You are labelling test items for a memory system. Read the file /tmp/claude-1000/-home-ubuntu-Documents-veracium-veracium-research/418bd771-544d-451a-a2d0-95f45d8d9a49/scratchpad/h5-labeller/h5_items_for_labelling.json (it is the only file
you need; read nothing else). It is a JSON array of items, each with "id", "relation", "object" and "text": a text a
person gave to an assistant app, and ONE candidate fact about that person.

For each item answer:

  GOLD QUESTION: In this text, does the USER, in their own words and in a part of the text that is theirs, state as
  currently true about themselves that <relation meaning>: <object>? A bare agreement with another speaker's wording
  is false.

The thirteen relations and what each means:
  avoids_tool    the user avoids, has stopped using, or refuses to use a tool, app or service
  client_of      the user is a client or customer of a named business or professional
  has_diet       the user follows a diet or an eating rule
  has_pet        the user has a pet (the object is the animal or its name)
  health_state   the user has a health condition or is in a health state
  located_at     the user lives in or is based in a place
  measures       a measurement of the user (height, weight, a size, a reading)
  partner_of     the user's partner or spouse is the named person
  prefers        the user prefers something
  relative_of    the named person is a relative of the user
  uses_tool      the user uses a tool, app or service
  works_as       the user's job, role or profession
  works_on       a project, product or piece of work the user is working on

Also give "value_source": where the object's words appear in the text: "user", "other" (only another speaker),
"both", or "neither".

Reply with ONE JSON array, one object per item, in the file's order: {"id": ..., "gold": true or false,
"value_source": ...}. Nothing else.
