Spaces:
Running
Running
| { | |
| "domain": "game", | |
| "scenario_id": "wikiwiper", | |
| "reset_payload": { | |
| "seed": 42, | |
| "scenario_id": "wikiwiper", | |
| "mission_id": "wikiwiper_game_sample", | |
| "options": { | |
| "env_id": "omnibench_aegis_env:game.wikiwiper", | |
| "max_steps": 8, | |
| "target_score": 1 | |
| } | |
| }, | |
| "action_examples": { | |
| "canonical": [ | |
| { | |
| "name": "inspect_objective", | |
| "args": {} | |
| }, | |
| { | |
| "name": "scan_zone", | |
| "args": {} | |
| }, | |
| { | |
| "name": "select_tool", | |
| "args": { | |
| "tool_slot": "hotbar.2" | |
| } | |
| }, | |
| { | |
| "name": "navigate_route", | |
| "args": { | |
| "route": "lower_corridor" | |
| } | |
| }, | |
| { | |
| "name": "engage_threat", | |
| "args": { | |
| "aggressive": false | |
| } | |
| }, | |
| { | |
| "name": "wipe_target", | |
| "args": { | |
| "target_id": "target-archive-core" | |
| } | |
| }, | |
| { | |
| "name": "verify_cleanup", | |
| "args": {} | |
| }, | |
| { | |
| "name": "submit_run", | |
| "args": {} | |
| } | |
| ], | |
| "shorthand": [ | |
| { | |
| "action": "inspect_objective" | |
| }, | |
| { | |
| "action": "scan_zone" | |
| }, | |
| { | |
| "action": "select_tool", | |
| "tool_slot": "hotbar.2" | |
| }, | |
| { | |
| "action": "navigate_route", | |
| "route": "lower_corridor" | |
| }, | |
| { | |
| "action": "engage_threat", | |
| "aggressive": false | |
| }, | |
| { | |
| "action": "wipe_target", | |
| "target_id": "target-archive-core" | |
| }, | |
| { | |
| "action": "verify_cleanup" | |
| }, | |
| { | |
| "action": "submit_run" | |
| } | |
| ] | |
| }, | |
| "expected_good_trajectory": [ | |
| "inspect_objective", | |
| "scan_zone", | |
| "select_tool", | |
| "navigate_route", | |
| "engage_threat", | |
| "wipe_target", | |
| "verify_cleanup", | |
| "submit_run" | |
| ], | |
| "expected_flow": [ | |
| "health", | |
| "reset", | |
| "step", | |
| "state" | |
| ], | |
| "notes": [ | |
| "Aligned fixture for the real wikiwiper domain.", | |
| "Uses domain-level semantic actions instead of raw control vectors.", | |
| "The selected tool slot and target_id match the default seeded wikiwiper variant used in local smoke/eval runs." | |
| ] | |
| } | |