MaxTask

The work is irreducibly distributed

The UK National Archives holds eleven million documents. The US National Archives holds thirteen billion pages. No organisation, model, or volunteer community can process all of it. Genealogical research will always be done by millions of people and machines in parallel, over decades. MaxTask is the standard way to describe one piece of that work.

What it is

A task is a request: "transcribe this image", "decide whether the Johann Müller in record X is the John Miller in record Y", "search the Württemberg emigration lists for Müllers who left between 1840 and 1860", "resolve these two contradictory birth years". A task can go to an AI agent, a volunteer, or a specialist organisation. It can carry a cost cap, a deadline, dependencies on other tasks, and a review step before its result is accepted.

What makes it different

• Acceptance criteria (acceptance_criteria[]): objective, checkable conditions that define "done" before work starts.
• Evidence required (evidence_required[]): the anti-hallucination rule. A task that needs a birth record must return a link to the birth record. A search must return the exact query and what it found. An AI cannot pass by producing plausible text.
• A structured result, not a text blob: what was tested, what was done, what was found, what changed, what to do next. Each step carries a verdict:

• An independent review (review{}): the reviewer must not be the contributor. Rejection reasons include fabricated_or_hallucinated and privacy_violation, which build a quality signal over time.
• A privacy gate (contributor_eligibility): defaults to first_party_only. Any task about a living person can never be sent to outside volunteers or third-party AI services. This rule is part of the standard.

What is deliberately not in it

Payment. An open standard must be neutral on business models. Who gets paid, how much, and how lives in extensions{}, so the standard describes the work and the quality gate while each implementation decides the economics.

Task types

extract_record validate_record merge_person judge_review resolve_conflict ingest_to_qdrant build_tree fetch_internet_archive cost_report other

Fields

Example

A conflict-resolution task escalated to a stronger model because two candidate mothers were within 0.05 of each other.

{
  "task_id": "c1000000-0016-4000-8000-000000000016",
  "schema_version": "1.13",
  "task_type": "resolve_conflict",
  "status": "done",
  "priority": 9,
  "payload": {
    "person_id": "b1000000-0012-4000-8000-000000000012"
  },
  "result": {
    "resolution": "adoptive_mother_confirmed",
    "winning_parent_id": "b1000000-0030-4000-8000-000000000030",
    "confidence": 0.72
  },
  "assigned_to": "opus-conflict-resolver",
  "cost_usd": 4.2,
  "cost_cap_usd": 5.0,
  "retry_count": 0,
  "max_retries": 1,
  "escalation_reason": "Conflicting biological vs adoptive mother with confidence gap 0.05",
  "created_at": "2026-05-04T15:00:00Z",
  "created_by": "judge-agent-sonnet-001",
  "started_at": "2026-05-04T15:10:00Z",
  "completed_at": "2026-05-04T15:45:00Z"
}

Listen

← All seven schemas