Public evals

Wren, graded by Wren.

Every week we re-run a small, scripted eval on Wren itself and publish the results — the numbers that look good and the one that doesn't. No third-party leaderboard, no cherry-picking. Just receipts we actually have.

This is a self-evaluation. These are product-quality checks Wren runs on itself through the same hosted model that powers the app — not an independent benchmark, and not a claim about frontier model research. Each item is graded automatically in your browser, so the numbers are reproducible. See how we run honest evals.

Latest run

No run yet — press “Run this week's eval”.

—

overall

Scheduled to re-run weekly via site cron. The harness above is the exact one the schedule triggers.

The harness

Four small suites, graded automatically

Deliberately small and deterministic — so anyone can re-run it and get the same numbers. Every claim carries a number; at least one number flatters nobody.

Run history

Every run, dated

Run date	Week	Instruction	Frames	Refusal	Precision	Overall
No runs recorded yet. The first run seeds this table.

Runs are stored and dated as they happen. A production Wren would keep a longer public ledger; this reference build starts its ledger the first time the harness runs. Read the monthly roll-up in The Wren Index.