fix(evidence): correct two LLM-extracted claims found by the auditor agent

Ran the new evidence-auditor agent against the two T5 arXiv records
docs/T5_HUMAN_EVAL_PROTOCOL.md cites for its attitudinal/behavioral trust
protocol change. Both papers confirmed authentic (re-fetched live arXiv
abstract pages, matched verbatim) and the protocol doc's own prose citation
is accurate to the real source -- no correction needed there.

But llama3.2's stored classification.extracted_claims had two issues the
auditor caught: arxiv_1a06e9d2bc41 (the attitudinal/behavioral paper)
extracted three generic bookend sentences from the abstract while
skipping the one sentence that actually states the distinction the
protocol change depends on -- not fabricated, just an extraction that
silently dropped the one claim that mattered. arxiv_ccfa1d1d73d9 (the
clarifying-questions paper) had one claim ("LLMs rely on user input to
generate code") that overgeneralizes what the abstract actually argues
(a specific proposed system, not LLMs in general).

Corrected both extracted_claims to reflect what the abstracts actually
say, and marked classification_reviewed=true on both -- the first
non-zero "Reviewed" rate anywhere in the evidence pipeline, from the
auditor-recommends/human-applies loop the agent was designed for working
end to end for the first time.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Author: Juanpacol

docs/EVIDENCE_COVERAGE.md

@@ -4,9 +4,9 @@ Auto-generated by `scripts/report_evidence_coverage.py`. See `.claude/agents/evi
 
 | Topic | Records | Sources | Validation pass rate | Median age | LLM-classified | Reviewed |
 |---|---|---|---|---|---|---|
-| T1 | 3 | arxiv | 100% | 0d | 100% | 0% |
-| T2 | 7 | arxiv, github_issues | 100% | 0d | 29% | 0% |
+| T1 | 23 | arxiv | 100% | 0d | 100% | 9% |
+| T2 | 6 | arxiv, github_issues | 100% | 0d | 17% | 0% |
 | T3 | 1138 | humaneval, mbpp | 100% | 0d | 0% | 0% |
 | T4 | 6 | github_issues, semgrep | 100% | 0d | 0% | 0% |
-| T5 | 7 | arxiv, github_issues | 100% | 0d | 29% | 0% |
+| T5 | 34 | arxiv, github_issues | 100% | 0d | 85% | 6% |
 | T6 | 18 | arxiv, github_issues, z3_docs | 100% | 0d | 6% | 0% |

docs/evidence/arxiv/arxiv_1a06e9d2bc41.json

@@ -41,12 +41,12 @@
       "T6": 0.0
     },
     "extracted_claims": [
-      "Trust is often cited as an essential criterion for the effective use and real-world deployment of AI.",
-      "Researchers argue that AI should be more transparent to increase trust, making transparency one of the main goals of XAI.",
-      "Properly distinguishing these two concepts provides a more comprehensive understanding of how transparency affects trust and reliance, benefiting future XAI research."
+      "The paper advocates for a clear distinction between behavioral (objective) measures of reliance and attitudinal (subjective) measures of trust.",
+      "Researchers sometimes use behavioral measures when intending to capture (attitudinal) trust, even though attitudinal measures would be more appropriate.",
+      "Properly distinguishing these two concepts provides a more comprehensive understanding of how transparency affects trust and reliance."
     ],
     "confidence": 0.8,
-    "classification_reviewed": false,
+    "classification_reviewed": true,
     "degraded_reason": null
   }
 }

docs/evidence/arxiv/arxiv_ccfa1d1d73d9.json

@@ -39,11 +39,11 @@
       "T6": 0.0
     },
     "extracted_claims": [
-      "LLMs rely on user input to generate code",
-      "Effective communication is key to increase confidence in generated code"
+      "The proposed system solicits clarifying user responses to refine LLM-generated code, rather than assuming LLMs inherently rely on user input.",
+      "Effective communication (via clarifying questions) is key to increasing confidence in generated code."
     ],
     "confidence": 0.7,
-    "classification_reviewed": false,
+    "classification_reviewed": true,
     "degraded_reason": null
   }
 }

docs/results/2026-07-16_evidence_coverage.json

@@ -0,0 +1,78 @@
+{
+  "topics": [
+    {
+      "topic": "T1",
+      "record_count": 23,
+      "sources": [
+        "arxiv"
+      ],
+      "validation_pass_rate": 1.0,
+      "median_age_days": 0,
+      "llm_classified_rate": 1.0,
+      "reviewed_rate": 0.08695652173913043
+    },
+    {
+      "topic": "T2",
+      "record_count": 6,
+      "sources": [
+        "arxiv",
+        "github_issues"
+      ],
+      "validation_pass_rate": 1.0,
+      "median_age_days": 0.0,
+      "llm_classified_rate": 0.16666666666666666,
+      "reviewed_rate": 0.0
+    },
+    {
+      "topic": "T3",
+      "record_count": 1138,
+      "sources": [
+        "humaneval",
+        "mbpp"
+      ],
+      "validation_pass_rate": 1.0,
+      "median_age_days": 0.0,
+      "llm_classified_rate": 0.0,
+      "reviewed_rate": 0.0
+    },
+    {
+      "topic": "T4",
+      "record_count": 6,
+      "sources": [
+        "github_issues",
+        "semgrep"
+      ],
+      "validation_pass_rate": 1.0,
+      "median_age_days": 0.0,
+      "llm_classified_rate": 0.0,
+      "reviewed_rate": 0.0
+    },
+    {
+      "topic": "T5",
+      "record_count": 34,
+      "sources": [
+        "arxiv",
+        "github_issues"
+      ],
+      "validation_pass_rate": 1.0,
+      "median_age_days": 0.0,
+      "llm_classified_rate": 0.8529411764705882,
+      "reviewed_rate": 0.058823529411764705
+    },
+    {
+      "topic": "T6",
+      "record_count": 18,
+      "sources": [
+        "arxiv",
+        "github_issues",
+        "z3_docs"
+      ],
+      "validation_pass_rate": 1.0,
+      "median_age_days": 0.0,
+      "llm_classified_rate": 0.05555555555555555,
+      "reviewed_rate": 0.0
+    }
+  ],
+  "total_records": 1194,
+  "generated_at": "2026-07-16T01:03:10.654378+00:00"
+}