Smoke runs, not benchmark results. These trials check the pipeline. They carry no ranking and support no comparison between configs or harnesses.
Smoke runs
Each run checked the path from a trial to the analysis summary to this page: Harbor ran each trial in a container, the injector logged its decisions, and the analysis read the job directories and wrote the summary shown here. Newest first.
track-s
12 trials, 32 faults applied, between 12 and 57 eligible tool calls per trial. Track S, fx-claude and cc-claude on three service tasks, one attempt per task and profile (D-0030). The tables pool both configs. One attempt per cell supports no comparison between them. The cost columns undercount fx-claude: some of its trials have no cost in result.json because the gateway settled them later; the settled costs are in settled-usage.json next to the summary. Source: results/smoke/2026-10-15-track-s/summary.json.
Per profile
Profile
Trials
pass^1
Total cost
Mean cost
Faults applied
clean
6
100.0%
$0.4988
$0.0998
0
moderate
6
100.0%
$0.4545
$0.1136
32
Per trial
Task
Profile
Reward
Cost
Eligible calls
Completion status
s-duplicate-merges-duplicate-merge__2HXwTSk
clean
1.00
$0.1163
51
success
s-duplicate-merges-duplicate-merge__HPFQagH
clean
1.00
-
51
success
s-followupss-followups__cPgaDXA
clean
1.00
$0.1171
15
success
s-followupss-followups__jcXcKHA
clean
1.00
$0.0703
12
success
s-p1-reports-p1-report__QeAsNYZ
clean
1.00
$0.1250
22
success
s-p1-reports-p1-report__o6FV6Bz
clean
1.00
$0.0700
22
success
s-duplicate-merges-duplicate-merge__J2Nuj3e
moderate
1.00
$0.1302
57
success
s-duplicate-merges-duplicate-merge__yXP3Tam
moderate
1.00
$0.1677
57
success
s-followupss-followups__G7CFa9x
moderate
1.00
-
21
success
s-followupss-followups__hkc4tfB
moderate
1.00
$0.0783
17
success
s-p1-reports-p1-report__R7UjSsp
moderate
1.00
$0.0784
26
partial
s-p1-reports-p1-report__sXP77Wn
moderate
1.00
-
26
partial
6 trials, no fault applied, between 2 and 4 eligible tool calls per trial. Track N, cc-claude only. The tasks of this run were later rebuilt under D-0023, D-0024 and D-0026, so its trials sit below the exposure floor of 8 in SPEC 10.3 and the moderate profile had few eligible calls to fault. Source: results/smoke/2026-10-11/summary.json.