E3 · LED wording: branch exp/led-wording (codes.md says led_output_low names the effect whatever the cause). Provider-error attempts are left out. Not part of the headline.

Compare models

6 agent(s) · 18 scored trial(s) · built from whatever records are in the runs directory

Leaderboard

#agent all pass scoredharness errorstasks mean end score (partial credit)range (partial credit) cost / trialtotal cost input tok / trialoutput tok / trial reasoning tok / trial
1opus
claude-opus-5-5
2 / 3 67%3—10.990.971–1$2.1070$6.32113840879339090
2argon
google/gemini-4-argon
2 / 3 67%3—10.9830.949–1$2.8301$8.490374238596434334186
3sol
gpt-6.1-sol
2 / 3 67%3—10.6670–1$0.4573$1.37202108674104374123
4astra
gpt-6-astra
1 / 3 33%3—10.990.971–1$3.4665$10.39952262920100893366
5sonnet
claude-sonnet-5-5
1 / 3 33%3—10.9830.968–1$1.4488$4.34635167717473530
6terra
gpt-6.1-terra
0 / 3 0%3—10.3330–0.998$0.2213$0.66381910358118065685

Ranked by all-pass: a session passes only if every check holds and nothing unsafe happened. The mean end score is partial credit, shown for context: a session can score high while failing a required check. These are the grades stored in the records.

All pass by task

Sessions that passed (every check holds, nothing unsafe) out of the scored sessions, per agent and task.

agentt4_takeover
argon2 / 3 67%
astra1 / 3 33%
opus2 / 3 67%
sol2 / 3 67%
sonnet1 / 3 33%
terra0 / 3 0%

Mean end score by task (partial credit)

agentt4_takeover
argon0.983
n=3
astra0.99
n=3
opus0.99
n=3
sol0.667
n=3
sonnet0.983
n=3
terra0.333
n=3

Progress reached

Scored trials that reached at least each stage of the task's progress ladder (grade.progress_stage).

t4_takeover

agentrig_checkedcalibratedfloodedsegmentedstimulatedanalysedvalid_report
argon3/33/33/33/33/33/33/3
astra3/33/33/33/33/33/33/3
opus3/33/33/33/33/33/33/3
sol3/33/33/33/33/33/33/3
sonnet3/33/33/33/33/33/33/3
terra3/33/33/33/33/33/33/3

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fault's tracking assumption stated as "no" at or after onset); named in report is scored; cleared means the fault was no longer active in the final truth.

agentfaulttrialsnoticed (assumption → no)median decisions to noticeworstnamed in reportcleared by the end
argonled_output_low33 / 3003 / 33 / 3
argonled_power_loose30 / 3——0 / 33 / 3
argonprojector_standby33 / 3003 / 33 / 3
argonroom_light33 / 3033 / 33 / 3
argonsample_cover_off30 / 3——3 / 33 / 3
astraled_output_low33 / 3003 / 33 / 3
astraled_power_loose30 / 3——0 / 33 / 3
astraprojector_standby33 / 3003 / 33 / 3
astraroom_light33 / 3223 / 33 / 3
astrasample_cover_off30 / 3——3 / 33 / 3
opusled_output_low33 / 3003 / 33 / 3
opusled_power_loose30 / 3——0 / 33 / 3
opusprojector_standby33 / 3003 / 33 / 3
opusroom_light33 / 3003 / 33 / 3
opussample_cover_off30 / 3——3 / 33 / 3
solled_output_low33 / 3003 / 33 / 3
solled_power_loose30 / 3——0 / 33 / 3
solprojector_standby33 / 3003 / 33 / 3
solroom_light33 / 3223 / 33 / 3
solsample_cover_off30 / 3——3 / 33 / 3
sonnetled_output_low33 / 3003 / 33 / 3
sonnetled_power_loose30 / 3——0 / 33 / 3
sonnetprojector_standby33 / 3002 / 33 / 3
sonnetroom_light33 / 3113 / 33 / 3
sonnetsample_cover_off30 / 3——3 / 33 / 3
terraled_output_low33 / 3003 / 33 / 3
terraled_power_loose30 / 3——0 / 33 / 3
terraprojector_standby33 / 3002 / 33 / 3
terraroom_light33 / 3223 / 33 / 3
terrasample_cover_off30 / 3——3 / 33 / 3

Cost against end score

One panel per agent on the same axes; one dot per scored trial (hover for the trial).

argon 3 scored trial(s)
00.510$2.093$4.186t4_takeover · t_present_faults · trial 0: end score 1, $3.1572, 82541 output tokenst4_takeover · t_present_faults · trial 1: end score 1, $2.9347, 77072 output tokenst4_takeover · t_present_faults · trial 2: end score 0.949, $2.3984, 33415 output tokenscost per trial (USD)
astra 3 scored trial(s)
00.510$2.093$4.186t4_takeover · t_present_faults · trial 0: end score 0.971, $3.7774, 10950 output tokenst4_takeover · t_present_faults · trial 1: end score 0.998, $3.9867, 10826 output tokenst4_takeover · t_present_faults · trial 2: end score 1, $2.6354, 8492 output tokenscost per trial (USD)
opus 3 scored trial(s)
00.510$2.093$4.186t4_takeover · t_present_faults · trial 0: end score 0.971, $2.1222, 34429 output tokenst4_takeover · t_present_faults · trial 1: end score 1, $2.0319, 28023 output tokenst4_takeover · t_present_faults · trial 2: end score 1, $2.1670, 39275 output tokenscost per trial (USD)
sol 3 scored trial(s)
00.510$2.093$4.186t4_takeover · t_present_faults · trial 0: end score 1, $0.4582, 8491 output tokenst4_takeover · t_present_faults · trial 1: end score 1, $0.4650, 14949 output tokenst4_takeover · t_present_faults · trial 2: end score 0, $0.4488, 7872 output tokenscost per trial (USD)
sonnet 3 scored trial(s)
00.510$2.093$4.186t4_takeover · t_present_faults · trial 0: end score 0.968, $1.5941, 50999 output tokenst4_takeover · t_present_faults · trial 1: end score 1, $1.4731, 51232 output tokenst4_takeover · t_present_faults · trial 2: end score 0.981, $1.2791, 39829 output tokenscost per trial (USD)
terra 3 scored trial(s)
00.510$2.093$4.186t4_takeover · t_present_faults · trial 0: end score 0, $0.1611, 9768 output tokenst4_takeover · t_present_faults · trial 1: end score 0.998, $0.2533, 12500 output tokenst4_takeover · t_present_faults · trial 2: end score 0, $0.2494, 13150 output tokenscost per trial (USD)