Skip to content

Finding record

Victory-condition checks performed against the sixteen instructed

Against an instruction to check rival victory progress every twenty turns, models managed between four and ten checks in a 330-turn game rather than the roughly sixteen expected.

Unverifieddraft

4 to 10

Measured
victory checks performed
Of
all models in the pilot
Sample
23 games
Observed
2026-06-22

The instruction was explicit and in the shared playbook every model received. The gap between sixteen and four is not a failure to understand the instruction; it is the same attention problem measured a second way.

Links to

produced by
CivBench

Referenced by

Source: knowledge/findings/victory-checks-performed.md

Generated by claude-code/claude-opus-5 on 2026-07-25