docs: commit eval summary; clarify critic as LLM-assisted-judge; fix test imports

- eval/results/summary.csv: un-gitignored and committed. All numeric
claims in the README eval table (52% staleness catch, 43.9% position
accuracy, 17.1s avg latency) are now reproducible on clone.
- README: clarified that the critic uses deterministic threshold
routing for STALE/INSUFFICIENT/PASS verdicts and an LLM-as-a-Judge
prompt only for pairwise CONTRADICTED detection.
- Test_Files/test_setup.py: removed — dev-time environment check script
(Phase 1 setup) with imports no longer in requirements.txt.

.gitignore

@@ -25,7 +25,8 @@ data/cache/
 *.db
 
 # RECON — eval outputs (large, reproducible)
-eval/results/
+eval/results/*
+!eval/results/summary.csv
 eval/survey_sources.json
 eval/survey_sources_summary.txt
 

README.md

@@ -43,7 +43,9 @@ session_loader → planner → retriever → critic → synthesizer
 - `CONTRADICTED` — claims conflict across retrieved sources
 - `INSUFFICIENT` — not enough high-quality evidence to synthesize
 
-If the critic flags anything other than PASS, the retriever tries again with a refined query (max 2 retries). This retry loop is what drives the staleness catch rate improvement.
+The critic combines deterministic threshold routing with an LLM-assisted contradiction check. STALE, INSUFFICIENT, and PASS verdicts are assigned based on hardcoded thresholds (mean paper age, minimum result count, score cutoffs). CONTRADICTED is determined by calling Groq with a structured pairwise prompt that returns a `{"contradicts": bool, "reason": "..."}` JSON verdict — a canonical LLM-as-a-Judge pattern applied to contradiction detection specifically.
+
+If the critic issues anything other than PASS, the retriever tries again with a refined query (max 2 retries). This retry loop is what drives the staleness catch rate improvement.
 
 **Synthesizer** — produces a structured research position: overview, key findings, active debates, and a per-claim confidence table with source attribution.
 
@@ -102,7 +104,7 @@ This is a **reference dataset used in evaluation** — not auto-generated from l
 | Fallback retrieval | DuckDuckGo (`ddgs`) + Tavily |
 | Embeddings | `all-MiniLM-L6-v2` |
 | Session memory | SQLite |
-| Eval | LLM-as-judge + custom staleness catch rate metric |
+| Eval | LLM-assisted contradiction detection (Groq structured prompt) + custom staleness catch rate metric |
 | UI | Gradio 6.10 |
 
 One deliberate choice worth noting: the `semanticscholar` PyPI library was explicitly avoided due to a pagination hang bug on large result sets. All S2 calls go through direct `requests.get()` to `graph/v1/paper/search`.

Test_Files/test_setup.py

@@ -1,28 +0,0 @@
-from dotenv import load_dotenv
-import os
-
-load_dotenv()
-
-print("=== Testing imports ===")
-import langgraph; print("✓ langgraph")
-import langchain; print(f"✓ langchain {langchain.__version__}")
-import gradio; print(f"✓ gradio {gradio.__version__}")
-import semanticscholar; print("✓ semanticscholar")
-from sentence_transformers import SentenceTransformer; print("✓ sentence-transformers")
-import networkx; print(f"✓ networkx {networkx.__version__}")
-
-print("\n=== Testing API keys ===")
-groq_key = os.getenv("GROQ_API_KEY")
-s2_key = os.getenv("S2_API_KEY")
-tavily_key = os.getenv("TAVILY_API_KEY")
-print(f"✓ GROQ_API_KEY: {'set' if groq_key else 'MISSING'}")
-print(f"✓ S2_API_KEY: {'set — will activate in 1-3 days' if s2_key else 'not set yet (pending)' }")
-print(f"✓ TAVILY_API_KEY: {'set' if tavily_key else 'MISSING'}")
-
-print("\n=== Testing Groq connection ===")
-from langchain_groq import ChatGroq
-llm = ChatGroq(model="llama-3.3-70b-versatile", api_key=groq_key)
-response = llm.invoke("Say exactly: setup confirmed")
-print(f"✓ Groq response: {response.content}")
-
-print("\n✅ Phase 1 complete — all systems go")

eval/results/summary.csv

@@ -0,0 +1,6 @@
+architecture,total_questions,position_match_rate,staleness_catch_rate,contradiction_catch_rate,avg_latency_ms,retry_rate,error_rate
+single_rag,130,0.3231,0.0,0.3,4786.3,0.0,0.0
+naive_multi,130,0.4462,0.0,1.0,23866.0,0.0,0.0
+recon_none,130,0.4769,0.42,1.0,21817.6,0.3923,0.0
+recon_linear,130,0.4385,0.52,1.0,17094.4,0.3385,0.0
+recon_log,130,0.4308,0.38,1.0,15943.1,0.3615,0.0