GS Consulting: RAG vs fine tuning research
Analysis date: October 27, 2026

This package contains GS calculations using published observations. It does
not contain a new training run, customer outcomes or a current model ranking.
The original experiments belong to Zhang et al. (RAFT v2, June 5, 2024)
and AWS (synthetic product-catalog example, May 28, 2025).

REPRODUCE THE NUMBERS
1. Keep the CSV files and reproduce_analysis.py in the same directory.
2. Run: python3 reproduce_analysis.py
3. The script uses only the Python standard library. It recalculates the
   interaction and comparator results with decimal arithmetic, checks them
   against derived_scores.csv and prints the threshold analysis.

The Excel workbook contains formula-driven calculations and the full source
register. Inputs are source observations, not a calibrated forecasting tool.
Different task metrics are not averaged. See methodology.txt for formulas,
exclusions, screening thresholds, missing data and transfer limits.

The SVG and PNG figures are under figures/. Figure 3 is a proposed diagnostic
workflow, not an experimental result. SVGs retain editable text and an embedded
GS logo. Figure numbers correspond to their placement in the article.

Source URLs and exact locations are in source_register.csv. The source manifest
records hashes of captured research materials. Full AWS article text and model
artifacts are not redistributed in this package.

GS Consulting derived planning analysis. Not an official legal, audit,
compliance, NIST, OWASP or regulatory determination. No statistical significance
or expected client improvement is established by these aggregate scores.
