Document EmpathRAG Core merged plan

Implement EmpathRAG Core hybrid router

Add Karthik eval harness and safety patches

Start V2.5 support navigator hardening

Checkpoint V2 curated support navigator

Add curated corpus integration scaffold

Start v2 safety hardening

chore: harden .gitignore — data artifacts + session artifact patterns

chore: add .claude/ to .gitignore

docs: correct metric attribution and condition count

docs: refine README structure and eval status

Revamp README: badges, pipeline diagram, results tables, eval status, known limitations

merge: resolve README conflict, keep university attribution

docs: update university attribution in README

Update course details in README

Clean repo: fix README with verified metrics, pin requirements, update gitignore, remove log from tracking

Add conversation memory, fix Gradio stack, improve SYSTEM_PROMPT, log human eval turns

Add sliding window conversation memory - 3 turns, n_ctx 4096

Tighten SYSTEM_PROMPT with few-shot example, reduce max_tokens to 200, fix paragraph post-processor

Improve SYSTEM_PROMPT for conversational peer support - validate, reflect, one question

Fix Condition C ablation - pure FAISS order, no safety score bias - mean=0.50

Add ablation Condition C eval - mean alignment 0.40 vs D=0.88

fix: MistralJudge schema detection for truths/claims/verdicts — matches DeepEval 3-call pattern (Day 15)

fix: MistralJudge uses from_json_schema grammar — correct structure for claim extraction and verification (Day 15)

fix: GBNF grammar-constrained JSON sampling in MistralJudge — eliminates DeepEval JSON parse errors (Day 15)

fix: MistralJudge JSON extraction for DeepEval faithfulness, commit Wilcoxon results (Day 15)

fix: wilcoxon sys.path for condition_a import, zero_method=pratt for binary scores (Day 15)

fix: Wilcoxon uses stub guardrail — tests retrieval quality on all 50 prompts not 13 (Day 15)

feat: DeepEval FaithfulnessMetric with Mistral judge, async_mode=False (replaces RAGAS, Day 15)

fix: guardrail dual-import path, bertscore key names, ragas reuse pipeline.llm (Day 14)

eval: save adversarial results at both t=0.50 and t=0.85, reset production threshold to 0.50 (Day 14)

fix: threshold sweep — calibrate DeBERTa guardrail threshold, patch eval imports (Day 14)

feat: eval scripts — BM25 baseline, adversarial probes, BERTScore, RAGAS, Wilcoxon (Day 14)

feat: Gradio demo app, DeBERTa Colab notebook, updated smoke test results (Day 13)

feat: real DeBERTa guardrail wired, skip_ig flag, smoke test updated

feat: wire real DeBERTa guardrail, fix smoke test for crisis intercepts

fix: wire real guardrail, anchor gitignore, fix dtype warning

fix: anchor /models/ gitignore to repo root only

fix: guardrail_ig.py — IG in embedding space, token_type_ids, correct dtype

fix: scope models/ gitignore to root only, add src/models package, remove .claude

fix .gitignore

Add Colab notebooks: emotion classifier + corpus annotation (Day 12)

Add eval suite: test prompts, adversarial probes, BERTScore references (Day 12)

Add pipeline orchestrator + smoke test — 4/5 emotion predictions correct (Day 12)

Day 3: correct gitignore, add data manifest

Day 3: proper Colab ipynb notebook for FAISS build

Day 3: FAISS index builder notebook for Colab A100

chore: keep requirements-lock.txt tracked for reproducibility

Day 2: NLI pairs built, fix gitignore for large files

Day 1: data pipeline, session tracker, query router, adversarial probes, Colab training notebooks