{
  "rubrics": [
    {
      "id": "rbc_cowork_feature_adoption",
      "title": "Cowork Feature Adoption Rubric",
      "summary": "Scores the second improvement batch this session: an explicit audit of which Claude Cowork features (scheduled tasks, persisted artifacts, model routing, memory/dreams) were and were not adopted while building the subagentjobs workers, with concrete deterministic follow-through rather than a vague \"consider using X\" note.",
      "criteria_md": "# Cowork Feature Adoption Rubric\n\n## Feature Gap Analysis\n- The set of unused Cowork features relevant to this repo (scheduled tasks, persisted artifacts, memory stores/dreams, model routing) is enumerated explicitly, not gestured at vaguely\n- Each feature is evaluated for actual applicability to this repo's architecture (Cloudflare Workers + D1, not MCP connectors) rather than assumed to fit\n- A feature found NOT to fit (e.g. artifacts' sandboxed network blocking non-CDN fetches, which blocks live-refreshing our own HTTPS APIs) is documented as evaluated-and-rejected with the specific technical reason, not silently dropped\n\n## Deterministic Follow-Through\n- At least one concrete, checked-in artifact results from the analysis (script, migration, code change) rather than the analysis being purely descriptive\n- The artifact is actually run and its output captured as evidence, not assumed to work\n- If an interactive Cowork tool call is rejected by the user/approval flow mid-task, the response pivots to a working alternative rather than stalling or silently dropping the goal\n\n## No Regression\n- The batch-1 fixes (Display-P3 + reduced-motion on all ~35 sites) are re-verified as still live before claiming batch 2 complete\n- The verification method is deterministic and repeatable (a committed script), not a one-off manual check that can't be reproduced next session\n\n## Honest Model-Routing Check\n- Confirms whether Fable 5 / Sonnet 5 routing decisions made earlier this session (engineering-coworker -> Fable5, Sonnet-main/Fable-advisor pairing) are still the correct assignment given no new model information, rather than re-litigating them without cause",
      "criteria_count": 9,
      "rubric_type": "text",
      "source_doc": "platform.claude.com/docs/en/managed-agents/define-outcomes.md",
      "source_site": "subagentrubrics.com",
      "created_at": "2026-07-01 23:43:40",
      "cite_as": "https://subagentrubrics.com/api/rubrics/rbc_cowork_feature_adoption"
    },
    {
      "id": "rbc_worker_quality",
      "title": "Worker Quality Rubric",
      "summary": "The rubric subagentevaluations.com actually scores this session's 9 shipped workers against. Structured as explicit, gradeable criteria per the real Managed Agents guidance: structure the rubric as explicit, gradeable criteria rather than vague ones (platform.claude.com/docs/en/managed-agents/define-outcomes.md).",
      "criteria_md": "# Worker Quality Rubric\n\n## Durable Persistence\n- If the worker is a primitive (owns real state), it is D1-backed with a real migrations directory, not in-memory or hardcoded\n- If the worker is an aggregator or docs site (no owned state), it correctly has no D1 binding rather than inventing one\n- Seed data is real and grounded in an actual source, never a synthetic placeholder row\n\n## Real API Contract\n- Exposes actual GET (and POST or PATCH where it owns state) JSON endpoints under /api/, not just server-rendered HTML\n- Write endpoints are gated behind a bearer-token secret named SITE_WRITE_SECRET\n- Responses are valid JSON with a consistent shape, the resource wrapped in a named key\n\n## Design-System Compliance\n- Shares the CSS custom properties used across the family (bg, panel, line, txt, accent) rather than inventing a new palette\n- Includes the Display-P3 widening block and a safe-area-aware sticky header\n- Any table element is wrapped in a table-wrap div for horizontal scroll on a 402px iPhone 16 Pro viewport\n\n## Provenance and Grounding\n- Content is grounded in a real, citable source such as a mirrored doc, a real API schema, or real account data, never invented to fill space\n- Where data could be mistaken for live or automated, a provenance field or badge honestly distinguishes manual entry from automation\n\n## Family Integration\n- Cross-linked from and to sibling site navigation and footers\n- Has at least one dogfood contract tracked on subagentcontracts.com",
      "criteria_count": 15,
      "rubric_type": "text",
      "source_doc": "docs/docs/platform.claude.com/docs/en/managed-agents/define-outcomes.md",
      "source_site": "subagentrubrics.com",
      "created_at": "2026-07-01 20:06:35",
      "cite_as": "https://subagentrubrics.com/api/rubrics/rbc_worker_quality"
    }
  ]
}