Initial import: XAUUSD news sentiment models, evals, and data tooling

TF-IDF baseline (news_sentiment.py, models/xauusd_sentiment.joblib) and the
two-head DeBERTa-v3-small transformer (transformer/, news_transformer.py)
with its Google News crawler and DeepSeek labeller (data_collect/), the
golden-set and archive evals (eval/), and the 2026-09-22 and 2026-09-27
results and progress notes.

Transformer weights (model.pt, model.onnx) and the crawl database are not
tracked; rebuild with transformer/train.py and transformer/export_onnx.py.
The DeepSeek label cache in data/ is tracked.

eval/README.md

@@ -0,0 +1,54 @@
+# ML-NEWS model vs DeepSeek — evaluation scripts
+
+Research scripts only. They live here, outside the ForexBot2 git tree,
+and nothing in ForexBot2 imports them. They read ForexBot2's live news
+code and its `logs/newsdata.db` read-only, and they never write to it.
+
+Run every script with the ForexBot2 venv (it has `requests`, `pandas`
+and the live news engine; the model loads there with identical output):
+
+```bash
+cd /home/harwinder/Development/ML-NEWS/eval
+/home/harwinder/Development/ForexBot2/venv/bin/python eval_golden.py
+/home/harwinder/Development/ForexBot2/venv/bin/python eval_archive.py --reprobe 100
+/home/harwinder/Development/ForexBot2/venv/bin/python eval_live.py --max-age-hours 24
+./run_all.sh            # the three in order
+```
+
+Add `--no-network` to any of them to report from the DeepSeek cache
+without spending. DeepSeek answers are cached in
+`results/deepseek_cache.json`, keyed by headline, prompt sha1 and model,
+so a headline is paid for once; a prompt change invalidates the cache
+by itself.
+
+## What each script measures
+
+| Script | Truth | Question it answers |
+|---|---|---|
+| `eval_golden.py` | 78 reviewer-labelled headlines (`data/llm_news_golden.csv`, the 2026-08-06 set where DeepSeek scored 78/78 and keywords 24/78) | Is the model *right*, per category, including negation traps and inverted metrics; and is DeepSeek still 78/78 today |
+| `eval_archive.py` | DeepSeek's own archived XAU scores | How well the recipe copies DeepSeek on days it never saw (expanding daily folds, headline+description vs headline only, keyword fallback for context); which confidence/margin thresholds keep how many headlines local at what agreement; how often DeepSeek agrees with itself on re-scoring (`--reprobe`), which is the ceiling for any student |
+| `eval_live.py` | DeepSeek live, on today's feed | Shadow mode: agreement on fresh headlines from the real feeds, split by gold-mentioning / not-in-training; accumulates in `results/live_shadow.jsonl` across days |
+
+Why three: the model was trained on DeepSeek's labels, so agreement on
+the archive says how well the student copies the teacher, not whether
+either is right. The golden set is the only independent truth. The live
+log grows the out-of-sample set over time.
+
+## Grading rules
+
+* Golden set: a continuous XAU score is bullish at ≥ +0.15, bearish at
+  ≤ −0.15, neutral in between (the August rule). The model is graded by
+  its native argmax label and, separately, by its score through the
+  same band.
+* Archive and live: DeepSeek xau → class at ±0.05, the ML-NEWS trainer's
+  own label rule; model = argmax label.
+* `dir` = expected non-neutral rows predicted with the right sign (a
+  neutral prediction is a miss). `sign` = same sign when both lean.
+  McNemar is exact and two-sided on discordant pairs.
+
+## Outputs
+
+`results/golden_<ts>.{json,csv}`, `results/archive_<ts>.json`,
+`results/archive_oof_<ts>.csv` (every out-of-fold prediction),
+`results/archive_reprobe_<ts>.csv`, `results/live_<ts>.json`,
+`results/live_shadow.jsonl` (append-only, deduplicated by headline hash).