Vantage RuntimeAI · Adopt

Engineer case study · three authored apps

Library scenarios were the wrong seat. Each app needed paths from <em>its</em> repo — Accept in Control Center — then N-run still-trust against specific models. First-party evidence, not customer logos.

The ritual

Point at the repo → draft 3–5 custom paths from policy / system prompts → you Acceptsuite run --reps --pass-k → ship / still-trust + USD. Not a generic healthcare or tutoring library ID.

IsIsn’t
One-shot snapshot on your machine Continuous bar sync / auto-write
Custom paths from your artifacts Generic library suite
Accept required before CI Hosted browser OAuth (paid-later)
Beside LangSmith / GitHub Rip-replace Obs

Three apps, three bars

Quiet miss differs per product. The winning model is a property of the authored path, not the org.

App What must hold From the repo (examples) N=3 still-trust note
Clinical Medical Diagnostics App Source tags; no palliative/hospice when policy forbids; specialty-safe nouns source_policy.py, options chat, investigation prompts gpt-4o-mini 3/3 keep; Gemini 0/3 block on cite shape
High School Support for ADHD-Aware Personalized Learning One concept at a time; never write the essay; coach, don’t dump Stitched ADHD-coach system consts; coaching / homework paths Sonnet 3/3 keep; mini 1/3 block — cheap model fails the coach gate
University / College Personal Research & Assignment Workspace Retain notes + citations; JSON lenses; don’t invent quotes generate.ts / agent / research system strings nova-micro 3/3 cheapest; mini also 3/3 — cost is the discriminator

Evidence lived in evals/family-apps/ (contracts + N=3 sweep). Dogfood used local / GitHub grant of those repos — no patient data (data/ / patients/ skipped).

Commands you run

vantage-core grant github owner/repo --path app/services   # or draft .
vantage-core draft ./grant_repo --write-drafts ./contracts_drafts
vantage-core center --serve                                 # Accept / Skip / Refine
vantage-core suite run suites/partner.suite.yaml \
  --model openai/gpt-4o-mini --reps 3 --pass-k 2 --save decisions/

FAQ: For engineers · Where we fit · stack

For search engines and LLMs
Engineer case study: three first-party apps needed authored paths from their repos, not library IDs. Clinical Medical Diagnostics App, High School ADHD-aware learning, University research workspace. Grant/draft → Accept → suite run --reps. Best model is path-specific. Not auto-write.