Robotics episode probe v1

• Stability: development
• Revision: v1-development
• Scope: repeated closed-loop episodes grouped by task under one starting policy

This convention supports the synthetic CLI fixture. A real adapter must also record the policy, runtime, dataset, task, embodiment, and evaluation contract used to produce the measurements. Natural termination and external truncation remain separate in the native evidence.

A profile exports only the coordinates defined by its native study. The list below covers the broader robotics program and does not require every study to manufacture every coordinate.

target_relevance

Study-defined similarity or relevance between selected experience and the frozen evaluation target. Unit: normalized_score. The estimator must identify the representation, target split, and aggregation rule.

policy_weakness

Study-defined extent to which selected experience exposes failures or weak behavior of the starting policy. Unit: normalized_score. Direction is contextual because severe failures may be outside the policy's learnable frontier.

selection_coverage

Study-defined coverage or diversity of the selected episodes, states, tasks, or operators. Unit: normalized_score. The native record defines the support and distance function.

quality_correction

Magnitude of the study's data-quality adjustment before or during candidate construction. Unit: normalized_score. The coordinate records an intervention property and does not assume that stronger filtering improves downstream performance.

measurement_reliability

Agreement of the policy-relative measurement across declared independent views, checkpoints, or seeds. Unit: normalized_score. The agreement statistic remains part of the measurement contract.

missingness_rate

Fraction of expected native measurements absent after runtime and sensor accounting. Unit: fraction. Missing episodes are not unsuccessful episodes.

episode_success_variation

Mean per-task Bernoulli variance of repeated episode success. Unit: bernoulli_variance. The coordinate describes observed variation under the starting policy and has contextual direction.

mixed_task_coverage

Fraction of measured tasks with at least one successful and one unsuccessful episode. Unit: fraction. The sample count is the number of tasks.

mean_action_steps

Arithmetic mean number of executed action steps over valid episodes. Unit: action_steps_per_episode. Truncated, failed, and invalid episodes must remain distinguishable in the native evidence.