falsepass_init
Initializes an audit directory for a specified RL environment and returns its audit name, so agents can start probing for false passes and false fails.
Instructions
Set up audits// for an environment and return the name to audit it with.
target: a Prime Hub slug (owner/environment), a Python file of reward functions written for TRL's GRPOTrainer (def accuracy_reward(completions, solution, **kwargs); or a module name such as trl.rewards with framework "reward") together with dataset (a .jsonl/.json/.csv/.parquet file, hf:owner/name, or call:function in the file; optional functions, weights, split, dataset_config, answer_column, prompt_column, map (a row mapper in the file), path (folders put on sys.path for a package that is imported rather than installed, e.g. a clone's src)), an OpenEnv Space (hf:owner/space, cloned next to the audit) or folder (openenv.yaml), a local folder with a pyproject.toml (verifiers or Inspect project), an Inspect task file, an Inspect registry name (inspect_evals/gsm8k), or a HUD environment folder. Options: name (folder name, default from the environment), version (Hub version), framework ("verifiers", "openenv", "reward", "inspect" or "hud" when detection is wrong), task (Inspect: which @task), answers (HUD and OpenEnv: path to a JSON file of reference answers, an object from task key to answer or a list in task order; without it only answer-free probes run, unless an OpenEnv task record carries an answer field), extra (more requirements to install, e.g. ["inspect-evals[math]"]; added to the environment's own), without (requirement names to drop from the environment's own, e.g. a pin that does not resolve), env_class / action_class (OpenEnv, "module:Class": the environment class or factory and the action class when detection picks the wrong one), manual (also write an adapter.py skeleton for a Q&A environment automatic mode cannot script), policies (the same for games, simulations and tool tasks: a skeleton built around policies; use it when an audit comes back "inconclusive", or for a verifiers v1 tool taskset once the automatic run of generic tool policies is done, to add a solver as the control and task-aware exploits; v1 episodes run through the taskset's own loop; for an OpenEnv MCP tool environment, or one whose reward comes after several steps, the policies play whole episodes through its reset() and step()), force (rewrite an existing audit's env.toml, e.g. to add answers; adapter.py is kept).
Tool call template
Replace each placeholder with a value matching the input schema. Only required parameters are included; add optional parameters as needed.
{
"id": 1,
"jsonrpc": "2.0",
"method": "tools/call",
"params": {
"arguments": {
"target": "<target>"
},
"name": "falsepass_init"
}
}Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| map | No | reward functions: a row-mapping function in the file | |
| name | No | folder name under audits/ (default: from the environment) | |
| path | No | folders put on sys.path for a package that is imported rather than installed, e.g. a clone's src | |
| task | No | Inspect: which @task of the file | |
| vars | No | environment variables the environment reads, set on every run, e.g. {"DATA_PATH": "/abs/path/data"} (a relative path is taken from workdir) | |
| extra | No | more requirements to install, e.g. ["inspect-evals[math]"] | |
| force | No | rewrite an existing audit's env.toml (adapter.py is kept) | |
| split | No | reward functions: the dataset split | |
| manual | No | also write an adapter.py skeleton, for a Q&A environment automatic mode cannot script | |
| target | Yes | what to audit: a Prime Hub slug owner/environment, an OpenEnv Space hf:owner/space, a local folder, an Inspect task file or registry name, a HUD folder, or a file or module of reward functions (with `dataset`) | |
| answers | No | HUD and OpenEnv: a JSON file of reference answers (task key to answer, or a list in task order) | |
| dataset | No | reward functions: a .jsonl/.json/.csv/.parquet file, hf:owner/name, or call:function in the file | |
| version | No | Hub version to install (default: the latest) | |
| weights | No | reward functions: one weight per function | |
| without | No | requirement names to drop from the environment's own, e.g. a pin that does not resolve | |
| workdir | No | the folder whose audits/ holds the audits (default: the current directory) | . |
| policies | No | write a policy-based adapter.py skeleton, for a game, simulation or tool task | |
| env_class | No | OpenEnv: "module:Class", the environment class or factory, when detection picks the wrong one | |
| framework | No | "verifiers", "openenv", "reward", "inspect" or "hud", when detection is wrong | |
| functions | No | reward functions: the functions of the file to use (default: all) | |
| action_class | No | OpenEnv: "module:Class", the action class | |
| answer_column | No | reward functions: the dataset column that holds the reference answer | |
| prompt_column | No | reward functions: the dataset column that holds the prompt | |
| dataset_config | No | reward functions: the Hugging Face dataset config name |