Skip to main content

Context design rationale and evidence

Layered context aims to keep the project policy visible while loading task-specific guidance on demand. This is a design choice to assess against actual tasks and tools, not a guarantee that a smaller rules file improves every model.

The production source authority is skills/process/prometheus-context-bootstrap/scripts/: bootstrap.sh handles legacy managed regions, bootstrap-v4.sh handles rules-source layout, and the corresponding verifiers inspect their own bounded predicates. UI routing and project-team adoption call the separate portable helpers documented in the canonical guide.

Historical model-specific prompt reductions, local word counts and skill-budget measurements describe their recorded environments. They are not current provider specifications, installed hook acceptance or release evidence. Consult current official harness documentation before asserting loading, compaction, imports or native control behavior.

Evaluate context changes with a representative real task set and recorded model routes after all production changes are coherent. Track source identity, actual loaded instructions, tokens, outcomes and independent review limits. A passing static layout check cannot verify hidden inference context or establish an unbiased judge.

Use UI routing, team models, local validation and the actual project policy to choose an evidence boundary. Do not publish historical measurements as fresh acceptance.