What this is

Claude (Anthropic) acts as the final AI reviewer for every auto-approved claim before SOL is dispatched. This is the last verification layer before funds move.

How it works

After the policy engine auto-approves a claim, Claude receives the complete verification payload plus cross-pipeline intelligence:

All 13 gate results (pass/fail + metadata)
Fraud scores from Gemini and Grok
Cross-validation signals between receipt, tweet, and wallet
Submission metadata and fraud flags
X account metrics (follower count, account age, engagement history)
Submission history for the wallet (previous claims, approval rate, flags)
Entity intelligence from mem0: trust trajectory (improving/stable/declining), cross-pipeline flags from referral/engagement/payout pipelines, recent Claude verdicts on this wallet, 7-day submission velocity, and notable behavioral patterns
Institutional context from knowledge base: dynamically selected gate rules, known fraud patterns, and current policy thresholds relevant to this specific claim's risk profile

What Claude returns

Verdict: approve / flag / reject
Confidence score: 0-1 scale indicating certainty
Audit narrative: Written explanation of the decision reasoning, stored permanently in the audit log

Flag behavior

If Claude flags a claim, it reverts from auto-approved to needs_review status. An admin must manually review the claim with full visibility into Claude's narrative and all underlying signals before funds can be released.

Graceful fallback

If the Claude API is unavailable (timeout, rate limit, outage), the policy engine's original decision stands. The claim is not blocked — it proceeds with the deterministic gate verdict. API availability is logged for operational monitoring.

Architecture summary

  Gemini sees ──→ Grok thinks ──→ mem0 remembers ──→ Claude decides
  (vision/OCR)    (reasoning)      (history)          (oversight)

Three independent AI systems from three different providers, backed by persistent cross-pipeline memory, ensure no single point of model failure or bias can compromise the verification pipeline. Claude's verdict is written back to mem0, building an ever-growing intelligence corpus that informs future reviews.