Reflective Context Governance Reduces Context-Borne LLM Failures — A Controlled N=120 RAW-vs-CLEAN Evaluation, with Reference Specification (v0.4) and Minimal Data Contract (v0.1)

Large language models are often evaluated as if the prompt were the primary safety boundary, yet real systems fail through context: copied secrets, stale policies, unauthorized claims, schema examples, and prompt-injection residue can all be placed into the model's reading field before the model answers. We evaluate a local-first Reflective Context Governor (RCGov) that transforms unstructured RAW context into a role-separated CLEAN Context Pack before model invocation. In a controlled N=120 benchmark (12 risk archetypes × 10 provenance/framing variants), RAW and CLEAN conditions are compared across four Groq-hosted model families (openai/gpt-oss-120b, openai/gpt-oss-20b, llama-3.3-70b-versatile, qwen/qwen3-32b). Across 480 paired model-case observations, RAW context produced 266/480 failures (55.4%) versus 10/480 (2.1%) under CLEAN — an absolute reduction of 53.3 pp and a relative reduction of 96.2% (paired McNemar: 266 RAW-fail/CLEAN-pass pairs vs 10 RAW-pass/CLEAN-fail; continuity-corrected chi-square = 235.60; no blank responses). Residual CLEAN failures concentrated in conservative legacy-appendix variants of safe-information and disputed-authority retrieval, not in secret leakage, destructive commands, review skipping, or schema canonicalization. The record bundles, as frozen reference appendices, the RCGov Technical Specification v0.4 and the Minimal Data Contract v0.1, so the evaluated system and its data model are reproducible from a single record. Deposit 2 of 2; the companion record is the concept/position paper Context Has Temporal Strata. AI co-observer: project AI systems assisted drafting; this deposit edition was prepared with Claude Fable 5 (Anthropic), working method only — the registered author is the human author alone.