Processed 1,383 of 1,550 official‑document items with the Markdown → JSONL Converter
Configured an F‑strategy that connects BM25 + embedding + reranker + Qwen direct‑evidence validator in the Source Flow Debugger
Added PostgreSQL search schema migration and embedding indexing scripts, enabling inspection of search candidates and reranker results in the debugger
Ran end‑to‑end tests on synthetic data items 36–45 using the Qwen Validation Debugger
Ensured common‑code candidates pass all Godot 3/4 engine checks to remain shared; otherwise they are split per version
Directly inspected generated code and the docs_chunks, api_mapping, label_prototypes JSONL in each chunk, confirming that expected cross‑validation results for Godot 3/4 match actual responses
Generated “irrelevant” JSONL from test items with the fewest overlapping title tokens, and improved structural checks to reject explanation comments and evidence symbols not present in the code
Updated API annotation checks to evaluate based on the target code’s identifiers and call structure, even when argument descriptions are present
Avoided hard‑coding syntax or answers in prompts; instead used only the observed structures from generated code and JSONL as validation criteria