xword

A crossword agent on the Nebius stack: Tavily looks up the clue, Jev ranks the shortlist, Token Factory writes only leftover wordplay, and Python owns the grid.

Jared Werba · source · jwerba.com

wall 0.0s
tavily 0
jev 0
tf tokens 0

puzzle

1 Engine

Slots, crossings, fill_slot. A wrong answer cannot corrupt the grid.

2 Tavily

Grounding. Search crossword clue "…" N letters. Blueprint recipe 03.

3 Jev

Ranks ≤12 candidates. Does not invent letters. Gateway evaluation, not chat.

4 Token Factory

Last resort for wordplay. Not used for routing or submit.

Eval vs the LangGraph baseline

Live race against nebius-xword.vercel.app (same ids, V4 Pro LangGraph): 3×3 12.1s / 3,116 TF tokens vs 6s / 0. Mini 5×5 15.2s / 12,764 vs 5.5–10s / 0. Trickier 5×5-b ~19s / 14.7k vs 7.9s / 0.

Harness: python -m eval.run_eval --solver empty|oracle|backtrack|search-jev. LangSmith project xword.

60-second demo

Record after the 5×5 is green. Four sentences: engine owns crossings; Tavily grounds; Jev ranks; Token Factory only writes leftovers.

Video slot — drop the take-home recording here