Feronix — Qwen3-4B LLM methodology Version: 30 September 2026 Objective Qwen3-4B LLM converts passages from official earnings releases filed with the SEC into a fixed set of financial features. The language model itself does not predict share prices. Input - Selected passages from the official earnings release only. - Up to 7,000 input characters across a maximum of six passage windows. - Ticker and SEC acceptance time for identification. - No future price information. Model settings - Model: Qwen3-4B via a local Ollama installation. - Temperature: 0. - Context window: 8,192 tokens. - Structured output following a fixed JSON schema. Extraction Scores are recorded for sentiment, management confidence, revenue, earnings, margins, demand, guidance, expectations, risk, liquidity, one-off items and capital distributions. Restructuring and customer, supply-chain and legal issues are also recorded as booleans. Quality - Every non-neutral conclusion must be supported by a short, verbatim excerpt. - Supporting excerpts are automatically located in the input. - Confidence and ambiguity are saved for each release. - Truncated JSON is retried with additional output capacity. - Each completed release is immediately saved in resumable storage. Evaluation Qwen features will be compared with price-based baselines and earlier text features only after processing is complete. The independent final test remains reserved until then. Limitation Verbatim evidence checks confirm provenance, not necessarily the correctness of an interpretation. A blinded human validation sample is therefore part of the protocol.