Implementation plan · revised 21 Aug 2026

The plan behind
tier-bench.

This page summarizes the decisions implemented in the current build. The complete implementation-ready plan lives in the repository.

Catalog

Nineteen initial product lines, no fixed model cap. OpenRouter is polled every ten minutes; exact regex rules select canonical releases. Admin can enable, disable, add, or remove entries.

Replacement

A newer matched release publishes automatically. The predecessor receives a 72-hour visibility deadline; no cleanup job is needed for correctness.

Experience

Six official boards: Overall, Chatting, Math, Code quality, Steerability, and Most value. No Performance-only board and no visible live/freshness decoration.

Privacy

Working benches are private by default. Share opens a choice between an immutable link and a generated picture.

Model pages

Sentiment distribution first, category placements second, anonymous-by-default comments third, factual metadata last.

Stack

Next.js on Vercel Hobby, Clerk, Neon Postgres, OpenRouter catalog, Turnstile, Resend, GitHub Actions, and private encrypted R2 backups.

Delivery state

The working vertical slice includes catalog sync, ranking, private revisions, sharing, comments, proposals, admin controls, email outbox, read-only API, local PGlite fallback, Neon production path, and scheduled workflows. Production launch still requires vendor keys, a Neon migration, Clerk configuration, restore testing, and a real beta security review.

Inspect the API