Plan amendment — add Session 6b, and open S0-F

• Date: 2026-09-20 · Branch: s6b-spec-amendments · PR: #13
• Documentation only: DESIGN.md, IMPLEMENTATION_PLAN.md, SESSION_PROMPTS.md, README.md, .gitignore and this entry. No file under app/, components/, content/, lib/, public/, scripts/ or tests/ was touched.

(redacted for publication, 2026-10-01 and 2026-10-02)

Goal

Two things, both requested by Paige on 2026-09-20.

1. Define Session 6b so an agent can build it: an editorial pass that names the thread already running through the work — a system's output can be checked, the checker is itself tested, and the check runs automatically or it does not count — on Home, in About and on /research, gives the GenAI interpretability literature review a deep page in place of a one-line card, and moves the thesis up one slot in the featured row. It changes emphasis, ordering and vocabulary over facts that are already in content/; it introduces no new fact and no new claim.
2. Open S0-F, a prerequisite of Paige's own, covering what the published build logs may say.

Session 6b could not simply be handed to a coding agent as a note. AGENTS.md lets an agent amend DESIGN.md, IMPLEMENTATION_PLAN.md and SESSION_PROMPTS.md only at Paige's request, and every session prompt tells the agent to read its section of the plan and build only that. Until the plan had a Session 6b section, there was no session to build — and one of 6b's tasks (the featured reorder) is pinned by a unit test that mirrors DESIGN §0, so an agent would have had to change a spec it was not allowed to touch in order to make the suite pass. The amendments come first, on their own branch, as the Session 3 split did in September.

What shipped

DESIGN.md

• Status: Sessions 1–6 and 3b merged (3b as PR #12, dc0a8af); Sessions 6b, 7 and 8 remain.
• §0 "Sections": four deep pages, not three.
• §0 "Deep project pages": the GenAI literature review added as the fourth.
• §0 "Featured order", a new row: StimMap3D, Honours thesis, Adolescent Spinal Curvature capstone. Three, not four — both the Home "Selected work" row and the /projects "Case studies" row are md:grid-cols-3, so a fourth featured card wraps onto a second row by itself. The row also records that tests/unit/content.test.ts pins the featured order to it, which was previously an unwritten coupling between a test and a prose table.
• §0 "Editorial pass", a new row: what Session 6b changes and the bound it works inside.
• §1 answer 1: rewritten, with the old answer kept beside it so the change is legible. It also records that profile.positioning feeds three surfaces — the Home hero, the site-wide Open Graph card's description and the JSON-LD Person description — and therefore has a length ceiling of about 280 characters, because the card is a fixed 1200×630 image with no overflow handling.
• §2: a site-map row for /projects/genai-literature-review; the /research row amended for the three themes and for the rule that the theme is a field on the content item, not a list typed into the page; the thesis row amended for its new tagline and its one added paragraph.

IMPLEMENTATION_PLAN.md

• §3: S0-F added, and the "every prerequisite is now closed" line corrected — S0-F is open and is the only one outstanding.
• §4: Session 6b added between Session 6 and Session 7: prerequisite, the five decisions Paige settled on 2026-09-20, eleven tasks, a latitude clause, verification and ten acceptance items.
• §4 Session 7: prerequisite updated (3b merged; 6b now runs first), with a note that 6b adds one HTML route to every sweep, adds no client JavaScript, and leaves one label-fitting observation for Session 7's polish pass.
• §4 Session 8: task 2 amended — S0-F must be done first, and the "what worked / what did not" synthesis is written as a taxonomy of what the agent got wrong and which mechanism caught it, with an explicit accuracy bound on what that record is and is not.
• §5 and §7: timeline row and acceptance-summary row for 6b; the hours estimate updated.

SESSION_PROMPTS.md

• Top banner: the new run order (S1 → S2 → S3a → S4 → S5 → S6 → S3b → S6b → S7 → S8), 3b recorded as merged rather than in review, and S0-F announced.
• A paste-ready SESSION 6b prompt before Session 7's, in the pack's existing shape.

README.md — status paragraph: 3b merged, 6b added.

Decisions and why

• Session 6b is its own session, not a fold into 7 or 8. Session 7 is accessibility and performance gates and Session 8 is the build story and launch; both are structural. 6b is content, and it has to land before 7 so the new route is inside Session 7's axe, responsive and Lighthouse sweeps, and before 8 so the launch snapshot and the asset reconciliation include it.
• Three featured cards, not four. Both grids are three-column. A fourth featured card makes the first screen worse, not better, and a literature review is a survey rather than a built artifact, unlike the deployed application and the capstone beside it. Moving the thesis from third to second instead costs little churn: two order fields and one test array.
• The skill relabel is specified as short, and the length was measured rather than assumed. lib/explorer/layout.ts places each label from labelWidth(n.label), which is a function of the label's character count, and scripts/layout-explorer.ts --check compares the committed layout byte for byte. Measured on 2026-09-20: "Interpretability (XAI)" (22 characters) regenerates the file byte-identically, while a 39-character alternative moves the xai label and its hub label, and the hub label is drawn in the Home hero teaser. The plan and the prompt both specify the short form and ask the session to prove --check passed without regenerating.
• The session may go beyond its task list, inside a written bound. Paige asked that the session be allowed to add what it finds, so the plan carries a latitude clause: sourced from content/ or an already-cleared source, no touching the verbatim resume bullets, profile.summary, the staged source folder, the approved wording or StimMap3D's disclaimer, no change to the skill graph or the spec files, verifiable by a check that runs in this repo, and disclosed under its own heading in the PR body and the build log. Anything failing one of those is a recommendation, not a change.
• S0-F is Paige's step, not an agent's. Rendering the build logs publicly is already in Session 8, and gate (b) forbids a staged-source path in a served page, so a scrub was required. Paige set its scope; any change to a repository's history is hers to run, and an agent may only list candidates and draft the request.

Tools and model

Claude Code with Claude Opus 5 (1M context). No code generation; documentation edits plus four read-only measurements against the working tree (below).

Constraints, human decisions and implementation direction

Paige decided, on 2026-09-20: the positioning line; the featured order; that the resume PDF may change; that she writes the About present-tense paragraph and approves its exact wording, and that the session ships without it if it is not approved in time; the agent-oversight wording; that Session 6b's prompt grants latitude for additional verifiable changes; and S0-F.

Paige cleared the agent-oversight wording on 2026-09-20: the multi-agent infrastructure-automation work involved guardrails, permissioning and oversight of what agents were allowed to do, and the grounding and evaluation work targeted hallucination and groundedness failure modes.

Generated work and rejected suggestions

Rejected, with reasons now recorded in the plan:

• Featuring the literature review as a fourth case study. Rejected for the grid geometry and for over-promoting a survey; superseded by the reorder.
• Re-tagging content so the skills constellation foregrounds evaluation. Measured and rejected: adding one skill to one entry re-lays out all 42 nodes, invalidates the Session 5 screenshots and changes the Home teaser, and the label in question still does not fit, so the picture does not actually improve. Recorded instead as a label-fitting note for Session 7.
• Grouping /research by theme inside the page component. Rejected against gate (e) — a theme is a fact about the work — and because it cannot be derived from the existing kind field, since the interpretability thread spans three kinds. The plan specifies a content field plus display strings in the labels module.

Verification (command, result, PR/commit)

Measurements taken on the working tree before writing the amendments, each restored afterwards:

Check	Command	Result
Baseline	npm test	17 files, 136 tests passed, exit 0
Layout baseline	npx tsx scripts/layout-explorer.ts --check	up to date (42 nodes, 173 edges)
Long skill label	relabel xai to 39 characters, then --check	exit 1, layout out of date; the xai label and its hub label both move
Short skill label	relabel xai to "Interpretability (XAI)", then regenerate	file byte-identical to the committed one
Skill-set change	add one skill to one entry, then regenerate	all 42 node placements change; the label still does not fit

Added by the review pass on the same day, after the first four commits:

Check	Command	Result
Both relabels together	xai → "Interpretability (XAI)" and grounding → "Grounding & claim verification", then --check	exit 0, layout up to date
Negative control	xai → a 45-character label, then --check	exit 1, layout out of date — the check is label-sensitive
Which labels are drawn	read labelAt.fit in content/generated/explorer-layout.json	eight do not fit: agent-loops, openai-agents-sdk, gemini-vertex-ai, bm25f, eval-harness, ci-gated-validation, access-control, embedded. grounding fits
Reference count	extract the shipped review PDF's reference list with pdf-parse	14 numbered references, [1]–[14]
Build-log entries citing staged-source paths	grep -l for the staged-source folder prefix over docs/build-log/*.md	11 of the 13 entries present before this one
Suite, after the corrections	npm run lint; npm test	clean; 17 files, 136 tests passed

After the amendments: npm test re-run — 17 files, 136 tests passed. Nothing in this branch is imported by code, and the featured-order test still matches content/, which this branch does not change, so the suite is unaffected. git status shows no staged-source file.

What the AI got wrong and how it was caught

• The first-pass review of this material asserted that a skill label change needs no explorer regeneration, twice. Caught by reading lib/explorer/layout.ts and then testing it: the long label made --check exit 1. The plan now specifies the short label and asks the session to prove --check passed.
• It also said the positioning line renders on Home "and nowhere else". Caught by grepping for the field: it also feeds the Open Graph card and the JSON-LD Person. That is why DESIGN §1 now carries a character ceiling.
• It said a fourth featured card would cost "a grid-class change" and named grid classes that exist nowhere in the repository. Both grids are md:grid-cols-3. It also missed that two unit tests pin the current three slugs.
• It said the new deep page needs nothing registered. The route and the sitemap do derive automatically, but ProjectPage.test.tsx requires a body test per deep page and seo.spec.ts requires a site-unique description. Both are now tasks.
• A first attempt to write this branch's longest file through a shell heredoc failed (ENAMETOOLONG at spawn). Rewritten with the file tool; no partial file was left behind.
• A first attempt to edit a file with a shell-quoted inline script broke on an apostrophe. Rewritten as a script file with the text read from disk, after backing the file up.

Found by a review pass over this branch on 2026-09-20, and fixed in it:

• S0-F said all thirteen present entries cite staged-source paths. Caught by running the grep: eleven do; plan-reorder-2026-09-15.md and s0e-vercel.md do not. The scrub is still required, but the count was wrong in three places (S0-F, the Session 8 amendment and the PR body).
• Session 6b task 5 sent the 14 metric to the wrong file, saying task 6 (the MDX) declares it. mergeResearch in lib/content/projects.ts replaces a researchId card's metrics with the research item's, so a metric declared in MDX frontmatter is discarded and the claims gate would never see it. The task now names content/research.ts and the mechanism.
• The Session 7 note listed grounding among the constellation labels that never get drawn. Caught by reading labelAt.fit in the committed layout: grounding fits. The note now names the four of the eight non-fitting labels it meant, and lists the rest.
• Task 1 called the skillsStrip change a reorder. It also drops python and adds grounding; the task now says so, and says why the strip is display-only.
• Acceptance 9 was worded too broadly. Read literally it failed on employers, institutions and venues already on the site. Reworded to what it meant: the session adds no fact and no organisation that is not already in content/.
• The "measured" claim covered one label, not two. The plan said both label lengths had been measured; only xai was. grounding has now been measured too, separately and together with xai, and the table above records it.

Screenshots

None — documentation only.