Feronix — Qwen3-4B LLM validation protocol Version: 30 September 2026 1. Technical schema validation All required fields, data types and score limits are checked. Invalid or truncated output is retried. 2. Evidence validation Exact supporting excerpts are located in the model input after whitespace normalisation. Untraceable evidence lowers the quality score and flags the analysis as uncertain. 3. Diagnostic benchmark Six compact, prelabelled financial cases were used as a technical content check. Qwen3-4B completed all six correctly, with a 100% evidence ratio. This small benchmark is not a full external validation. 4. Blinded human sample A reproducibly drawn sample from both datasets will be manually labelled without revealing future price outcomes. Agreement will be reported for each field. 5. Stability Quality and model outcomes will be broken down by company, year, release length and confidence. An average figure must not conceal unstable subgroups. 6. Final predictive test Qwen features will be compared with a price-based baseline and earlier text features using the same chronological splits and horizons of 1, 5 and 20 trading days. 7. Reporting Positive, negative and neutral outcomes will be published with sample sizes, uncertainty, limitations and prespecified costs.