Skip to main content

One Semantic Model, Multiple Representations

MRS is a semantic score model plus protocols for safe, bounded editing. It is commonly represented in two complementary forms: An implementation may keep canonical state in a database/graph form for performance, then export MRS-S as needed. Agents read MRS-S content in Working Set Envelopes. Agents write MRS-Ops operations. The orchestrator validates and applies operations to canonical state.

File Types

  • MRS-S: .mrs or .mrs-sapplication/vnd.mrs+sexpr
  • MRS-Ops: .mrs-opsapplication/vnd.mrs-ops+sexpr
  • Encoding: UTF-8 (no BOM)
See: /MRS-Specification-RFC#16-mime-types-and-file-extensions

Minimal MRS-S Score

Key points:
  • Every measure has :id (UUID) and :number (display)
  • Every event has :id (UUID) — no :at required in storage (orchestrator computes it)
  • Players own instruments; instruments appear in measure content

Working Set Envelope

For agent tasks, the orchestrator extracts a Working Set Envelope:
Key features:
  • Scope uses UUIDs — not measure numbers
  • :bundle grants lane permissions
  • :context-views provide task-relevant information
  • :content is valid MRS-S the agent can read

Agent Output: MRS-Ops

Agents write typed operations, not complete fragments:
Key features:
  • Explicit operations: create-event, not “here’s the whole fragment”
  • Temporary IDs: "e1", "s1" — orchestrator assigns UUIDs
  • No derived fields: No :at required — orchestrator computes it
  • Spans reference tmp-ids: "e1" in same envelope is valid

Orchestrator Response

The Workflow

Why This Works

See: /MRS-Specification-RFC#5-mrs-ops-the-mutation-protocol