subagentrubrics

.com durable rubric primitive
grounding

The real rubric field shape, and what this site generalizes from it

Anthropic's Managed Agents user.define_outcome tool accepts a rubric field shaped as a small tagged union, documented at platform.claude.com/docs/en/managed-agents/define-outcomes.md:

ShapeMeaning
{"type":"text","content":"# Rubric\n..."}The rubric body is inline markdown, passed directly as a string.
{"type":"file","file_id":"file_abc123"}The rubric body lives in an uploaded file, referenced by id rather than inlined.

The doc is explicit that “a rubric is a markdown document describing per-criterion scoring” and that “the rubric is required” when defining an outcome — there is no outcome without one. It also gives the one piece of authoring guidance this site takes most literally: “Structure the rubric as explicit, gradeable criteria,” not prose that merely gestures at quality.

How this generalizes into the rubrics table

The mapping is close to 1:1, the same move subagenttasks.com made generalizing the real TaskCreate/TaskUpdate schema:

Real fieldColumn here
rubric.type ("text" or "file")rubric_type — same two values, CHECK-constrained
rubric.content (inline markdown, when type: "text")criteria_md — the markdown body itself
rubric.file_id (when type: "file")would map to criteria_md holding a reference rather than a body — not exercised by the current seed row, which is type: "text"
(new, not in the real schema)criteria_count — a simple integer count of gradeable bullet points, added so a rubric's rigor is queryable without parsing markdown
(new, not in the real schema)source_doc / source_site — provenance, the same discipline every primitive in this family applies

Why this, and why now

Of the abstractions this family has already turned into primitives — identity, durable persistence, contracts, tasks, citations, cache analytics — the rubric was the one still living only as prose inside a doc mirror, never queried or referenced by anything. subagentevaluations.com needs a rubric to score against; rather than hardcoding one inline, this site is where that rubric now durably lives, the same way tasks live on subagenttasks.com and contracts live on subagentcontracts.com.