Sonnet won 1 of 4 comp1 games (25%)

small-sample mixed play result

Source paper

extracted_from

(2026) · Robert Müller · Clemens Müller

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Sonnet 4.5 win rate=35.7% (n=14)finding0.791
Sonnet's win rate in exploratory games
Haiku wins 7.1% of mixed gamesfinding0.727
Very low win rate against code agents.
Sonnet 4.5 TrueSkill μ=26.4 ± 4.9 (n=14, 35.7% win rate)finding0.725
Mid-field performance with larger uncertainty due to small sample.
G3.1-FL wins 50.0% of 28 mixed gamesfinding0.719
Half the games won against code agents.
GPT5.4-N wins 14.3% of mixed gamesfinding0.715
Similarly poor against code agents.
TrackerAgent win rate 53.6% in 98 canonical gamesfinding0.712
TrackerAgent won over half of the combined-comp1 games.
Claude 3.5 Sonnet shows higher rate of alignment-faking reasoning than Claude 3 Opus in helpful-only setting but almost none in animal welfarefinding0.699
Linked to Claude 3.5 Sonnet not exhibiting pro-animal-welfare preferences
G3-F wins 67.9% of 28 mixed games (vs three code agents)finding0.698
Robust performance against algorithmic baselines.