How this stays honest: a monthly job validates every workflow against the live registries, benchmarks the ones that run, and flags drift. This page is generated from that run — no stale claims. Badges: measured = clean wall-clock from a real run · estimate = derived from measured parts · perf pending = built & runs, not yet timed · blocked = a known dependency blocker. Raw numbers: creative benchmarks.csv · audio benchmarks.csv. Scriptwriting throughput from the sovereign-stack stress tests. Lip-sync / avatar: all three local engines now run as dedicated pinned services — Mouthpiece (MuseTalk, fast draft), Persona (LatentSync 1.6, production naturalism), Thespian (Hallo2, cinematic) — local output tops at social-grade; broadcast-grade uses a cloud path.