Thevindu-models Dont merge just check only

[ThevinduFernando2003]

Summary

Do not merge thevindu-models into main. This branch records a laptop drop-in bake-off. Live defaults stay where they are.

• Generator: keep Qwen 2.5 3B Q4. Gemma JSON validity 0.00; Phi grounded hits 0.00 and JSON 0.00 (garbage tokens under empty chat_format). Phi must beat Qwen on both metrics — it did not.
• Judge: keep Llama 3.2 3B. Granite matched 5/5 gold labels at ~12% slower (promotable candidate, not flipped). Gemma-as-judge is not a drop-in (SystemMessage rejected).
• Retrieve: keep MiniLM-L6 + L-6 reranker + hybrid BM25. The real company-law PDF reversed the toy-fixture E5 and L-12 wins. One BM25-only chunk survived rerank top-3 — leave hybrid on.
• Components: each comparison model now has Dinura-style chat_agent/ + evaluator/ files (re-exports of live integrated-backend/learnmate, bound to that model id) and Thevindu-style thresholds / checklist / model card / version registry.

.env was not edited. No new selectable_default. No GGUF or secrets in this diff (*.gguf / models/ / .env are gitignored).

Decision rules (applied, not implied)

• Phi / Gemma promote only if they beat Qwen on both grounded hits and JSON validity.
• Granite promotes only if it matches/beats Llama on gold agreement and has no large latency regression. Match + ~12% is a candidate, not a flipped default.
• Retrieve verdicts use the real PDF (results/real_pdf.json), not the 36-chunk toy fixture.
• Same-family generator+judge is refused before load.

Machine

Windows laptop CPU (AMD64, Intel Family 6 Model 186), backend venv Python 3.13.7. Not Colab T4. Not uvicorn chat p95. End-to-end chat p95 was not measured (stack was down); retrieve ms must not be labeled as serving p95.

Raw evidence

Table	Backing file
Generators / judges	thevindu-models/results/gguf.json
Real-PDF retrieve	thevindu-models/results/real_pdf.json
Toy retrieve (history only)	thevindu-models/results/retrieval.json
Written verdicts	thevindu-models/RESULTS.md
Per-model cards	thevindu-models/01_generators/*/RESULTS.md, 02_judges/*/RESULTS.md, 03_embeddings/*/RESULTS.md

Four-eyes checklist (reviewer: confirm in writing)

• Every number states which machine it was measured on
• The real-PDF result (not the toy fixture) is what the retrieve verdict is based on
• The stated decision rules above were actually applied
• No selectable_default or .env value changed anywhere in the diff
• Raw results/*.json backs every table number, not just the summary sentence

Test plan

• Open thevindu-models/RESULTS.md Decision table and §4–§6; check each cell against the JSON files
• Confirm git check-ignore still catches integrated-backend/models/*.gguf and .env
• Confirm models_registry.yaml still has only qwen25-3b as selectable_default: true
• Optional: from integrated-backend, run eval_components.py (live generate_node / Judge) if you want app-prompt numbers in results/components.json
• Do not merge after review