Use the task-package specification to choose the package layout and record evidence for the exact revision.
Start with an original tool-use scenario in which a missing prerequisite causes an observable failure. A recovery trajectory records the attempt, the repair and the checkpoint. The worked example uses a new temporary directory and a deliberately missing file.
Author and verify
Define the observable state before and after recovery. Capture ordered tool events and actual exit statuses. A Tier 1 verifier should require the failed attempt and a successful final checkpoint in the same run. A solver's claim of recovery alone carries no result. A judge can separately evaluate the explanation.
Calibrate and correct
Replay the failed start and recovery with the same inputs and then remove the failure event as a negative control. Pin the environment, capture format, verifier revision and trial identity. A dropped event is incomplete capture. Save the verifier outcome separately from any cause or repair notes, and keep raw terminal output private.
Worked example
Use Python 3.12. The script runs three real subprocesses in a temporary directory. It checks that the final read succeeds only after the recorded failure and repair.
python3.12 - <<'PY'
from pathlib import Path
from subprocess import run
from sys import executable
from tempfile import TemporaryDirectory
with TemporaryDirectory() as directory:
root = Path(directory)
read = "print(open('ready.txt').read(), end='')"
commands = [read, "open('ready.txt', 'w').write('ready')", read]
events = [run([executable, "-c", command], cwd=root, capture_output=True, text=True)
for command in commands]
def passes(trace):
return [event.returncode for event in trace] == [1, 0, 0] and trace[-1].stdout == "ready"
assert passes(events)
assert not passes(events[1:])
print("ordered recovery pass, missing failure event fails")
PYWhen packaging a trajectory set, retain verifier outcomes and stable evidence references for each run. Label this fixture as scripted. It establishes the checkpoint behavior for one controlled scenario.