Evidence Verify

Changelog / 变更日志

EvidenceVerify

证据验证

Cross-source verification, credibility scoring, and Twitter signal validation for debate claims. 辩论声明的跨来源验证、可信度评分和 Twitter 信号验证。

When to Use / 何时使用

• When Judge audits a round (primary use) / Judge 审计回合时（主要用途）
• When Orchestrator needs to re-verify claims during refresh / Orchestrator 在刷新时需要重新验证声明
• When a debater wants to pre-verify their own evidence / 辩手想预验证自己的证据时

Input / 输入

• claim: A ClaimItem with evidence_ids[]
• evidence_store_path: Path to evidence_store.json
• independent_search: Boolean — if true, run fresh WebSearch (used by Judge)

Output / 输出

• Updated status recommendation for the claim
• verification_notes: Explanation of verification result
• confidence_level: "high" | "medium" | "low"

Core Workflow / 核心工作流

Step 1: Gather Referenced Evidence / 收集引用的证据

Resolve evidence_ids to full EvidenceItem objects from the evidence store.

If any evidence_id is not found in the store, flag as evidence_missing and note in verification.

Step 2: Source Independence Check / 来源独立性检查

For critical claims (claims central to the argument or with high impact):

• Require ≥ 2 independent sources (different publishers, different URLs)
• Sources from the same publisher group do NOT count as independent
• Wire services (AP, Reuters) redistributed by multiple outlets count as ONE source

For non-critical claims (peripheral supporting points):

• 1 reputable source is acceptable

Use LLM judgment to determine criticality based on how central the claim is to the debate.

Step 3: Twitter Signal Validation + Corroboration (v3) / Twitter 信号验证 + 关联确认

If ANY evidence supporting a claim has source_type = "twitter":

1. Check social_credibility_flag (set by SourceIngest):

   • If likely_unreliable: Flag claim as high verification priority. Auto-set status ceiling to unverified unless strong independent evidence exists.
   • If needs_verification: Standard verification process, but actively search for corroboration.

2. Corroboration search (enhanced from v2):

   • For verification_priority = "high" Twitter claims: ALWAYS run an independent WebSearch query derived from the claim text
   • Search specifically for non-social (tier1/tier2/tier3) sources that confirm or deny the claim
   • Update the Twitter EvidenceItem's corroboration_status:
     • corroborated: Found ≥1 independent non-social source confirming
     • uncorroborated: No independent confirmation found
     • contradicted: Found independent source(s) that directly contradict the claim

3. Status determination (unchanged logic, new data):

   • Twitter + corroborated → eligible for verified (if other conditions met)
   • Twitter + uncorroborated → maximum unverified with note "twitter-only, uncorroborated"
   • Twitter + contradicted → contested with note "twitter claim contradicted by [source]"
   • Twitter + likely_unreliable + uncorroborated → unverified with note "flagged (suspected misinformation)"

Step 4: Cross-Source Agreement / 跨来源一致性

Use LLM to compare snippets across evidence items:

1. Do the sources agree on the core factual claim?
2. Are there contradictions between sources?
3. Do sources add complementary details or conflict?

If sources agree: strengthens verification If sources partially conflict: flag specific points of disagreement → contested If sources fully contradict: → contested with detailed note

Step 5: Independent Re-Search (Judge Only) / 独立重新搜索（仅 Judge）

When independent_search = true (invoked by Judge):

1. Run fresh WebSearch with queries derived from the claim text
2. Compare results against debater-provided evidence
3. Check: Does independent search confirm, contradict, or add nuance?
4. This is the Judge's independent verification — do NOT trust the debater's citations alone

Step 6: Determine Status / 判定状态

Based on all checks:

Error Handling / 错误处理

• Cannot fetch original source for re-verification → note as "source_unreachable" but don't auto-invalidate the claim
• Independent search returns no results → note "no independent confirmation found" but don't auto-reject