Skip to content

VARSITY v1.0.0: first public release

Choose a tag to compare

@StephenSook StephenSook released this 05 Jun 21:50

VARSITY is a real-time, screen-reader-native, IFAB-Laws-grounded AI that explains VAR and offside decisions to blind football fans, built on IBM Granite for the IBM SkillsBuild AI Builders Challenge (June 2026). This is the first public release: the deployed, CI-gated build.

Try it live

What it does

A VAR review fires. VARSITY computes the real offside geometry, retrieves the exact IFAB Law, has IBM Granite explain the why, has Granite Guardian check it stays grounded, and the fan's own screen reader speaks it, in their language, before the broadcast picture catches up.

Verifiable highlights (all run live)

  • Real StatsBomb 360 offside geometry: 5.69 m offside / onside / too-close, the margin computed not hardcoded.
  • IFAB-Laws RAG (Docling to FAISS, Granite embeddings), Hit@5 = 1.00.
  • Calibrated uncertainty: ECE 0.34 percent vs a 4.16 percent overconfident control, spoken on an IPCC-style scale, with a too-close call that withholds the number and defers to the official.
  • All-IBM: Granite reasoning, Granite Guardian, Granite embeddings, Granite Vision, and Docling, federated through Context Forge with an A2A narrator.
  • Fully on-device offline mode: Granite Nano in the browser via WebGPU, zero backend calls.
  • SHA-256-signed IFAB corpus that fails closed; the free-text oracle survives a 13/13 red-team.
  • WCAG 2.2 AA enforced in CI; a watsonx outage degrades to a Law-citing floor instead of crashing.

In this release

  • A knowledge map of the project: a Mermaid concept map in the README, a 199-concept auto-extracted graph, and a cinematic 3D architecture view, all under docs/knowledge-graph/.
  • Attached media: a Gource time-lapse of the full build history (1080p60, 197 commits over 4.7 days) and a 10-second B-roll cut, in the brand palette.

Built solo to a team-of-four standard. VARSITY complements audio description, it does not replace it.