chat model, answer-letter probabilities · v1 set (earlier version)

Gemma 4 31B

Earlier version

Results f24d1f41, updated 16 November 2026, 22:20 UTC; scored by GNU Backgammon 1.08, 2-ply.

How we called it

Model ID
gemma-4-31B-it
Provider or endpoint
our GPUs, FP8 (Cygnet's letter-reading server)
Date run
Missing
Price
no price per token; own hardware; price date not recorded
Model type
chat model, answer-letter probabilities
Weights
open
Runs on
our GPUs
Reasoning setting
none (letter read directly)
Confidence source
letter probability
Rules text
short (the v1 rules text; the same stored request text for every model)
Option groups
groups of up to 26 options; no decision in this data was split
Sampling settings
Missing
Notes
calibration setting fitted for Cygnet's 12B model: changes reported confidence, not the play picked

Missing B7 fields: sampling, date_run

Results in v1 set (earlier version)

Earlier version: plays were listed in the runner's order, which tends to put the best play early. When Gemini 4 Argon answered the same 1,000 positions with the plays shuffled, its error rate changed by +1.9 (95% interval −0.2 to +4.2), so the effect is not proven.

Error rate (PR)
23.520.1 to 26.9
Decisions
1,000 (1,000 scored, 0 invalid)
Engine's best play
54.4%
Blunder rate
17.5%
P95 equity loss
0.251
Mean loss, worst 5%
0.411
Median / P90 s per decision
0.59 / 0.87
$ per 1,000 decisions
Not measured
Risk judgement, margin 0.001
63.3 (95% interval 60.3 to 66.3); risks found 36.3% (129 / 355), right safety 90.2% (333 / 369)
Risk judgement, margin 0.02
63.9 (95% interval 60.9 to 66.9); risks found 38.4% (122 / 318), right safety 89.4% (363 / 406)
No trade-off
276 (left out of every risk metric)

JSON

Risk call card

63.3risk judgement, margin 0.001

right risks129
wrong risks24
needless risks36
missed risks202
right safety333

Hit risk only: being hit. Gammon risk and priming are not in it.

By game phase

PhaseError rate (PR)Decisions
OpeningToo few decisions143
Contact29.0641
Holding gameToo few decisions20
RaceToo few decisions136
Bear-offToo few decisions60