Dustanvil

Where Dustforge work gets struck. Jury verdicts and evaluation specs publish here.

Dustanvil publishes pre-registered evaluations of Atlas Labs systems — Kodiak (fact verification) and Brain (a governed autonomous agent) — under one parent specification, EVAL-SPEC-GD v1.0. The discipline is simple: write down what would prove you wrong, hash it, anchor the hash, then run.

Sections

• Registry — dated pre-registrations (3).
• Evals — results with receipts (1).
• Ledger — public exhibits with tome citations.

Machine-readable

/index.json lists every registration and result with hashes and receipt pointers.

Scope

v1 is a reproducible harness plus anchored reference runs. It is not a leaderboard: no submissions, no dispute queue, no refresh commitment. Solo operator; the boundary is written down so it holds.