Let me understand what's being asked here. The system prompt says: "You file one Markdown note. Output only the note. Do not call tools. Do not use XML. Do not describe how you will write it."

The conversation shows a series of "filed notes" for a Mind about whether universities grow or shrink due to AI. The final instruction is a consolidation/revision instruction: "Revise the notes as a set: keep every claim that still holds, drop repetition, sharpen wording. Output one Markdown note that is the current best statement of what is established. Start with a # title, then short paragraphs, then: What would change this: … Keep sources or Unverified: tags. Do not mention tools, sessions, or that you are revising."

So I need to write one comprehensive consolidated Markdown note representing the current best statement of what is established across all the filed notes. This is a synthesis task — the deliberation in the prior turns (Notes 12, 15, 16's thinking) has already compiled the full inventory of durable claims. I should draw on that compiled inventory.

Let me compile the durable claims carefully:

**Core answer:**
- "Both" — the system bifurcates and thins; region- and stratum-specific. Not uniform shrink or growth.
- Total enrollment drifts ~19M → ~14–15M by 2045 (Unverified)
- Degrees conferred roughly steady while institution count falls — "thins, doesn't shrink"
- AI is an accelerant/sorter of existing trends, not primary cause
- Two-front squeeze: fewer 18-year-olds AND weakened degree signal
- Barbell: system shrinks at the median while top decile grows; median campus roughly halves (aggressive claim)
- Cost-efficiency claim: AI cost efficiency doesn't shrink the sector, it reallocates capacity — total roughly flat while geography rotates south (attrition-timing mechanism)

**Demography (sourced):**
- US births fell ~17%, 4.32M (2007) → ~3.59M (2023) (CDC/NCHS)
- HS grads peak ~3.9M/yr ~2025 (WICHE "Knocking at the College Door"); birth dearth hits college age 2029–2037
- WICHE: ~13% decline in HS grads by 2041; NE/Midwest/Plains lose 10–20% college-age cohorts; TX/FL/TN/GA/Carolinas grow
- Net domestic migration to Sun Belt (Census Bureau)

**Credential value / signaling:**
- Screening vs insurance decomposition: screening dies first (AI-verified work samples cheaper than proxy-filtering); insurance (blame-shifting/liability cover) is sticky residue through 2040s
- Post-2035 degree demand tracks a sector's legal/liability regime before it tracks region
- Licensed/statute-gated fields hold (nursing, law, accounting, teaching, engineering, medicine); super-selectives hold
- Mid-tier unlicensed BA premiums compress 30–50% real by mid-2030s (Unverified)
- Fastest collapse in generalist white-collar fields (business, communications, generic liberal arts)
- By ~2035 median degree = positional luxury marker; degree reprices from entry ticket to exit credential
- "AI doesn't kill the degree; it kills the undifferentiated degree"
- Credential-value decay hits by major before by institution
- Pre-AI drift evidence: Burning Glass Institute "The Emerging Degree Reset" (~46% of middle-skill postings dropped degree requirements 2021–22); HBS "Dismissed by Degrees"
- Stanford Digital Economy Lab "Canaries in the Coal Mine" (2025) — early-career employment declines in AI-exposed occupations
- Total degree-seeking enrollment falls 15–25% below demographic baseline by 2045 (Unverified)
- Mid-market enrollment contracts 25–40% by ~2035 (Unverified)
- Micro-credentials are a middle-market illusion — complement at top, fail to stabilize bottom
- Two brakes: AI application floods push employers back to crude degree filters; elite alumni networks
- Elite degrees gain value BECAUSE verification makes mid-tier signals redundant
- Degree survives as insurance (sorting, licensure, networks), not instruction
- Demand migrates to "protected degree + AI skill stack"
- "Bachelor's required" listings fall 30–50% in AI-exposed occupations by late 2030s (Unverified)
- Statutory gatekeeping: residual degree value set by statute, not brand signal
- Binding constraint flips from demand to physical attestation capacity (labs, clinicals, proctored assessment)
- Tiered online-plus-residency models (Georgia Tech OMSCS, ASU) absorb demand and grow
- Licensure premium vs signaling premium: signaling dies, licensure survives
- Hiring decomposes by