NoeFlandre/audit-trail

Audit Trail tasks

This dataset lists the tasks of the Audit Trail OpenEnv environment. The environment code is in the GitHub repository.

What the environment teaches

The agent gets one log and a question. The question states a rule. The agent searches the log with grep and gives one answer. The rule has two steps: find the right lines, then compare them. The logs come from three families:

Family	Domain	Rule
auth	Cybersecurity	The earliest successful login from a source IP that had at least K failed logins before it. Answer: the user name.
ledger	Finance	The earliest debit that repeats an earlier debit with the same account and amount, 1 to 90 seconds before it, where the earlier debit was never reversed. Answer: the txn id.
sensor	Natural science	Among stations whose first reading at or after 06:00 of 30.0 C or more is followed by another reading of 30.0 C or more, the station with the earliest such reading. Answer: the station name.

Each task has a level from 0 to 7. Levels 0 to 2 are easy, 3 to 5 medium and 6 to 7 hard. Harder levels have longer logs, more decoys and a higher K.

Files

File	Contents
tasks.jsonl	One task for each line: task_id, split, family, level and the accepted seed.

The dataset does not contain the logs. The server builds each log from its seed, so the same seed always gives the same log and answer. If a seed does not give a log that passes the checks, the server tries the seed plus 100, then plus 200, and so on. The accepted seed is in tasks.jsonl.

Reward

• 1.0: the answer is correct. Case does not matter.
• 0.0: the answer is wrong, or the agent uses all 10 steps.

Splits

All 24 tasks are in the train split.

Limits

• The logs are synthetic. They do not come from a real system.
• A grep returns at most 10 matching lines.
• Each episode has 10 steps.

License

Apache License 2.0.