RWA Token Risk Evaluation & Anomaly Monitoring with Claude Agents SDK and Braintrust
Real-World Asset (RWA) tokenization requires AI-native security monitoring frameworks capable of detecting collateral drift, smart contract parameter manipulation, and fraudulent valuation metrics. By building an agentic audit runner with the Claude Agents SDK and evaluating workflow step performance via Braintrust, engineers can construct continuous compliance guardrails for tokenized assets.
What you are building
The core problem, expected build, and operating context for this challenge.
Develop an automated RWA collateral auditing agent with Claude Agents SDK and Braintrust to detect valuation anomalies and issue protocol risk alerts.
How work is evaluated
Evaluates the agent's accuracy in detecting artificial asset valuation inflation and triggering contract pause recommendations.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
emergency_trigger_accuracy
Agent must recommend TRIGGER_EMERGENCY_PAUSE when valuation diverges > 25%.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
oracle_divergence_detection_rate
Accuracy in detecting malicious oracle reports. • target: 0.98 • range: 0.9-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Build continuous collateral auditing workflows using Claude Agents SDK extended thinking
Log and evaluate agent tool invocations and decision reasoning in Braintrust
Implement multi-oracle cross-referencing to detect asset valuation manipulation
Generate automated smart contract emergency pause recommendations when risk thresholds breach
Reference links and supporting material
Dataset of 300 tokenized RWA asset pool states with injected oracle manipulation attempts, property appraisal discrepancies, and liquidity stress events.
Challenge setup
Evaluates the agent's accuracy in detecting artificial asset valuation inflation and triggering contract pause recommendations.
Challenge input
JSON containing pool_id, onchain_minted_tokens, independent_valuation_reports, and oracle_feeds.
Claude Agents SDK
Provides extended thinking for deep forensic collateral evaluation.
Braintrust
Traces agent step decisions and measures tool invocation performance.
Evaluated output
JSON with collateral_health_index, risk_flag, recommended_action, and braintrust_eval_score.
- Agent must recommend TRIGGER_EMERGENCY_PAUSE when valuation diverges > 25%.
- Accuracy in detecting malicious oracle reports. • target: 0.98 • range: 0.9-1
- Benchmark: DeFi Security & Audit Bench
- Oracle Divergence Detection Rate target: 98%
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Claude Code · Required
- Claude Code · Required
- Claude 4.1 Opus · Optional
- Claude 4 Sonnet · Optional
Evaluation contract
- emergency_trigger_accuracy · Weight 1
- oracle_divergence_detection_rate · Weight 1
Metrics
- oracle_divergence_detection_rate · Target 0.98
Benchmarks
- DeFi Security & Audit Bench
- security
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing pool_id, onchain_minted_tokens, independent_valuation_reports, and oracle_feeds.
Output: JSON with collateral_health_index, risk_flag, recommended_action, and braintrust_eval_score.
Execution harness: Python
Evaluation tools
- Claude Agents SDK · Autonomous Agent Engine
- Braintrust · Agent Evaluation & Observability
Recipe freeze marker
No recipe freeze time is recorded.