Research question: Does adding retrieved context have the same effect before and after ordinary domain fine tuning? Does a hybrid advantage survive a change of comparator?

Candidate selection: Chose within-task interaction analysis over an arbitrary readiness index or cross-study average. Published matched configurations support direct arithmetic without invented analyst ratings.

Unit and inclusion: All five task columns and all five LLaMA2-family configurations in RAFT v2 Table 1. GPT-3.5 reference row excluded from the main contrasts because the model family differs.

Public observations: Preserved 25 reported scores. Different task metrics stay in separate rows. No cross-task or cross-study quality average.

Base retrieval change: Base plus RAG minus base. Source-scale score points, not relative percent.

Tuned retrieval change: DSF plus RAG minus DSF. DSF is ordinary domain specific fine tuning.

Interaction contrast: Tuned retrieval change minus base retrieval change. Describes these reported configurations, not a causal effect isolated from every prompt/training difference.

Hybrid comparator: RAFT minus max(DSF, base plus RAG) within the same task. RAFT is a context-trained hybrid recipe, not every hybrid implementation.

Sign reversal: Base retrieval change > threshold AND tuned retrieval change < negative threshold.

Sensitivity: Thresholds 0, 1 and 5 source-scale points. Equality counts as within the deadband. These are analyst screening choices, not significance tests. No weights or normalization.

Manual check: HuggingFace: base retrieval 26.43 − 0.22 = 26.21; tuned retrieval 42.59 − 61.06 = −18.47; interaction −18.47 − 26.21 = −44.68.

Missing data: No missing included scores. Per-example predictions, run variance and uncertainty estimates were not available; no bootstrap, statistical significance or error-cause attribution claimed.

AWS cross-check: Keep BERTScore, LLM evaluator and latency on their native scales. The 2025 synthetic vendor experiment is a separate illustration, not a pooled replication of RAFT.

Original GS contribution: New reproducible extraction, interaction and comparator calculations; threshold analysis; a proposed diagnosis worksheet linking the evidence to pilot choices. Source experiments belong to their authors.

Limitations: Five selected task columns are not five independent enterprise deployments. Prompt, retriever, training data and evaluation choices limit transfer. Newer models and customer workflows may behave differently.

Use: Choose which comparisons to run in a pilot. Do not use these counts as the probability that RAG or tuning succeeds.

Caveat: GS Consulting derived planning analysis of cited public results. Not an official legal, audit, compliance, NIST, OWASP or regulatory determination. No new model training or inference was run.

Figure scope: Three figures: paired retrieval changes, hybrid comparator sensitivity, and diagnosis workflow. Six would repeat the small evidence set without adding a useful decision.