Kai Ase Siren
Skip to content

Boundary and content ownership

Who owns a boundary, and who owns created work.

Boundary owners

Every boundary names the role holding the other side with a required owner. A boundary without one is a roster-wide rule that some roles care about more than others, which belongs in the layer that owns rules rather than here.

The owner is a relationship, not authority. It grants no permission and no executable capability. It does two mechanical things: the owner receives the body without declaring it, and loading fails when an owner declares its own boundary, since that would place one role on both sides.

Two-sided bodies

One body carries both halves under conditional headings, so the reader self-selects before reading a word of prose:

# Boundary: modify live system

Who changes running systems, and who hands that change to the role that owns it.

### If you own this boundary

You own live system modification...

### If you defer this boundary

Your clone is sealed against live mutation...

The owner section comes first, so the deferral reads as the consequence of the allocation rather than as a bare prohibition. Both sections are required, each is bounded separately at 200 words, and the whole file goes to both sides so each role can read what the other was told.

Nothing parses these headings. The roster already records who owns and who declares, so delivery, the identity card, and evaluation coverage all key off that. Headings exist for the glance.

Sections identify by relationship, never by role name. A role list in prose beside the same list in KDL is the drift this design removes.

Reading the allocation

agent-compose catalog boundaries prints one verb per cell in stable boundary order, so two runs line up beside each other:

boundary                     platform  sysadmin  science  frontend  gamedev  director  advocate
modify-live-backend          scope     OWNS      defers   defers    scope    defers    defers
build-foundational-software  OWNS      scope     scope    scope     scope    defers    defers

describe carries the same facts in a sentence past column 90, in per-role order, so it cannot be compared across roles. Four boundaries against seven roles means three roles own nothing, and a role may hold more than one scope.

Product and engine ownership

Agent Compose keeps executable parsing, validation, selection, deterministic rendering, and diagnostics in Go. Product prose and profile policy belong to reviewable local data assets.

Boundaries

  • Engine assets - generic evaluation behavior and the native adaptation policy. internal/roster/definitions/NATIVE-ADAPTATION.txt is embedded data, not a profile override.
  • Profile assets - role skills, role-bound methods, structured role metadata, role identity, invariant, copy contracts, and optional complete behavior board, which lives here under evalkit/ and evaluations/.
  • Personality-library assets - personality bindings, aliases, identity primitives, and definition skills.
  • Consumer configuration - local profile and library roots only. Agent Compose does not fetch URLs, clone repositories, resolve releases, or read git references.

Evaluation matrices

The engine supplies its complete generic matrix when a selected role has no profile asset. A profile matrix replaces that matrix as one complete unit. The loader does not merge fields, and a role cannot silently opt out.

Developer Advocate owns one connected audience loop: reusable source artifacts, provenance, claim discipline, editorial recommendations, audience research, channel adaptation, community continuity, durable feedback, qualification, discovery support, evidence selection, and decision records. These responsibilities no longer cross artificial role handoffs.

It owns recommendations about communication addressed outward, not every human-readable artifact. suggest-external-comms is the single source for which records a deferring role retains, which words a scoped seat may write inside its own artifact, and for the separate authorization publishing and sending need. This page does not restate those lists, because three hand-maintained copies had already drifted apart before the boundary existed.

Omitting a boundary

A deployment with no seat for the owning role composes without that boundary. See boundary omission.

Reviewed production locations

  • internal/evaluation/evaluation.go - typed evaluation pack rendering, generic fallback, profile matrix parsing, and validation remain executable. Role-specific profile matrices load from data assets.
  • internal/evaluation/result.go - result decoding, score validation, and canonical v2 pack digesting remain executable behavior.
  • internal/roster/roster.go - rendering remains executable. The long-form native-adaptation policy lives in definitions/NATIVE-ADAPTATION.txt.
  • internal/person/person.go - KDL parsing, local library merge, conflict detection, and copy-contract validation remain executable.
  • cmd/agent-compose/main.go - CLI help and command wiring remain executable glue. No profile, library, or customer-specific prose belongs there.

Generated artifacts retain logical source IDs and content digests. They do not publish local filesystem paths.