Pattern Proposal: Grounded Prompt Fact Projection

Summary

Extracts validated facts from code changes and architectural registers to project a deterministic factual basis into agent prompts.

Candidate Pattern

• ID: grounded-prompt-fact-projection
• Classification: intelligence
• Maturity: established
• Type: implemented

Observed Signals

No specific signals listed.

Evidence

Supporting Evidence

• src/cas_runner/steering/models.py (PromptFact, SteeringRuntimeIndex, ArchitecturalDelta) - Defines typed prompt facts and index schema grounding LLM prompts
• .github/workflows/cas-build-steering-projection.yml - CI workflow triggering projection generation

Scope Assessment

• Evidence Strength: high
• Confidence: high
• Applies to: ideas-to-life-cas

Rationale

• Intent: Prevent LLM hallucinations during governance checks by strictly grounding reasoning on verified capability diffs and historical rules.
• Context: Invoked during steering runtime indexing and prompt preparation for architectural decision agents.
• Force: Full repository context exceeds LLM prompt token windows and introduces irrelevant noise
• Force: Raw Git diffs lack architectural semantic context

Traceability

Artifact source: proposed-patterns