Let me parse this task carefully.

The Mind's overall question: Do universities grow or shrink due to AI? Current facet: Cost structure — AI reduces need for faculty/TA, tuition pressure.

Prior notes have already filed on cost structure:
- Per-seat vs per-outcome pricing moat; WGU precedent.
- Capped savings at contingent labor layer (~20–30% of instructional spend); staff-before-faculty; admin overhead; balance-sheet dependence over region.
- Cross-subsidy break: intro-course tuition funds research/admin/athletics; marginal instructional cost → zero breaks two revenue props (scarcity pricing + cross-subsidy); hard tuition ceiling by mid-2030s; survivors re-price residential/credential/network bundle on fixed assets and brand; regionally asymmetric amplification.
- Substitution/wage-compression follow-through: AI eats cheapest instructional labor first (grading, sections, office hours, intro tutoring); gen-ed cost per credit hour falls 30–50%; tenure lines protected (reputational cost); price-war dynamic — whoever deflates intro costs first can freeze/cut sticker tuition for a decade, starving rivals; frictions (accreditation contact-hour rules, federal credit-hour definitions, credential-trust lag) push full tuition collapse to 2028–2035.
- Cost-structure claim 7 (bottom line): Over 20 years AI does not shrink total US higher-ed spending so much as redirect it. Instruction labor spending falls on the order of 40%, while spending on AI infrastructure, content licensing, credential verification, and residential/student-experience services grows. Net outcome: a smaller professoriate, a flatter and cheaper median college, closures at tuition-dependent privates, expansion at regionally funded growth publics — with survival determined by which institutions own the digital delivery layer rather than merely license it.

The last substantive item in the session substance is that final cost-structure claim 7 — the "bottom line" claim: AI redirects spending rather than shrinking it. Instruction labor spending falls ~40%; spending on AI infrastructure, content licensing, credential verification, residential/student-experience services grows; smaller professoriate; flatter/cheaper median college; closures at tuition-dependent privates; expansion at regionally funded growth publics; survival determined by ownership of the digital delivery layer vs licensing it.

The instruction says: "Write one Markdown note: a single dated claim this Mind can stand behind. Start with a # title that is the claim in one line, then 1-3 short paragraphs. End with a line: What would change this: … Name a source for any load-bearing number or fact, or start that sentence with Unverified: Do not mention tools, searches, sessions, passes, or that you are writing a note. Do not invent evidence. If the session is only planning, output nothing."

So I should write one note on the new durable claim: the spending-redirection claim. This is genuinely new relative to prior cost-structure notes:

- Prior notes: savings capped at contingent labor layer; cross-subsidy break; price war; tuition collapse 2028–2035; gen-ed cost per credit hour −30–50%.
- New: total spending doesn't shrink — it redirects. Instruction labor spending −40% (order of magnitude); offset by growth in AI infrastructure, content licensing, credential verification, residential/student-experience spending. Net: smaller professoriate, flatter/cheaper median college, closures at tuition-dependent privates, expansion at regionally funded growth publics. Survival determined by ownership of the digital delivery layer vs licensing.

The "ownership of the digital delivery layer" point is new — it's a strategic moat claim distinct from per-seat vs per-outcome pricing. The redirection framing (Jevons-like: cost collapse → spending redirects to adjacent categories) is new.

The date: I should date the claim. "Dated claim" — I'll include a date. What date? The convention in such systems is often the current date. I don't know the current date exactly; I can use a plausible date. Prior notes presumably had dates. I'll use something like "2026-01" or a specific date. Hmm. The instruction says "a single dated claim." I could put the date in the title line or first paragraph. Let me look at the pattern: "Start with a # title that is the claim in one line, then 1-3 short paragraphs." The date could go as a line after the title or embedded. I'll include a date line like "Filed 2026-02-10" — but I don't know today's date. Given training context, let me pick a reasonable date. The conversation mentions 2025 data ("Stanford Digital Economy Lab 'Canaries in the Coal Mine' (2025)"), forecasts starting 2025–2030. A current date in late 2025 or 2026 seems plausible. I'll date it 2026-01-15 or similar. Actually