Build the platform that will judge you: how we built DOGFOOD 2026

DOGFOOD 2026 in one line: build the platform that will judge you. Every team ships a submission and judging portal, then an automated acceptance checker grades their own portal against it.

188 teams submitted. That constraint killed demo-ware early: if peer-score reads do not 403, if closed events accept submissions, if CSV export breaks — you fail, live, in front of the judges.

T1 core: JWT httpOnly-cookie auth with five roles, events with server-enforced deadlines, invite-code teams, draft-to-submitted projects, and a public searchable gallery.

T2 judging: track assignments, weighted rubrics frozen at exactly 100%, backend-enforced judge isolation, conflict and self-review blocks, progress dashboard, z-score calibration with Bradley-Terry and Borda views plus a ROBUST/FRAGILE boundary verdict, audit trail and CSV export.

T3 voting and T4 stretch (API keys, HMAC webhooks, Ed25519 certificates, gallery widget, bulk import/export) round it out — see the product posts for those.

Claimed T1+T2, verified 7/7 PASS on run.py. Two reproducible demo events prove the awkward cases: Calibration Cup (4 judges × 30 projects) and Grand Jury (30 judges × 40 projects), both ending FRAGILE — raw #2 falls after calibration.

The full 3,500-word build log is in the repo at blog/build-the-platform-that-will-judge-you.md and was submitted to the Write-Up Quest tally form.