Implementation: SHADOW (context compiler) · IMPLEMENTED (contracts) · Exposure: INTERNAL. See status labels. A projection is the exact set of chunks one component sees, with a fingerprint that identifies it. This page covers the records and the reasoning behind hide, short, long and full.

​
Architecture contract

  • ContextChunkV1 carries content by reference, its source provenance, trust class, sensitivity and per-model-family token estimates.
  • ContextVisibilityDecisionV1 sets each chunk to HIDE, SHORT, LONG or FULL with a relevance score and a reason code. Hiding is a view, not deletion.
  • ContextProjectionV1 lists the selected decisions, the pinned instruction fragments, a token estimate and a context_fingerprint.
Pipeline: state, projection, retrieval, dedup, ranking, diversity, dependency expansion, provenance and trust checks, compression, token budgeting, serialization for the target component. Compression happens after the next operation is known. Conditional instructions are pinned by predicate so they appear exactly when the condition holds. If the build fails, the fallback is a minimal projection or deterministic retrieval, never a transcript dump. Why the fingerprint: it lets a step be explained (“this component saw these chunks at this visibility”) and lets a replay check that it rebuilt the same input.

​
Current implementation

Only the schemas exist on main. The compiler that would produce projections is in draft behind an off-by-default flag and is not merged as an enforced path.

​
Availability

​
Developer relevance

If you build a retrieval source, emit chunks with provenance and a trust class. Do not depend on ordering or on chunks being present; visibility is decided per step.