Experiments
Selected measurements from the execution simplification work, with scope and limitations.
Recorded direct runs
One direct artifact run recorded 10 generations, 4 tool calls and 198.05 seconds. A repair run recorded a failed test, a model-owned repair, a passing test and an unchanged-hash check in 11 generations, 6 tool calls and 209.15 seconds. A shell-produced CSV run recorded 8 generations, 4 tool calls and 158 seconds, with a verified total of 42 across two rows.
Recorded managed run
One durable managed run recorded 11 generations, 7 tool calls and 217.43 seconds, with persisted plan completion and retained artifact evidence. These are bounded runs, not a general performance benchmark.