Ghostlight/docs/architecture/projected-local-context.md
2026-05-02 04:20:18 +01:00

3.2 KiB

Projected Local Context

The projected local context is the missing seam between canonical Ghostlight state and a character agent.

Canonical state is storage. It can contain means, plasticity, activation, relationship stance, memories, culture, goals, and author-facing source paths. A character response model should not have to interpret that machinery. It should receive a compact operating context: what this character knows, wants, misreads, can do, sounds like, and must not resolve.

Contract

Pipeline:

canonical state + scene + reviewed annotation
  -> projector/state interpreter
  -> ghostlight.projected_local_context.v0
  -> response-model prompt
  -> generated action proposal

The projected context may retain source references for audit. The rendered prompt text must not expose raw state internals such as current_activation, plasticity, means, or selected numeric dimensions.

Artifact Shape

Schema: schemas/projected-local-context.schema.json

Validator: tools/validate_projected_contexts.py

Generator: tools/project_local_context.py

Current examples:

  • examples/projected-contexts/scene-02-sanctuary-intake.maer_tidecall.projected-context.json
  • examples/projected-contexts/scene-02-sanctuary-intake.sella_ren.projected-context.json

The context contains:

  • speaker identity and setting
  • known facts and speaker beliefs
  • current stakes
  • embodiment and interface affordances
  • active inner pressures rendered as ordinary language
  • relationship read and known biases
  • tensions
  • action affordances
  • projection controls
  • voice surface
  • likely response moves
  • do-not-invent boundaries
  • rendered prompt text

Runtime Boundary

The character model receives local operating context, not omniscient truth. If Maer acts, Sella's next pass receives only the observable event and Sella's own local context. Maer's private intent is not included.

The sequential runtime still owns:

  • tool contracts
  • action enums
  • event proposal capture
  • participant appraisal
  • reviewed mutation
  • Ink materialization

The projector owns the translation from Ghostlight's internal state language to character-local operating prose.

Current Limit

The first projector is deterministic and deliberately simple. It uses threshold language like "dominant", "strong", "present", and "background" rather than numeric scores. This is enough to remove raw state soup from Qwen prompts and create inspectable training artifacts.

Later versions can become smarter, but they should still preserve this boundary: response models act from projected local context; they do not become ad hoc interpreters of the canonical schema.

Embodiment Boundary

Embodiment is part of the prompt surface, not flavor garnish. If a character is a cetacean Navigator, the response model must know what kind of body is acting and what infrastructure surrounds it.

For Navigator scenes, project habitat-native affordances directly: channels, wet/dry workstations, shared displays, control rails, acoustic or translation systems, and other infrastructure built around cetacean movement and communication. Do not let the model default to hands, pockets, standing, walking, leaning, or doorway blocking unless the fixture explicitly establishes that visible interface.