GASCOIN's verification system operates in seven independent layers. Each layer catches a different class of attack. All layers work together to ensure no fraudulent submission reaches payout.

  LAYER 1   Gemini Vision               Receipt OCR + tamper detection + EXIF forensics
  ─────────────────────────────────────────────────────────────────────────────────
  LAYER 2   Grok Reasoning              Cross-validation + tweet quality + fraud reasoning
  ─────────────────────────────────────────────────────────────────────────────────
  LAYER 3   X API v2                    Tweet verify + account quality + social graph
  ─────────────────────────────────────────────────────────────────────────────────
  LAYER 4   Claude Oversight            Final review + audit narrative + payout gate
  ─────────────────────────────────────────────────────────────────────────────────
  LAYER 5   Referral Pipeline           Ring detection + anti-farm + auto-verify
  ─────────────────────────────────────────────────────────────────────────────────
  LAYER 6   mem0 Intelligence           Cross-pipeline memory + trust trajectory
  ─────────────────────────────────────────────────────────────────────────────────
  LAYER 7   Knowledge Base              Institutional rules + fraud pattern library

Layer 1 — Gemini Vision

Google Gemini processes every receipt image: OCR extraction of structured data (date, amount, station, wallet characters), tamper detection scoring, EXIF forensics analysis, and image integrity verification. This is the first line of defense against fabricated or manipulated receipt images.

Layer 2 — Grok Reasoning

Grok (xAI) performs cross-validation of all extracted signals, tweet quality analysis, pattern detection across submission history, and fraud reasoning. Two independent scores (AI probability and tamper score) must both fall below their thresholds. This layer catches sophisticated forgeries that pass visual inspection.

Layer 3 — X API v2

The submitter's X account is evaluated for authenticity signals: follower count, account age, posting history, engagement patterns, and network quality. xAI-powered scoring identifies bot accounts, purchased followers, and coordinated networks. This layer makes it economically infeasible to create fake accounts at scale.

Layer 4 — Claude Oversight

Claude (Anthropic) reviews every auto-approved claim before SOL is dispatched. It receives all gate results, fraud scores, cross-validation signals, mem0 entity intelligence, and knowledge base context, then returns an approve/flag/reject verdict with a confidence score and written audit narrative. Flagged claims revert to admin review. Every verdict is written back to mem0 to inform future reviews.

Layer 5 — Referral Pipeline

Graph analysis maps referral relationships and identifies circular referral patterns, mutual-referral schemes, and coordinated sign-up networks. Detected rings are flagged and excluded from referral rewards. Ring signals are written to mem0 and propagated to the payout worker — if a wallet is flagged for ring activity, its pending payouts are blocked.

Layer 6 — mem0 Cross-Pipeline Intelligence

Every pipeline writes signals to mem0 persistent memory: submission verdicts, payout results, engagement anomalies, referral rings, handle changes, and auto-bans. This creates a longitudinal profile for each wallet and X account. The payout worker checks mem0 flags before dispatching SOL — a wallet flagged by the referral pipeline after a claim was auto-approved will be caught here. Claude receives the full mem0 entity profile during review, giving it awareness of behavior across all pipelines. A daily intelligence worker synthesizes cross-pipeline patterns and refreshes Redis flag caches.

Layer 7 — Institutional Knowledge Base

Gate rules, fraud patterns, policy thresholds, and decision history are stored in a knowledge base synced from operational documentation. Claude's review prompt dynamically pulls relevant entries based on the claim's risk profile — high-risk claims get fraud pattern context, failed gates get rule explanations, referral flags get ring detection rules. Weekly intelligence reports generated from cross-pipeline analysis are stored back into the knowledge base, creating an ever-growing institutional memory.