Phase 2: UI reliability badges, changelog, eval runner — Phase 2 complete

CHANGELOG.md

@@ -2,6 +2,49 @@
 
 ## [Unreleased]
 
+### Phase 2 — Edge Reliability Scoring + OpenAlex Augmentation
+- 2.1: Created src/openalex_utils.py
+  - `search_openalex(query, max_results=5)` — keyword search, returns Paper-compatible dicts
+  - `get_openalex_by_doi(doi)` — singleton DOI lookup for citation enrichment
+  - `get_citation_centrality(doi, citation_count)` — `min(1.0, log1p(count) / log1p(10000))`
+  - Abstract reconstruction from OpenAlex inverted index format
+  - DOI URL stripping (`https://doi.org/` prefix removed)
+  - API key auth via `OPENALEX_API_KEY` env var; graceful degradation without key
+- 2.2: Added DOI extraction to retriever_utils.py
+  - Added `"externalIds"` to Semantic Scholar API field list
+  - Extracts `doi = (r.get("externalIds") or {}).get("DOI", "") or ""`
+  - Passes `doi=doi` to Paper constructor
+  - Added `doi: str = ""` field to Paper dataclass in state.py (after references, before hybrid_score)
+- 2.3: OpenAlex augmentation in retriever.py
+  - Triggered when S2 retrieval returns < 12 papers
+  - Searches top 2 sub-questions, max 3 results each
+  - Deduplicates by DOI and paper_id before merging
+  - OpenAlex papers get hybrid_score from semantic sim (0.3) + authority + recency
+- 2.4: Created src/reliability.py — three-signal paper reliability scorer
+  - Signals: centrality (40%), recency (30%), coherence (30%)
+  - Centrality: `min(1.0, log1p(citation_count) / log1p(10000))`
+  - Recency: `max(0, 1 - (age_years / 10))`
+  - Coherence: batched LLM call (one call for all papers) — scores 0–1 relevance to query
+  - Dominant signal labels: FOUNDATIONAL / CURRENT / DECLINING / SUPERSEDED
+  - FOUNDATIONAL: `centrality >= 0.60 AND (coherence >= 0.65 OR coherence == 0.0)` — coherence passthrough for LLM-off case
+  - Returns `{paper_id: ReliabilityScore}` dict
+- 2.5: Wired reliability scorer into critic.py
+  - `score_papers()` called after high-score paper check
+  - STALE threshold changed from `mean_age > 24` to `mean_reliability < 0.40`
+  - Falls back to age-based threshold if scorer returns empty
+  - `paper_reliability_scores` added to ResearchState TypedDict
+  - Scores serialised to `{pid: rs.__dict__}` in all non-INSUFFICIENT critic return paths
+- 2.6: Added per-paper trust summary to synthesizer output
+  - Reads `paper_reliability_scores` from state after synthesis
+  - Appends `## Evidence Trust Summary` block to `synthesized_position`
+  - Color-coded signal labels: HIGH (FOUNDATIONAL/CURRENT), MEDIUM (DECLINING), LOW (SUPERSEDED)
+  - Trust summary included in `export_md` via `_build_export_md`
+- Warning fix: `_headers()` in openalex_utils.py downgraded from WARNING to DEBUG when API key absent (was spamming 14+ warnings per query during centrality scoring)
+- Eval results (recon_linear_v2_full, 130Q):
+  - Staleness (STALE verdict): 32.3% (42/130) — down from 52% v1; reliability-based threshold is more precise, fewer false-positive STALE calls
+  - Contradiction catch rate: 0% (retriever still returns adjacent-topic papers, not opposing-camp pairs — known gap, deferred)
+  - Position accuracy (MATCH): 44.6% (58/130) — unchanged vs v1 43.9%
+
 ### Phase 1 — Integrity + Critic Fix
 - 1.1: Archived patch_contradiction.py to eval/archived/ with README
 - 1.2: Fixed critic_node — STALE and CONTRADICTED checks now run in parallel

app.py

@@ -43,7 +43,14 @@ def _highlight_citations(text: str) -> str:
         text
     )
 
-def _paper_cards_html(papers) -> str:
+SIGNAL_COLORS = {
+    "FOUNDATIONAL": "#22c55e",
+    "CURRENT":      "#3b82f6",
+    "DECLINING":    "#f59e0b",
+    "SUPERSEDED":   "#ef4444",
+}
+
+def _paper_cards_html(papers, reliability_scores: dict = {}) -> str:
     """Render retrieved papers as styled cards."""
     if not papers:
         return "<p style='color:#6b7280;font-style:italic'>No papers retrieved.</p>"
@@ -54,15 +61,29 @@ def _paper_cards_html(papers) -> str:
         authors = ", ".join(p.authors[:2]) + (" et al." if len(p.authors) > 2 else "") if p.authors else "Unknown"
         abstract_preview = (p.abstract[:180] + "...") if p.abstract and len(p.abstract) > 180 else (p.abstract or "")
 
+        rs = reliability_scores.get(p.paper_id)
+        if rs:
+            dominant = rs.get("dominant_signal", "DECLINING") if isinstance(rs, dict) else rs.dominant_signal
+            sig_color = SIGNAL_COLORS.get(dominant, "#6b7280")
+            signal_badge = (
+                f'<span style="background:{sig_color}22;color:{sig_color};border:1px solid {sig_color}44;'
+                f'padding:2px 7px;border-radius:12px;font-size:0.75em;white-space:nowrap;font-weight:600;margin-left:6px">'
+                f'{dominant}</span>'
+            )
+        else:
+            signal_badge = ""
+
         cards.append(f"""
 <div style="border:1px solid #2d3748;border-radius:8px;padding:12px 14px;
             margin-bottom:8px;background:#1a1f2e;">
   <div style="display:flex;justify-content:space-between;align-items:flex-start;margin-bottom:4px">
     <span style="font-weight:600;font-size:0.92em;color:#e2e8f0;flex:1;margin-right:8px">{p.title}</span>
-    <span style="background:{score_color}22;color:{score_color};border:1px solid {score_color}44;
-                 padding:2px 8px;border-radius:12px;font-size:0.78em;white-space:nowrap;font-weight:600">
-      {p.hybrid_score:.3f}
-    </span>
+    <div style="display:flex;align-items:center;flex-shrink:0">
+      <span style="background:{score_color}22;color:{score_color};border:1px solid {score_color}44;
+                   padding:2px 8px;border-radius:12px;font-size:0.78em;white-space:nowrap;font-weight:600">
+        {p.hybrid_score:.3f}
+      </span>{signal_badge}
+    </div>
   </div>
   <div style="color:#94a3b8;font-size:0.8em;margin-bottom:6px">
     {authors} · {p.year} · {p.citation_count:,} citations · <span style="color:#64748b">{p.source}</span>
@@ -205,7 +226,7 @@ def run_query(query, session_id, decay_config, history):
     verdict_html  = _verdict_badge_html(verdict, critic_notes, retry_count,
                                          papers_used, latency, decay_config, rewritten)
     claims_html   = _claims_html(result.get("claim_confidences") or [])
-    papers_html   = _paper_cards_html(result.get("retrieved_papers") or [])
+    papers_html   = _paper_cards_html(result.get("retrieved_papers") or [], result.get("paper_reliability_scores") or {})
     session_ctx   = load_session(session_id)
     session_html  = _session_html(session_ctx, session_id)
     export_md     = result.get("export_md", "")