Own the hardware.
Cloud for when you need it.

UVU's instinct is right: at $25 per seat per month, covering 48,669 students runs $14.6 million a year — $16.4 million with all 6,123 employees added (54,792 × $300) est — four to five percent of the university's entire Education & General budget. Nobody should sign that.

But the real choice is not "expensive cloud or nothing." Three facts, all verified this week, change the picture:

1. Tools that add no license cost already exist. UVU's own guidance pages list Copilot Chat as included for students and staff under the current Microsoft license, and Google's Gemini for Education base tier costs education institutions nothing. Whether Gemini is actually switched on for UVU's account is something UVU's Digital Transformation office must confirm. fact license inclusion · tenant scope
2. The biggest signed education deals run ~$19–30 per user per year. Cal State pays ~$19.20/user·yr across 675,000 people; Colorado ~$20; Maine ~$23 per billed FTE. Against those, the $25/month fear is off by 10–16× — though mid-size schools have been quoted far more, which is why week one requests UVU's own numbers. fact
3. Owning the floor is now cheap. The best free-to-download models score six points behind the best paid ones on the 0–100 capability score explained above, run on a $2,139 Mac mini, and improve at no new license cost — each upgrade still earns a qualification pass. Faculty/staff serving hardware is roughly $64k upfront (~$80k over three years with power and care); operating staff and adoption support are the recurring costs, and they're priced in the open throughout this guide. fact est

Downloaded models have become useful enough for daily campus work. Two years ago, running a useful model on a desktop was a hobby. Today the best free-to-download models — Kimi K3 and GLM‑5.3 — score 60 on the Artificial Analysis index against 66 for the best paid-only model, and the lag behind the strongest paid models has fallen from about twelve months to three or four. The strongest paid models still do some work better. fact The chart below is the whole hardware argument: the score of the best model that fits in a 64GB Mac mini, measured on the same index.

The hardware just reset. Apple announced new Mac minis (M6 / M5 Pro) and Mac Studios (M5 Max / M5 Ultra) on August 25, first arrivals September 22, with a 512GB flagship following in late October. Buying last-generation hardware right now would be the one unforced error — the previous models are no longer sold new, and the survey in §04 and Appendix O found only single refurbished units and reseller backorders; timing the validation to the first shipment wave costs nothing extra. fact

Utah is funding exactly this. The 2026 Legislature put $15M one-time into a shared higher-ed AI research data center (administered under the University of Utah's budget item), created a research fund that reserves 15–25% for institutions like UVU, and UVU's own research office lists a $5M "Utah AI Moonshot" and a $45M Higher Education Research Fund with full proposals due October 6. A new president took office August 10. The October 6 round is available only if UVU filed the required notice of intent by August 28; if it did not, the next round is the target. fact the dates · unknown whether the notice was filed

So cloud is not unaffordable. The honest problem is different:

Owning trades that recurring bill for operating work: hardware is mostly one-time money (the kind universities actually have), capacity grows as free model upgrades land, conversations stay on campus under UVU's own policies, and the machines double as teaching objects for the AI programs UVU already runs — but somebody has to run them, and that staffing cost appears beside every hardware number in this guide. Rent where elasticity matters. Own where usage is daily and data is sensitive. Decide with UVU's own quotes and telemetry, not anyone's brochure — including this one.

The serving unit is unglamorous and right-sized: the new Mac mini M5 Pro with 48 GB of shared memory, a 512 GB disk and 10-gigabit networking — $2,399 retail, $2,229 education — likely purchasable on Utah's existing cooperative contract, with Procurement confirming the route (§09). The plan prices the 10-gigabit machine because its own rack kit assumes 10-gigabit switching; the same mini without it is $2,299 retail / $2,139 education and ships three weeks sooner. Published tests on this exact chip class measured the fast-lane model family reading prompts at 2,100–3,700 tokens per second (a token is roughly a word piece) and writing replies at 97–105. The higher-quality 27B model runs slower per word, so the router sends each job to the lane built for it (§11). Every number here gets re-proven on the delivered minis before any purchase beyond the four-unit validation. fact prices & published measurements · est service planning

The newest lineup beside the generation it replaced

A fair question from the first read-through: the plan is priced on Apple's newest machines, so what did the generation before them cost, what does it cost today, and how much faster should the new one really be? Here is the answer, priced both ways, with the grades. The short version: the previous machines are no longer sold new; single refurbished units exist at prices near or above what they cost new; and the new ones should carry about 10% more conversations per box (40% more for the Ultra) — credits we set well below Apple's headline claims. fact prices, availability, published measurements · est speed credits and like-for-like deltas

What it means. At education pricing the newest fleet costs about 14% more than refurbished previous-generation units and should carry about 10% more — roughly 4% more per conversation carried — and it is the only one UVU can order in quantity, with a warranty, on contract. That is why the plan is priced on the newest lineup, and why we call buying the old generation the one unforced error. One caution for Facilities: Apple's maximum power ratings rose (mini 140 → 155 W; Studio 145 or 270 → 480 W). Those are electrical ceilings for circuit planning, not expected draw — the one measured AI load on the previous mini was 46 W. The budget simulator now has a generation switch and the configurator prices every configuration both ways, so no comparison is hidden. est

What each machine can actually run — the matrix, not one number

A fair challenge from the read-through: why not put the most intelligent open models on 512GB Studios, and did we test every model against every machine, capacity included, with two models on one box? We had not. Appendix Q now does it — ten open models × six machines, each cell with memory fit, speed, conversations at once, cost per conversation, and an intelligence-per-dollar index. The short version is below; scores are the same 0–100 index explained in §02. fact measured or published · est derived, basis in Appendix Q

Want to try a different mix of machines? The hardware configurator starts from the computers instead of the dollars: pick how many of each, and it shows people served on an ordinary day and in exam week, the model sizes it can run, cost to buy and to keep for three years, staff, and power — and prices that same configuration on the previous generation beside the newest.

Reality check kept in view: peak-hour demand for a 5,000-person population models at ~7 simultaneous conversations (base case); for the whole campus, ~65–70. Demand, not headcount, sizes the fleet — which is why the pilot's telemetry matters more than any projection in this guide. est

There is no single right budget. There is a curve: every dollar buys coverage in a particular school, and every cut has a name. We built the curve into an instrument. Type any budget — or set each school's own budget, since money inside a school stays inside the school — and it solves the per-school supply-and-demand equations live: who is covered, where the first bottleneck forms, what falls below the line, and what the same coverage would cost to rent forever.

Prefer paper? Budget scenarios at a glance freezes the simulator onto one printable sheet: the coverage curve and four worked budgets.

These are estimates of UVU demand, not UVU measurements: no UVU usage data exists yet, and the pilot's own figures will replace every number here. They are anchored to the best published whole-campus results in the country — a full-campus service at Cedarville (82% of eligible people activated, 75% used it in a given week, 49 messages per active person per week), Virginia Tech's controlled pilot (78% weekly, median 11 messages a week), and 2026 national surveys (32% of students and 25% of faculty now use AI daily). The busiest hour of the day carries about 8% of that day's traffic. fact the other campuses' figures · est everything applied to UVU

The planning insight: for the phase-one population, base-case peak load is about 53 simultaneous conversations — roughly 65–70 if concurrent-enrollment students join later. The financial risk is not under-buying — it is over-buying before measuring. est

The growth story, from the federal database — with a twist that matters. UVU's headline growth is real (32,670 → 48,669 since 2010; Utah's largest and fastest-growing big institution). But the government data shows 76–87% of the last decade's net growth came from concurrent-enrollment high-schoolers; the phase-one population — degree-seeking students plus staff — grew only ~0.6–1.3% a year. So year-three capacity planning is driven by adoption deepening, not headcount: the sourced band is 1.3–2.4× year-one demand, base case 1.9×. Utah is also more resilient than the national "enrollment cliff" (high-school graduates projected −5.6% by 2041 vs −10.3% nationally). Design integration and facilities for the 2.4× envelope; buy compute on the measured triggers. fact IPEDS/USHE/WICHE · est band

We stress-tested this model until it broke, and it broke in instructive places. The base rate is anchored to measured whole-campus telemetry, not surveys: Cedarville's full-year deployment works out to 2.15 requests per eligible person per day — 1.43 peak streams per 1,000 people against our 1.44 — and Berkeley's course bot and Michigan's token volumes land inside the same envelope. Two cases demand far more capacity: a 60-second reasoning reply doubles the streams needed (response time is the single biggest lever), and one 300-student class starting an assignment at 2pm needs ~143 slots to keep waits near a minute. So the design shows the queue, spills only public-class overflow to approved cloud routes, and carries numeric reorder triggers: buy six more minis when the weekly 95th-percentile load — the level exceeded only 5% of the time — holds above 70% of capacity for two weeks; open the overflow valve when any 15-minute wait passes 30 seconds. Automated/agent traffic gets its own queue class entirely — the one European campus that published full telemetry saw 99% of volume come from scripts, not chats. est model · fact anchors

One more honest correction, from the engineering review (Appendix R): the model's fixed 30-second service time is right for chat and wrong for a finals-week mix of chat, coding, documents, and research writing — the weighted service time is about 51 seconds, so the same arrivals occupy about 93 conversations instead of 54, a 70% understatement. That sits inside the 9× stress envelope the fleet is already sized for, so the numbers stand; the reason is now stated plainly. What fits inside no envelope is coding agents admitted freely: one agent turn uses about 580,000 prompt tokens, roughly 1,200 chats' worth, and if 5% of users started agents in the peak hour a 26-mini fleet would run at 133% of capacity with 16-minute waits. Agents therefore run in their own queue with a metered allowance and overflow to the frontier pool before utilization reaches 85%. est · fact agent telemetry (GitHub production trace, 3.2 million users)

Machines are the easy half. The national record is brutal about the hard half: Cal State bought 500,000 seats and had activated at least 250,000 by spring 2026; voluntary training completion was 0.7% of students and 16% of faculty — and activation is not the same as use. UK government trials left ~20% of Copilot licenses untouched. Seats do not create usage. fact

What does? The programs that worked share five moves, and the evidence is consistent:

• Pay for a finished artifact, not attendance. CSU Bakersfield's $500-on-completion program hit 86% completion with a waitlist. Hawai'i paid 50 faculty $1,000 each to redesign an assignment. Small money, real deliverables. fact
• Gate access with a short training, then support weekly. Virginia Tech's trained pilot ran 78% weekly-active. Manchester reported 90% adoption in 30 days with cohort support. fact
• Teach inside the discipline. Faculty ignore generic prompt classes; they show up for "AI for your anatomy course." (Ithaka S+R, 19 institutions.) fact
• Design adjunct-first. In fall 2025 UVU counted 1,310 part-time and 802 full-time instructional faculty fact. That is 62% part-time est. It counts people, not full-time jobs. Any program built around salaried faculty misses most of the people who teach. Asynchronous modules, evening clinics, paid completion. UVU Common Data Set 2025–26, §I-1 (fall 2025); the narrower count is in Appendix J.
• Address integrity fears head-on. In every faculty survey we found, academic dishonesty outranks job loss as the top concern by a wide margin — but no single survey pins one share, so this guide no longer quotes one. The integrity evidence itself is calmer than the fear: no clean post-2022 jump in proven cases where institutions published totals, and AI detectors too unstable and uneven to prove anything (Appendix V). Lead with assessment redesign and authorship policy, never with "AI is inevitable." fact survey ranking · unknown any single share

Budget shape: adoption money is its own budget, never carved from hardware numbers. A one-college test costs $20–40k (30 faculty × $500 + two paid department leads); a cross-campus program runs $150–250k. At scale budgets the simulator holds an adoption line (default 18%) — because the evidence says the machines fail without it. est

Every component is open, inspectable, and swappable; the stack fits UVU's existing systems like it was designed for them — because it was:

Built in after the engineering, safety, and law reviews

Logging, precisely: chat history lives in the user's own workspace on UVU-controlled systems (that's what makes conversations resumable), with user deletion and a published retention rule — stated exactly: "Chat content logging is off by default. UVU retains only the data listed in its records notice, except when a user saves a conversation or a scoped legal or records hold requires preservation." A saved chat tied to a student is a FERPA education record; retained chats and metadata are records under Utah's open-records law — classified and, where required, produced, never promised "confidential" (Appendix T). Operational analytics see metadata only — tokens, latency, model, department — never prompt text; capacity planning and department accounting run entirely on those counters. Security-event logs required by Policy 447 capture system events, not conversation content, and the forensic procedure for a suspected leak is a written runbook with CISO sign-off. The full data-flow diagram, retention table, and runbooks are the first deliverables of the validation phase — drafted by us, owned and approved by Dx before any pilot user logs in. The control plane (the small databases and router state behind the curtain) is named, priced, backed up, and owned by the service's operator — with Dx holding approval, audit, and takeover rights — in that same package; no invisible single points of failure, and no borrowed Dx staff time. est

Stands on its own; fits the house: this service is designed to exist independently — its own hardware, front door, model operations, staffing, and budget — because everything UVU already runs (the employee AI Gateway launched in July, Ask Wilson in 44 courses, the shared state compute) is spoken for by the people using it today. Nothing here borrows their capacity. What it needs from UVU is small and specific: a room with power and a network drop, sign-on federation with the campus Entra tenant, a data-class ruling, and a sponsor. Three harness tiers ride the same router: a chat front door for everyone (LibreChat), a sandboxed coding agent for builders (OpenCode or Cline class), and a governed research search for scholars (Onyx class) — compared product by product in Appendix L, with the whole market of 171 harnesses — adoption, growth, funding, stability, and the dead-or-declining list — in Appendix M. The surfaces it plugs into are already mapped from public sources — identity, Canvas, portal, ticketing, security tooling, network — and the questions only UVU can answer are down to five (Appendix K). Where UVU wants the two worlds to meet, the connection points are ready — the fleet can serve as a local engine behind the Gateway, and course assistants like Wilson can call it — as optional integrations, never as dependencies. Dual sponsorship still fits the governance pattern UVU's documents describe: the college owns the academic mission; Dx owns identity, security review, and the Policy 445/447 package (data flows, classifications, runbooks), which arrives drafted for Dx and the CISO to own, change, and approve before any pilot user logs in. fact Gateway/Wilson · est arrangement

• No new bid expected. Apple minis and Studios are on NASPO ValuePoint contract MA 23003 / Utah addendum PA4282 — the competition requirement is already satisfied through June 2027; UVU Procurement confirms the SKUs and process (their call, made easy). fact
• Utah's trial-use statute (§63G‑6a‑802.3) was built for this: minimum quantity, defined test measures, no obligation to scale. Last-published signature levels: dean/executive at $25k+, VP at $50k+ — a $12k validation sits comfortably low (current schedule to be confirmed; the live PDF is down). fact verify current schedule
• Week-one, in parallel, at zero cost: request binding seat quotes at four scopes (500 / 2,112 instructors / 6,123 employees / all-campus) so the cloud comparison becomes like-for-like; apply for a Redtail early-access allocation (Utah's new 264-GPU statewide AI system — UVU faculty are eligible today) for research workloads; ask USHE whether a system-wide seat deal is in motion (Burns sits on the statewide AI task force); and stand up the governance charter (§07) with Faculty Senate at the table. fact eligibility · est plan
• Reviews to plan for, not fear: academic tech committee (ATSC), IT Oversight, security (Policies 445/447), and accessibility as a tested release gate — the web accessibility standard (WCAG 2.1 AA) with disabled-user testing, a funded remediation owner, and an equally effective alternative for every required task (Policy 452). UVU's campus AI policy (draft 441) is still in the approval pipeline; the pilot charter follows the enacted policies and adopts 441 when it lands. Timeline estimate for a funded on-premise pilot with no student-record data: 8–16 weeks to users, owned by those reviewers, not promised by us. Campus-wide: 6–18 months, phased. est
• Two gates added by our own gap-hunt: a one-page Facilities + Network + CISO signoff (named room, measured power/cooling headroom — 26 minis draw ~4kW and ~1.15 cooling tons — and the 10GbE build-to-order option specced, +$100/box) before any purchase order; and counsel-approved interim service terms (acceptable use, a "report this response" route, an incident taxonomy that carries UVU's existing 24-hour reporting duties into the service, and a narrow rule that the assistant never makes official decisions — it explains, cites, and hands off) before the first user. est
• One scheduling courtesy: Dx cuts the campus portal over to Pathify on November 16 — the pilot goes live before or after that window, not during it. fact
• Three rules that arrived in 2024–2026 (Appendix T): Utah's AI disclosure law — the service labels itself "UVU AI assistant — not a human" before and throughout every chat, the statute's safe harbor, and keeps counseling, clinical, legal, and financial advice out of the general assistant; since May 6, 2026 a state-funded university may ask Utah's Office of AI Policy for a joint interpretation of the one open question (is a free campus assistant a "consumer transaction"?). The federal ADA Title II web rule — WCAG 2.1 AA for the real chat, uploads, streaming answers, and generated documents by April 26, 2027, tested with assistive technology, not a page scanner. Records — a records map before launch: which chats are FERPA education records, which retained data are open-records material, which retention series applies (advising records: five years after separation), and how a legal hold stops deletion. fact law and dates · est design
• Ride the state's programs; count none as funding yet: the Utah Board of Higher Education's statewide AI direction (December 2025) and its AI Task Force (May 2026 — Dr. Burns is a named member); the free statewide AI Workforce Credential (available since July 1, 2026) goes into faculty and student onboarding instead of a competing certificate; the 2026 budget's $15 million shared AI research data center for Utah's public universities and $3 million ongoing AI workforce accelerator — UVU's share is unknown until an allocation exists; the NSF State and Regional AI Infrastructure Hubs solicitation (deadline November 4, 2026; $4–12 million over five years; one award per state) as a Utah consortium, not a solo request. fact programs and dates · unknown UVU's share
• Who pays, and who runs it (Appendix U): $0 from general student fees — Utah's fee rule bars them for instruction, academic support, and administration; central IT funds the shared base (fleet, identity, security, monitoring, refresh reserve, baseline research pool, the accountable owner); colleges and grants buy marginal nodes and reserved capacity; showback first, chargeback only for reserved or unusually heavy use; a college joins only after a signed three-year commitment covers its hardware and half its marginal staff time. Staffing: a professional owner with a paid student operations team beneath — more coverage at the smallest band, about a quarter off staffing cost at the larger bands (the simulator's staffing switch shows both). Buy the first fleet outright unless a written education-lease quote beats ownership; plan a year-three review and a year-four refresh, with three-year resale values near half of purchase price. est · fact fee rule, wage anchors, financing terms
• Buy in gated waves, never 26 at once (Appendix Z): four minis after the delivered-unit test (load the model; 1k/8k/32k-token tests; four streams at the 10-token-per-second floor; batch 1/2/4/8, cold-load, peak-memory, and 24-hour soak; identity, logging, isolation, accessibility, and recovery checks); up to eight more only after half the pilot returns at their next natural task within 30 days, measured demand uses at least 60% of capacity for four teaching weeks, and support stays inside the staffed plan; up to the planned 26 after spring results and a refreshed local-versus-cloud comparison. Waiting for a rumored next chip means about 15 months — 60 validation device-months and 390 fleet device-months forgone — while a machine 25% faster would repay no more than 9.6 months of waiting inside a 48-month life; so delay only for a failed must-have, an official Apple announcement within 90 days that clears a stated bar (the required memory, or 35% more measured throughput for no more than 20% more cost), a dated grant award within 60 days, or a blocking accessibility or security result. The 512 GB line stays unpriced until Apple posts it and two approved research workloads need it. A dated 24-month watchlist — Apple events, model checkpoints, grant deadlines, state milestones, the April 26, 2027 accessibility deadline — closes the appendix. fact cadence and arithmetic · est gates
• The full engineering design lives in its own document: the architecture dossier — component contracts and a substitution ledger, the data-class routing with per-contingency behavior, the model-provenance and release-integrity rules, 24 pre-decided contingencies ("if the CISO rejects a component… if quotes come back cheap… if one department eats 80% of usage…"), five ready decision memos, and the eight weekly tripwires with owners. est
• Money paths beyond the college budget: the $5M Utah AI Moonshot and $45M HERFP — full proposals due Oct 6, but only for teams whose notice-of-intent was filed by Aug 28, so the first question for Burns is whether UVU filed one (if not, target the next cycle); the state's $15M shared AI-compute appropriation; NSF NAIRR compute; the Kahlert relationship for a named lab; and the free research programs in §11. fact

Said plainly: in every search we ran, no public example surfaced of a university running campus-wide AI on Apple hardware — the schools that self-host (Münster, UCSD, Florida, Indiana, Purdue) all run Linux/NVIDIA on central research computing. UVU would likely be first, which is exactly why the plan buys four boxes before it buys twenty-six, measures everything, keeps the hosted lane running beside it for honest comparison, and publishes exit thresholds up front. If the validation says the Macs lose, the same gateway, sign-on, governance, and adoption work carries straight onto whichever backend wins — nothing is wasted but the price of four excellent computers UVU keeps anyway. fact search result · est framing

For the mathematicians and physicists — Burns's hardest requirement — the frontier pool routes heavy research to the strongest current models: Claude and GPT‑5.6-class models now solve International Math Olympiad sets essentially perfectly, score ~94% on graduate-level science exams, and produced at least one passing referee grade on 7 of 10 unpublished research problems in a controlled trial this summer. Metered access for 200 heavy researchers models at ~$22k/year — and the free lanes come first: Anthropic's scientist program (free seats + up to $50k API credits per project), OpenAI's researcher seats, NSF NAIRR compute. fact scores & programs · est pool cost

Any hardware review will ask "why not NVIDIA?" Here is the honest comparison with NVIDIA's own campus-friendly box, the DGX Spark, using verified prices and published measurements:

Where Spark genuinely wins — and earns a place in this plan: it runs CUDA, the language of every datacenter GPU a UVU graduate will ever touch. For the ML courses, fine-tuning labs, and NVIDIA-certification work the Smith College already teaches (UVU has an active NVIDIA partnership), one or two Sparks as named lab machines are the right purchase — for teaching, not for serving. For serving inside a Jamf-managed campus, the published evidence favors the Macs on management fit, power, and single-user speed, while NVIDIA/Linux keeps the more mature high-concurrency serving stack — which is precisely what the validation phase measures head-to-head before any fleet money moves. fact capabilities · est verdict

Every option on one page — where each honestly wins

How this was made. Twenty-seven independent research passes ran against the live web across September 2–3, 2026 — model leaderboards, Apple's store, signed contracts, Utah statutes, UVU's own policy manual, published campus telemetry — plus direct pulls from the federal IPEDS database (enrollment trajectories for UVU and eight Utah peers; UVU degree production by field). Every research thread is logged as its own tracked issue. Claims carry grades: fact verified at a primary source that day; est derived, with the math shown; unknown honestly unresolved. Then the work attacked itself, twice: an adversarial pass re-verified the eighteen most load-bearing numbers (ten confirmed, eight sharpened, zero broken); a paid-skeptic pass built the best case against this plan (its strongest points now live inside §09 and §12); three independent adversarial reviews — a blocking CIO, a skeptical CFO, a wary Faculty Senate — attacked the draft; a second audit round then graded the fixes and re-checked every number, catching an engine bug and several inconsistencies, which were corrected before publication. A fresh-eyes review of the published site on September 3 found more, and those corrections are in this version; the calculators and the written figures were reconciled against one another as part of it. The demand model was stress-tested to its breaking points (§06), and the proposed stack was physically benchmarked on our own hardware, isolation battery included (§04). Known remaining gaps are stated where they live, not hidden. Failed searches are logged, not papered over. This is how we believe AI-era consulting should be done — and the method is itself a demonstration of what UVU's own people can build with these tools.

Key sources

Live leaderboards: Artificial Analysis, LMArena, Epoch AI lag analyses. Hardware: Apple newsroom (Aug 25, 2026), Apple education store configurations, NVIDIA DGX Spark documentation, llama.cpp/oMLX published benchmarks. Contracts & pricing: CSU–OpenAI signed agreement, CSU renewal reporting, U. Maine System, U. Colorado, Microsoft education pricing, Google education pricing. UVU & Utah: UVU institutional data, Kahlert Applied AI Institute, UVU Policies 445/447/452, NASPO Apple contract, Utah Code §63G‑6a, UVU HERFP / AI Moonshot notice. Adoption evidence: Virginia Tech pilot report, Cedarville telemetry, Tyton 2026, Ithaka S+R, CSU rollout reporting. Precedents: Münster uniGPT, UCSD TritonGPT, U‑M ITS reports. Teaching, trust, accessibility, and timing reviews (Appendices V–Z): randomized trials and meta-analyses on AI tutoring 2023–2026 (Harvard, Turkey, Nigeria, Khan Academy evaluations), the Challenge Success and ICAI integrity data, independent AI-detector tests, the AI Assessment Scale, W3C WCAG 2.1 and DOJ guidance, LibreChat's public issue tracker, Apple's release history and newsroom, the Artificial Analysis index history, Wikimedia Commons image licenses. Engineering, safety, law, and money reviews (Appendices Q–U): Artificial Analysis model pages and methodology, the oMLX and llama.cpp benchmark databases, MLX-LM quantization benchmarks, GitHub's production coding-agent telemetry (arXiv 2608.00101), Utah Code Titles 13, 53E, 53H, 63G and 78B, the DOJ ADA Title II rule and its 2026 interim rule, Department of Education FERPA/PPRA guidance, HHS HIPAA guidance, USHE and Governor's Office announcements, Apple Financial Services terms, USHE R516, the NSF 26-513 solicitation, EIA electricity data. The complete source log — several hundred dated citations with the searches that failed as well as succeeded — is available on request.