Agent
read 20 s
write 45 s
broad 75 s
production change
failed call
wait
World
person
page
world or bot
Checks
pending
pass
fail or gate
call that bears on it
reference run
Step inspector ‹ Prev call Next call › Follow playhead
About this data (T03). Every lane comes from a recorded run of the AwsBench T03 environment: the agent lane from the aws server's meter and action log (each call's virtual cost of 20, 45 or 75 s) with arguments and results from the agent's session; routing share, per-cell ledger lag and backlog, the ASG, concurrent restarts, cellmap load-shed, the canary and settlement from the world's exported series (one sample a minute, ten seconds for routing share and restarts, shown every 30 s); host states by state and the sample host's threads against its limit_nproc from an in-process replay of the run's own calls through the world package as the run's build had it, which matches the run's exported host counts at every minute; the hold's armed state from the same replay; people, pages and world events from the timeline and the workspace's records; checks from the grader's rows. The stretch before 08:15 is the scenario's own presimulation from 06:00, the same in every run. Which checks a call "bears on" is a heuristic over its tool and arguments. The reference is the scripted oracle path replayed in process; its outcome rows are the grader's own rules on the replayed world. Runs marked dev are development runs on older builds whose env and grader bugs are fixed since; the final grader's score, where it differs, is under the verdict.
About this data (T04 ). Every lane comes from a recorded run of the AwsBench environment, not a simulation of one: the agent lane from the aws server's meter and action log (each call's virtual cost) with arguments and results from the agent's session; the world lanes from the world's own exported series (one sample a minute, shown every 30 s); people, pages, approvals and emails from the workspace's records; checks from the grader's rows. Which checks a call "bears on" is a heuristic over its tool and arguments, not the grader's attribution. T04, T05 and T06 have no reference run here: their scripted oracle paths were graded on deployed environments (D12 1.0), and those exports were not kept. Runs marked dev are development runs on older builds whose env and grader bugs are fixed since; the final grader's score, where it differs, is under the verdict.
About this data (T01). Every lane comes from a recorded run of the AwsBench T01 environment, not a simulation of one: the agent lane from the aws server's meter and action log (each call's virtual cost of 20, 45 or 75 s) with arguments and results from the agent's session; checkout, ledger-writer load and capacity, backlog, host recovery and the DNS record from the world's exported series (one sample a minute, ten seconds for the DNS state, shown every 30 s); people, pages and world events from the timeline and the workspace's records; checks from the grader's rows. The 02:30–02:40 stretch is the scenario's seeded prologue, back-filled. Which checks a call "bears on" is a heuristic over its tool and arguments, not the grader's attribution. The reference is the scripted oracle path replayed through the world package in process. Runs on older builds are development runs: their env bugs are fixed since. The final grader's score, where it differs, is under the verdict.