Measured against the same two benchmarks used for every prior backend
swap this project has made:
- Urgency noul: von 0.9995 (vs. laya's 0.77, kev's inverted 0.17) --
essentially matches Jev's own claimed ~0.999, and correctly separates
calm (0.059) from urgent.
- costslayer's L1 task-scope Choice classification (external stress test,
costslayer itself unchanged): von scored 0.737 average, actually
*behind* laya's 0.77-0.80, with a real 25% under-classification rate.
No backend wins both axes. Swapping anyway because Noul was this
project's clearest weak spot (kev inverted it, laya was only middling),
and because von ships its own /v1/systemone HTTP server -- removes the
need for a custom wrapper (serve_laya.py deleted; laya had no server of
its own). README's Known Limitations section now documents the Choice
weakness plainly rather than only reporting the win.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
zzgandClaude Sonnet 5
15704f4ada