# Shared evaluator context; never inject this file as instructions to the model.
schema_version: 1
status: planned
default_context:
  surface: >-
    A signed-in person, Alex, is using the production in-app Agent Chat with the
    CEO of the selected synthetic company Acorn. Chat is enabled. The standing
    conversation already exists and is distinct from ordinary delegated work tasks.
  people_and_resources: >-
    Names are fixture aliases, not real user data. Morgan is a prospective or
    existing engineer as specified by each scenario; Riley is an eligible peer,
    manager or reviewer only where specified. Launch is a synthetic project.
    A mentioned task, account, document or attachment exists only when the
    scenario says it does; do not seed the requested outcome in advance.
  authority: >-
    The human is an active company operator and the CEO has the effective
    agent grants needed for the scenario's permitted agent operations. Neither
    identity implies instance administration. The CEO remains an agent, not a
    board session. Responsible-user, company, private-resource, ownership, work
    mode, budget, approval and runtime admission rules still apply. Scenario
    denials and narrower authority override these defaults.
  runtime: >-
    A supported provider, model, account and execution target are ready for
    ordinary work unless the scenario tests a missing prerequisite. Only real
    production-admitted tools, REST fallback or installed CLI/MCP surfaces count.
    Availability of a route or a declaration in an eval does not install a tool.
    Feature-specific experiments are enabled only where the scenario specifies.
  starting_state: >-
    Each scenario starts independently in an owned isolated fixture. Connections,
    policies, existing records, audiences and pending decisions are explicit in
    its context. Other scenarios are not prerequisites and their side effects
    must not leak into this one. Stated previous turns are persisted fixture
    history produced through supported surfaces, not imagined model memory.
  time: >-
    Freeze a fixture clock and record the timezone, defaulting to America/Chicago.
    Today, this week, last week and latest refer to that clock and seeded dates,
    not the date this catalog was written. Include older distractors where
    recency matters. Test daylight-saving boundaries when promoting schedules.
  external_services: >-
    Use synthetic accounts, pages, repositories, recipients and files, with
    controlled provider fixtures or separately authorized owned test accounts.
    example.test is a fixture recipient, not a real delivery destination. Keep
    secrets in supported secure setup, never queries, YAML, logs or public evidence.
  interpretation: >-
    The scenario's context overrides these defaults. Its query is a human turn,
    not a system instruction. Some context specifies a later human decision,
    browser action or fault; that must actually happen in the eventual journey.
    Pre-chat, external-client, channel, operator and developer boundary scenarios
    state their surface explicitly and must not count as direct Agent Chat passes.