E5 · Scenario sweep: 65 test scenarios × 1 trial on the frozen code (no change to what the model sees); terra and sol, experiments/sweep1_openai. All-pass: Terra 44/65, Sol 48/65. Not part of the headline.
Compare models
2 agent(s) · 128 scored trial(s) · 2 harness failure(s) excluded
· built from whatever records are in the runs directory
Leaderboard
Agents were not run on the same task and scenario set, so their means are not directly comparable; use the per-task table.
#
agent
mean end score
range
all pass
scored
harness errors
tasks
cost / trial
total cost
input tok / trial
output tok / trial
reasoning tok / trial
1
terra
gpt-6.1-terra
0.951
0–1
44 / 65
65
—
5
$0.1534
$9.9713
1279094
8948
4192
2
sol
gpt-6.1-sol
0.923
0–1
48 / 63
63
2
5
$0.2661
$17.9984
1222832
7384
2472
Mean of each trial's end grade over the scored trials. Harness failures are counted in
their own column and never scored. Means over a handful of trials are noisy; read the range and the per-task table
before ranking.
Mean end score by task
agent
t1_connectivity
t2_setup
t3_light
t4_takeover
t5_traces
sol
0.84
n=14
0.964
n=15
0.923
n=14 · 2 harness err.
0.909
n=11
0.998
n=9
terra
0.984
n=14
0.951
n=15
0.935
n=16
0.905
n=11
0.98
n=9
Progress reached
Scored trials that reached at least each stage of the task's progress ladder
(grade.progress_stage).
t1_connectivity
agent
rig_checked
calibrated
flooded
segmented
stimulated
analysed
valid_report
sol
14/14
14/14
14/14
14/14
14/14
14/14
14/14
terra
14/14
14/14
14/14
14/14
14/14
14/14
14/14
t2_setup
agent
calibrated
segmented
submitted
valid_report
sol
15/15
15/15
15/15
9/15
terra
14/15
13/15
13/15
8/15
t4_takeover
agent
rig_checked
calibrated
flooded
segmented
stimulated
analysed
valid_report
sol
11/11
11/11
11/11
11/11
11/11
11/11
11/11
terra
11/11
11/11
11/11
11/11
11/11
11/11
11/11
t5_traces
agent
read_docs
analysed
submitted
correct
sol
9/9
9/9
9/9
9/9
terra
9/9
9/9
9/9
8/9
Recovery by fault
Per fault that occurred in a scored trial. Noticed is the grader's time-to-notice diagnostic
(the fault's tracking assumption stated as "no" at or after onset); named in report is scored; cleared means the
fault was no longer active in the final truth.
agent
fault
trials
noticed (assumption → no)
median decisions to notice
worst
named in report
cleared by the end
sol
bump_drift
1
1 / 1
3
3
1 / 1
1 / 1
sol
calibration_missing
2
2 / 2
0
0
2 / 2
2 / 2
sol
calibration_not_applied
2
0 / 2
—
—
0 / 2
2 / 2
sol
calibration_stale
1
0 / 1
—
—
0 / 1
1 / 1
sol
dead_culture
4
4 / 4
13
25
2 / 4
4 / 4
sol
dropped_frames
5
2 / 5
8
16
5 / 5
5 / 5
sol
hdmi_loose
3
0 / 3
—
—
0 / 3
3 / 3
sol
hdmi_signal_lost
3
3 / 3
1
14
3 / 3
3 / 3
sol
led_output_low
6
2 / 6
0
0
5 / 6
6 / 6
sol
led_power_loose
2
0 / 2
—
—
0 / 2
2 / 2
sol
leds_aged
2
0 / 2
—
—
0 / 2
0 / 2
sol
lens_distortion
2
1 / 2
3
3
1 / 2
0 / 2
sol
loose_wire
5
5 / 5
0
4
5 / 5
5 / 5
sol
mask_flip
3
3 / 3
2
3
3 / 3
3 / 3
sol
photobleaching
4
0 / 4
—
—
4 / 4
4 / 4
sol
projector_standby
3
2 / 3
2.5
5
3 / 3
3 / 3
sol
red_led_aged
2
0 / 2
—
—
0 / 2
0 / 2
sol
room_light
8
4 / 8
1
2
7 / 8
8 / 8
sol
sample_cover_off
4
0 / 4
—
—
4 / 4
4 / 4
sol
sample_moved
2
0 / 2
—
—
2 / 2
2 / 2
sol
slide_damaged
1
1 / 1
3
3
1 / 1
0 / 1
sol
stale_frames
9
5 / 9
1
15
9 / 9
9 / 9
sol
storage_pressure
2
0 / 2
—
—
2 / 2
2 / 2
sol
trigger_cable_loose
2
0 / 2
—
—
0 / 2
2 / 2
sol
uneven_illumination
2
2 / 2
6.5
13
2 / 2
0 / 2
sol
usb_loose
2
0 / 2
—
—
0 / 2
2 / 2
terra
bump_drift
1
1 / 1
4
4
0 / 1
1 / 1
terra
calibration_missing
2
1 / 2
0
0
2 / 2
2 / 2
terra
calibration_not_applied
2
0 / 2
—
—
0 / 2
2 / 2
terra
calibration_stale
1
0 / 1
—
—
1 / 1
1 / 1
terra
dead_culture
4
4 / 4
9
19
2 / 4
4 / 4
terra
dropped_frames
5
2 / 5
6
12
5 / 5
5 / 5
terra
hdmi_loose
3
0 / 3
—
—
0 / 3
3 / 3
terra
hdmi_signal_lost
3
3 / 3
1
13
3 / 3
3 / 3
terra
led_output_low
6
5 / 6
7
30
6 / 6
6 / 6
terra
led_power_loose
2
0 / 2
—
—
0 / 2
2 / 2
terra
leds_aged
2
0 / 2
—
—
0 / 2
0 / 2
terra
lens_distortion
2
0 / 2
—
—
0 / 2
0 / 2
terra
loose_wire
5
5 / 5
0
4
5 / 5
5 / 5
terra
mask_flip
3
3 / 3
0
5
3 / 3
3 / 3
terra
photobleaching
3
0 / 3
—
—
3 / 3
3 / 3
terra
projector_standby
3
1 / 3
0
0
2 / 3
3 / 3
terra
red_led_aged
2
0 / 2
—
—
0 / 2
0 / 2
terra
room_light
8
4 / 8
0.5
2
7 / 8
7 / 8
terra
sample_cover_off
4
0 / 4
—
—
4 / 4
4 / 4
terra
sample_moved
2
0 / 2
—
—
2 / 2
2 / 2
terra
slide_damaged
1
1 / 1
4
4
1 / 1
0 / 1
terra
stale_frames
10
6 / 10
0
13
10 / 10
10 / 10
terra
storage_pressure
1
0 / 1
—
—
1 / 1
1 / 1
terra
trigger_cable_loose
2
0 / 2
—
—
0 / 2
2 / 2
terra
uneven_illumination
2
1 / 2
18
18
2 / 2
0 / 2
terra
usb_loose
2
0 / 2
—
—
0 / 2
2 / 2
Cost against end score
One panel per agent on the same axes; one dot per scored trial (hover for the trial).