompts; the Builder refused every command and kept the canvas unchanged. A follow-up probe found five uncovered forms: passive permission questions such as “Can the callout be removed?” and recommendation questions such as “Would you recommend removing the callout?” The proposal-boundary guard now rejects these alongside “Should the callout be removed?” and “Would the callout look better at full width?” The reasoner suite passes all 51 cases, and the 390 by 844 browser test confirms these additional questions create no proposal even when the model returns a valid command. This closes the identified question-form gaps, while broad natural-language reliability remains open. These scripted checks do not establish how the live model interprets these phrases, so broad natural-language reliability remains open. The 390px browser review satisfies responsive layout and visitor-flow acceptance. A physical-phone cold-start measurement is separate device-performance evidence and can be added if that becomes a product requirement. The run report records the exact scope and evidence.

Builder composition schema follow-up, October 14

The build path now sends a closed JSON schema to the local model. It allows only registered block names and supported spans, and rejects generated copy fields. The repeated six-scenario WebGPU matrix produced schema-conforming raw responses in all 12 runs, and all 12 visitor-facing drafts passed review. In the two intro-only exclusion cases, the raw model still selected an extra card; the reconciler removed it before review. Inspection found that the prompt example itself included a lede and card, contradicting the intro-only briefs. Removing that example led to four live repeats returning only a lede, with no block removed by reconciliation. The recheck records the raw outputs. The expanded visitor-flow matrix then repeated five brief shapes twice: negative-only and mixed-prohibition intros, one card, one callout, and two titled cards with separate phrases. All 10 reviewed drafts passed. The reconciler removed unrequested raw structures in the single-card, callout-only, and two-card cases. A copy-targeting fix now keeps the callout phrase in its body and routes each repeated card phrase to its own card; the captured drafts show both card bodies correctly. The small model still adds default blocks to some raw plans, so this evidence supports the guarded, reviewable visitor experience for these prompts, not independent model reliability. The expanded capture records the raw outputs and reviewed drafts. Broader language coverage remains open. The tracked matrix, exclusion recheck, and run report preserve the prompts, outputs, and gate limits. Responsive browser review remains the device criterion; physical-phone cold-start performance is separate and is not required for this goal.

Builder live composition follow-up, October 14

Five fresh runs on the deployed Qwen2.5-1.5B-Instruct-q4f16_1-MLC profile covered a pottery studio, a mixed-prohibition bakery brief, two bike-shop phrasings, and a repeat bakery brief. All five reviewed drafts passed: requested names and phrases appeared in the intended fields, explicit prohibitions were absent, and no unrequested block remained. In the pottery case, the raw answer added unsupported promotional language; the Builder's source-phrase review removed it before the visitor could approve the draft. That distinction remains central: this run supports the reviewed visitor flow for these scenarios, not reliable raw model writing or broad semantic understanding. The five-case live report from this earlier pass is retained locally in .cache/desk-audit and is not a published audit artifact. The newer tracked matrix and follow-up above preserve the current schema-constrained evidence.

The production Builder was reviewed at 390 by 844 in a real browser. The page fit the viewport at 390 CSS pixels, showed its heading, composer, empty canvas, and block rail, and logged no browser errors. The owner confirmed that responsive browser behavior is sufficient for device acceptance; physical-device model cold-start performance is outside this goal. The review tour points to the live-evidence review and the phone-width canvas. The expanded live composition matrix passed all 10 reviewed drafts, while the smaller model still sometimes proposes extra blocks and broader language reliability remains unproven. Stronger privacy-safe outcome proof also remains unavailable for several project claims. The page audit records those evidence limits and keeps the claims attributed; it does not turn private or unavailable records into a portfolio action item.

Previous release acceptance, October 14

The previous release passed the criteria recorded at that point in this plan. Responsive browser checks at desktop and 390-pixel widths satisfy the Builder's device acceptance. No physical phone or cold-start measurement is required for this goal. The measured Builder workflow meets its completion criteria: 12 raw composition responses matched the closed schema, 10 of 10 reviewed drafts passed across five shapes, supported edit and refusal scenarios were repeated against the live model, and the review gates keep unsupported copy and structure out of the approved page. Broad language coverage and raw-model accuracy outside these scenarios remain unproven and are not claims the portfolio makes. Project-outcome evidence was reviewed page by page. Public artifacts are linked, while customer use and private classroom operations remain attributed or labeled according to their limits; no private data is exposed to strengthen those claims. Commit 9b70254 is on main. Hosted CI passed typecheck, unit tests, lint-baseline validation, and the full browser suite. The Pages build and deployment succeeded, and the Builder, Plans, and live-evidence tour routes returned HTTP 200. The checkout was clean and tracked origin/main at the time of that release.

Builder negative-block reconciliation follow-up, October 14

A code review found a gap between the composition prompt and reconciliation: the prompt told the model that a block named in a prohibition was not a request, but reconciliation treated that same mention as a positive block count. A model could therefore return a prohibited block and have it survive the review draft. Reconciliation now removes prohibited block mentions from the positive count, filters a type that remains prohibited from the model's plan, and ignores quoted copy while classifying structural requests. If the brief later explicitly requests that same type, the positive request takes effect. The first adversarial live brief also found that the welcome-intro parser could absorb the next callout's requested wording and leave its body empty. Copy extraction now stops at a new block mention, while the explicit welcome sentence and callout phrase are restored to their own fields. The live grader now reads the actual review-panel headings, including blocks with no text fields. Unit coverage checks direct prohibitions, a prohibition without other blocks, a later positive override, quoted wording, Unicode before exclusions, and adjacent block copy.

The live seven-case WebGPU Builder matrix passed all seven reviewed outcomes. It covered the prior intro, single-card, callout, two-card, mixed-prohibition, and pottery briefs plus a new brief asking for an intro and callout while prohibiting a card. In five of the six existing briefs, the model proposed one to three unrequested blocks; reconciliation removed them before review. The sixth model answer matched the requested intro. On the new brief, the model returned only the intro; code restored the requested callout with “Open daily” in its body and left the prohibited card out. The direct unit case supplies the adversarial model answer containing a prohibited card and confirms that reconciliation removes it. The tracked live capture records the raw plans and visitor-facing drafts. The full Builder browser suite passed 81 cases, the full unit suite passed, and type checking, link lint, and diff checks passed. Responsive browser coverage remains the device criterion; physical-phone performance is outside this goal.

This closes the identified negative-block reconciliation gap. It does not establish broad language understanding or reliable raw model writing; those remain open and require more held-out tasks before the Builder reliability item can be closed.

Portfolio AI reliability closeout, October 14

The full live WebGPU scenario set passed 58/58 on the current code. It records the selected model profile, raw answer, visible result, timing, and browser errors for 23 DESK and navigation scenarios and 35 Builder scenarios. The latter include deletion, width, movement, explicit negative requests, and reviewed composition with required phrases, field placement, and prohibited blocks. In several composition cases, the model proposed extra blocks; reconciliation removed them before review. In the relative-placement case, the model returned no usable edit, so the Builder proposed the single adjacent move stated by the visitor. The proposal remained review-only until approval.

That held-out placement check found an uncovered explicit form: “Put the callout after the second card.” The request guard now resolves both blocks and recovers a move only when the destination is exactly one position away; ambiguous targets and longer moves remain with the model. A live edit also confirmed that an incorrect model target for “Set the intro to half width” is corrected to the explicitly named intro and shown for review. “Do not remove the callout” left the canvas unchanged. Focused reasoner tests passed 52 cases, and TypeScript checking passed. The full raw capture preserves the run.

This completes the measured portfolio-local AI workflow described in this plan. The DESK remains on the shared 0.5B profile because the 1.5B comparison did not improve the portfolio-wide task set enough to justify its cost. The Builder uses the 1.5B profile only behind its existing capability gate, and continues to build visitor-facing copy from supplied phrases because the tested models invent claims and unrequested structure. The review step exposes omissions; it does not certify truth. The audit does not claim open-ended language understanding, raw generated-copy quality, or physical-phone performance.

Done means

The home page gives a clear first read. The project stories and stack pages explain the work in plain language and lead to inspectable evidence. Notes, talks, and teaching work feel like part of the same portfolio rather than separate shelves. The builder's AI path has measured evidence and honest limits. Every public route in scope has been reviewed at desktop and narrow widths, the navigation has no dead ends, and the final deployment matches the reviewed result.

Builder alias and docs presentation follow-up, October 14

The no-model Builder path previously used a separate word matcher and could lose explicit requests or prohibitions when another block type was named. The bounded block vocabulary now supplies the aliases used for matching, reconciliation, and copy placement. Announcement maps to a callout, and contact-form aliases remain explicit alongside other requested blocks. The no-model path and the model-error fallback now pass their word-matched plan through the same reconciliation and source-copy checks as model plans. A live Qwen2.5-1.5B run passed both announcement and contact-form scenarios. In the announcement case, code removed three unrequested raw model blocks; in the contact-form case, the requested name and email controls remained. A three-requirement no-model browser case also verifies intro copy, the exact announcement phrase, two contact inputs, no card, and honest offline status. The initial announcement attempt hit a WebLLM constrained-decoder error before returning a response; the fallback path handled the unavailable answer, and two warm repeats plus the combined two-case run passed. The raw reports are tracked in audit/, including the initial runtime error. The full Builder browser suite and broader held-out language coverage still determine whether the reliability item can close.

A presentation issue reported on the CRUMB docs exposed a shared MILL-rendering gap: fenced source blocks had language labels but no syntax tokenization. The portfolio's shared Markdown adapter now renders common TypeScript, JavaScript, JSON, shell, YAML, Markdown, HTML, CSS, Python, SQL, and diff fences with server-side token spans. Unknown languages and Mermaid source remain escaped text. The syntax colors use the site's theme tokens, with weight and muted tones preserving distinctions in the monochrome theme. A browser assertion checks the actual CRUMB TypeScript example; unit coverage checks escaping and unsupported-language fallback. The rendered CRUMB code example was reviewed in the browser at desktop width.

The combined Builder and documentation browser run passed 94 of 94 cases. The full portfolio run passed 439 of 442 cases with four workers. Three timing-sensitive checks failed under that load: a DESK capability answer had not finished streaming, a same-page spotlight had not finished scrolling, and the plan-detail route hit the server's ten-second request timeout. All three passed when rerun serially on a fresh test server. The Builder and CRUMB cases passed in both runs. This local result records the parallel-load limit. Hosted CI passed all 442 browser cases in 5.6 minutes, along with type checking, unit tests, and the lint-baseline check. The Pages build, export verification, and deployment also passed for commit cbfba34.