Session 6b — The headline and the interpretability thread

• Date: 2026-09-20 · Branch: s6b-positioning · PR: #14

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

Goal

An editorial pass, not a feature: emphasis, ordering and vocabulary over facts already in content/. One thread runs 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. The session names that thread on Home, in About and on /research, gives the 2024 literature review on interpretability in generative AI a deep page in place of a one-line card, and moves the thesis to second of three case studies. It changes no fact and adds no claim.

What shipped

• Home (content/profile.ts): Paige's positioning line; a "Currently" line that carries what the positioning line does not (the MCP server under access-control filters, the agent loop with citation guardrails, the scheduled audit job that opens its own remediation pull requests) and keeps the 220+ metric; a skills strip that leads with evaluation and grounding, with python off it and grounding on it.
• About (content/about.ts): the second paragraph now opens on "making a system's claims checkable instead of merely fluent", says what the evaluation work targets (hallucination and groundedness, as a check on a product's output), and carries the agent-oversight sentence (wording cleared by Paige, 2026-09-20). The third paragraph ties the anomaly-detection work and the GMLE certification to the same thread in plain terms. A present-tense fourth paragraph, added after the first review at Paige's request, is on oversight of AI agents: Paige chose its content from options the agent offered, and the agent drafted the wording for her approval in the PR. To keep five paragraphs, the Egypt passage is condensed into one sentence of the closing paragraph, which already names the valedictorian address it comes from.
• Skill labels (content/skills.ts): xai is "Interpretability (XAI)" and grounding is "Grounding & claim verification". Label text only.
• /research: three themed groups in DESIGN §2's order — Interpretability and evaluation, Medical and neural AI, Applied LLM systems — then Reviewing. The theme is a field on each research item; the headings are in lib/content/labels.ts and the order in RESEARCH_THEME_IDS. Newest first inside each group; all seven items are still listed, each once.
• The literature review's summary (content/research.ts): rewritten from the review, with its reference count, 14, declared as a metric on the research item and sourced to the review ("Literature review, 21 April 2024").
• /projects/genai-literature-review: a fourth deep page, written from the review itself: why it was written, the failure modes it covers, the distinction it draws between interpretability and explainability with a note that other literatures use the two words the other way round, all eleven proposals it describes reported as their authors' work, where the field fell short, and its conclusion. No figure, and no numeral in the body except the date; the count renders through MetricStat.
• Featured order: StimMap3D, thesis, capstone. Three cards, one row, on Home and /projects.
• Thesis page: the tagline asks the method question, and one closing paragraph says what the approach carries and what it does not. No measured claim changed.
• Metadata: the root, About and /research descriptions name the subject. /projects' follows the featured order by construction, and the Open Graph card draws the new positioning line.
• Tests: see Verification.

Decisions and why

• /research finds a deep page through the pointer that already exists. The literature review's card points at its research item with researchId. Adding a projectSlug on the item as well would have typed the pairing twice, and the explorer's pair count caught it (below). deepPageForResearch() in lib/content/load.ts resolves either pointer.
• A review is not called a case study. /research links the thesis and the capstone with "Read the case study" and the literature review with "Read the summary" (deepPageLinkLabel in labels.ts). The page is a survey of other people's work and says so in its second paragraph.
• The deep page lists every proposal the review describes, in the review's order. Listing only the ones closest to the site's thread would misrepresent the source.
• The new page points at the review's sources rather than restating them. It refers to the review's reproduced chart as "a chart in one of the surveys it cites" and links the PDF, where the reference list is.
• About stays at five paragraphs. AboutSchema caps the narrative at five, per DESIGN §2's "3–5 short paragraphs", so the new sentences went into the second and third, and the present-tense paragraph took the Egypt passage's place (Paige's choice over raising the cap).
• About no longer carries StimMap3D's disclaimer callout (Paige, 2026-09-21, after the first review). Session 2's code review had added it by reading gate (c) as "wherever StimMap3D appears", and About names the project in passing. DESIGN's wording is "shown or embedded"; the gate now says so in AGENTS.md, DESIGN.md and the plan, and lists where the disclaimer stays: the card, the deep page, the embed, the timeline entry, the Open Graph card and the resume.
• Both relabels, not only xai. The strip now shows grounding second on Home, and "claim verification" is what the role's first bullet describes. Both leave the explorer layout untouched.

Tools and model

Claude Code with Claude Fable 5.1 (claude-fable-5-1). Text extraction from PDFs with pdftotext; Playwright for the e2e suites and screenshots; a separate review agent (feature-dev:code-reviewer) for the review pass.

The follow-up after the first PR review (the literature-review sentence fix, the doc status updates and About's present-tense paragraph) used Claude Code with Claude Opus 5 (claude-opus-5), with web search.

Constraints, human decisions and implementation direction

Decided by Paige on 2026-09-20 and implemented as given: the positioning line, word for word; the featured order and the count of three; that the resume PDF may change; the wording for the multi-agent work (guardrails, permissioning and oversight of what agents were allowed to do) and for the evaluation work (hallucination and groundedness); and that the session ships without About's present-tense paragraph unless she supplies it. After the first review Paige asked for it to be folded in, with options to choose from; she chose its content from the options the agent offered, and chose to condense the Egypt passage. The agent drafted the wording from those choices, for her approval in the PR.

Untouched, as instructed: the resume bullets in content/experience.ts, profile.summary, every entry's skill set, the staged source folder, StimMap3D's disclaimer.

Generated work and rejected suggestions

Generated by the agent: every wording change above except the positioning line; the deep page's body; the schema field, labels, loader helper and page grouping; the tests.

Beyond the task list (the plan's latitude clause). Each is sourced from content/ or the review, touches nothing the clause protects, and is checked in this repo:

Change	Check
/research links the literature review to its new page, labelled "Read the summary"	ResearchPage.test.tsx
The new page's facts column links the review PDF (the link already shipped on /research)	ProjectPage.test.tsx; the loader's document check
The new page ends by linking the thesis page	renders in the body test; the claim restates the thesis page's own "the tool must itself be tested"
A unit test caps the positioning line at 280 characters	content.test.ts
A unit test checks content text against two word lists	content.test.ts
A unit test pins the three themes' membership and order, and "14 papers", never "14 studies"	content.test.ts
SESSION=6b screenshots also save the two Open Graph cards	s6b-og-*.png

Rejected or left as recommendations:

• Adding eval-harness to the thesis so the constellation shows evaluation: a skill-set change, which this session may not make.
• Raising About's paragraph cap to six in advance: a DESIGN §2 decision, not an editorial one.
• Pinning the resume's literature-review line to its old wording with resume.bullets: Paige chose to let it change.
• A closing heading on the thesis page: the acceptance allows a tagline and one paragraph.

Verification (command, result, PR/commit)

Command	Result
npm run lint	clean
npm run typecheck	clean
npx tsx scripts/layout-explorer.ts --check	up to date (42 nodes, 173 edges), without regenerating; content/generated/explorer-layout.json is not in the diff
npm run resume	2 pages
npm test	17 files, 142 tests passed (136 before the session)
npm run build	clean; /projects/genai-literature-review and its Open Graph image prerender
npm run e2e:a11y	97 passed (one more route in every sweep)
npm run e2e:features	42 passed, 1 skipped (the opt-in live-embed test, as before)
SESSION=6b npm run e2e:screenshots	28 passed
node scripts/js-budget.mjs /	137.1 KB gzipped before and after (459.7 KB raw, 7 scripts)

The resume PDF changed in three places, not one. The plan expected the literature-review bullet to change. The skill labels also print in the resume's "Technical skills" lines, so "Interpretability (XAI)" and "Grounding & claim verification" replace the resume of record's "Explainable AI (XAI)" and "Grounding & Citation Verification" there. Still two pages; the revision pin on its assets.json record is updated in a follow-up commit.

Acceptance 5 was checked with git diff main -- content/projects/interpretable-medical-imaging-thesis.mdx: the tagline, order, and one added paragraph; nothing else.

Claims drawn from the literature review (shipped PDF, "Literature review, 21 April 2024"): the industries and the black-box framing, p. 1; governance and the two definitions, p. 2; the chart and the survey's finding on evaluation procedures, p. 2; the eleven proposals, pp. 3–8; "no method has been widely deployed", p. 8; the conclusion on standardized metrics, pp. 8–9; the reference list, fourteen entries, pp. 9–10.

Not done here: the Vercel preview check in an incognito window needs Paige's login, and CI runs on the PR.

What the AI got wrong and how it was caught

• Typed the research-to-page pairing twice. To link /research to the new page the agent first set projectSlug on the research item, although the card already points back with researchId. explorer.test.tsx counts one pair per pointer and failed (seven, expected six). Replaced with a loader helper that resolves either pointer.
• Missed two of three references to the old skill label. The first replacement changed the quoted label in explorer.spec.ts and left two regular expressions. Caught by grepping for the old label before running anything.
• Wrote three sentences the review does not support in the deep page's first draft: a motive ("I wanted to know…"), a cause ("come from how the models learn") and an inference ("two of its findings explain part of why"). Caught on a re-read against the extracted text; each was cut or rewritten as something the review says.
• The plan's task list was incomplete in three places, and the agent found them by running the suite rather than by reading. Two tests named the old xai label; three tests assumed the featured projects and the deep pages are the same set; and the skill labels reach the resume PDF.
• Left a module comment false. content/research.ts still said every summary uses the resume's wording after one was rewritten from the review. The agent found it while checking for stale comments, and the review pass reported the same line independently.
• "For a security team" was looser than the resume. The role was on a security team; About now says "on". Caught on a second read of the new sentences against the role's bullets.
• Attributed the review's definitions to its sources. The page said the interpretability and explainability distinction followed "the explainable-AI sources the review cites"; the paragraph that defines them (page 2 of the PDF) cites nothing. Caught by a second review of PR #14 against the extracted text; the sentence now says the review states the distinction without a citation.
• The review pass was narrower than asked. The review agent had no shell, so it read the working tree and the review PDF rather than the diff, and could run nothing. It checked the new page against the PDF sentence by sentence and reported one finding, the comment above. The commands in Verification were run by the implementing agent, not by the reviewer.
• A shell heredoc holding test code failed to parse, and a first edit script assumed Python was installed. Both were rewritten as Node scripts in a scratch folder; no partial edit was left.

Screenshots

docs/screenshots/s6b-*.png: every route at 1280 px and 390 px (Home in both themes), plus s6b-og-home.png (the positioning line inside the 1200×630 card) and s6b-og-projects-genai-literature-review.png.