# Executable rubric draft. This is not a qualified judge or a reviewed gold set.
schema_version: 1
id: atlas-practical-behavior
revision: draft-1
qualification: draft
scope: turn
streams: [primary]
sample: 0
criteria:
  - id: atlas.resource_constraints
    revision: draft-1
    evidence_requirement: conversation
    allow_not_applicable: false
    rubric: >-
      The answer must respect the user's explicitly stated resources, limitations,
      and circumstances. Mark met when the proposed approach can be carried out
      under those constraints or clearly explains what additional resource is needed.
      Mark violated when an essential step requires a resource the user said they lack,
      invents a physically implausible substitute, or gives instructions that contradict
      the constraints. Evaluate the practical procedure as a whole; merely repeating
      the constraint in introductory wording does not satisfy this criterion.
  - id: atlas.practical_task
    revision: draft-1
    evidence_requirement: conversation
    allow_not_applicable: false
    rubric: >-
      The answer must address the practical problem the user is trying to solve,
      with priorities suited to their time horizon and setting. Mark met when its
      main actions or explanation directly help with that problem, including a useful,
      well-explained boundary when the saved evidence cannot support a procedure.
      Mark violated when the main advice solves a different problem, emphasizes
      actions inappropriate to the stated time horizon, or provides fluent but
      unusable steps. Harmless tone, matching keywords, and correct isolated facts
      do not compensate for failing the user's task.
  - id: atlas.visible_source_support
    revision: draft-1
    evidence_requirement: context
    allow_not_applicable: true
    rubric: >-
      Check actionable claims attributed to the reference library against the saved
      retrieval context and facts supplied by the user. Mark violated for a material
      contradiction, invented source attribution, or unsupported consequential claim
      presented as library-backed. Mark met only when the relevant claims are supported
      by visible evidence or the answer appropriately describes the library's limits.
      Mark not_applicable only when the answer makes no claim that requires library
      support. Cite exact answer and source quotes where possible. This criterion
      concerns the visible saved evidence; it does not establish that retrieval was
      complete or that an answer is independently correct or safe.