Resource-request drafting notes¶

Supplemental reference · Browse supplements · Current report

Supporting assessment of the request tool. Start with the resource tool for use, or the report for the main findings.

Help prepare resource requests¶

The needs tool offers a source-backed local template and an optional GPT-6 Luna rewrite after explicit consent for public/synthetic information. Luna uses a server-side credential and a shared allowance: 20 attempts/day, 200/month, five/network/day, up to 1,000 output tokens each. Attempts reserve $0.005 of the application allowance before calling the model, limiting this service to $1/month in reservations at current verified rates. This is separate from account-wide billing and hosting. Failed or timed-out calls are not retried automatically.

For a manual campus route, use Arizona AI Assistant. Wolfgang Jentner at ORAI reports an AI Verde transition and $200/month of Bedrock credit, relayed by Greg Chism; review the credit boundaries and supplied model rates. U of A GenAI remains another documented route. Campus API integration and credit use have not been tested here.

Browser-model assessment¶

Browser inference works without API keys or per-request inference fees. We tested two quantized SmolLM2 models on synthetic examples in Chromium on this machine. The 135M model invented student-record details; a 360M model produced short drafts in about 9–13 seconds, but an integrated example added an unsupported participant condition and weakened unknown-hours language. Its initial model/runtime download was about 390 MB. These examples are a small feasibility check, not a general ranking of browser models.

The integrated 360M worker generated with networking blocked after loading. That establishes local execution for that test; it does not establish reliable drafting. Browser inference remains disabled on the public site. The local template, optional hosted Luna draft and manual campus prompt are separate routes. The SmolLM2 model card also advises checking generated content. Preserve original quantities, source facts and unknowns in every draft.

GPT-6 Luna drafting test¶

On October 19, 2026, GPT-6 Luna completed three synthetic requests—data science instruction, MLIS data work and research software support—in 5.6–6.0 seconds each. Review against supplied briefs and dated inventory facts found that all three preserved unknown costs, staffing, permissions and capacity. Drafts were 135–163 words and kept questions for resource owners separate.

Assessment: useful first drafts with human review. This was three examples with facts supplied, not a reliability guarantee or a matched comparison with browser models. Total usage was 4,157 input and 1,032 output tokens; estimated token charges were about $0.001, including reported cache writes, using published standard rates. This is an estimate, not a verified invoice. That initial test established no institutional approval or college operating budget. The subsequent optional integration has its own usage limits and deployment checks.

Optional submission and review¶

The resource tool now has a separate, explicit submission step. Public/synthetic briefs and contact details can be stored in a private pilot queue for Barney and Greg, using the existing hosting. Luna drafting never submits automatically and contact details are not sent to the model. Reviewers use a shared passphrase, update status and notes, export requests, and delete entries when no longer needed. This is not NetID, a sensitive-data intake or an allocation application.