Problem
AI agents using OKX Trade Kit operate autonomously with access to 164 trading tools. Current safety controls (--read-only flag, per-module filtering, rate limiter) gate execution permissions but do not verify reasoning quality. When the agent is allowed to trade, nobody checks whether its thesis is sound before the order goes out.
In live experiments, AI trading agents fabricate data in 22% of trade proposals — inventing price levels that don't exist, confusing one asset's data with another's, or citing volume figures that contradict the actual feed. This is consistent across GPT-5.4, Gemini 3.5 Flash, Claude Opus 4.8, and Kimi K2.6.
Proposed Solution
A new verify module with a verify_trade_decision MCP tool that validates agent reasoning before order execution.
How it works
Agent proposes trade → verify_trade_decision → Sentinel API (~3s) → ALLOW / BLOCK / UNCERTAIN
- ALLOW: Reasoning is sound, proceed with execution
- BLOCK: Critical flaws detected, reject
- UNCERTAIN: Issues found, return structured objections for agent re-planning
Single HTTPS call to sentinel.thoughtproof.ai/sentinel/verify. No SDK, no library dependency, fail-closed on error.
Input/Output
// Input
{
action: "BUY BTC-USDT $500",
thesis: "BTC showing bullish divergence...",
reasoning: "RSI at 28.5, volume up 15%...",
marketContext?: "optional market data"
}
// Output
{
verdict: "UNCERTAIN",
confidence: 0.67,
objections: [
{ severity: "high", explanation: "Claimed RSI of 28.5 not found in provided market data" }
]
}Re-Plan Loop (key differentiator)
On UNCERTAIN, the agent receives specific, actionable objections — not just "rejected." In live experiments, agents self-corrected in 52% of cases: standing down, revising size, or rewriting their thesis around actual evidence.
Live Experiment Data
489 trading cycles, 7 days, real money on Coinbase Spot:
- Unverified agent: -44% ($1,035 → $563)
- Sentinel-verified agent: -9% ($1,035 → $937)
- 22% outright fabrication (invented numbers not in any data source)
- Currently also running on Binance Testnet — same patterns, confirming this is model-level, not platform-specific
Proof of Concept
We have a working TypeScript implementation of the verify module: verify.ts (~150 lines, follows existing module patterns).
Full design document with architecture details available at the same repo.
Cost
- <$0.01 per verification
- ~3 second response time
- No infrastructure required — single HTTPS call
Links
Happy to contribute a PR if the maintainers think this fits the module architecture. The verify module is designed to complement existing safety controls (read-only, rate limiting) rather than replace them.
Problem
AI agents using OKX Trade Kit operate autonomously with access to 164 trading tools. Current safety controls (
--read-onlyflag, per-module filtering, rate limiter) gate execution permissions but do not verify reasoning quality. When the agent is allowed to trade, nobody checks whether its thesis is sound before the order goes out.In live experiments, AI trading agents fabricate data in 22% of trade proposals — inventing price levels that don't exist, confusing one asset's data with another's, or citing volume figures that contradict the actual feed. This is consistent across GPT-5.4, Gemini 3.5 Flash, Claude Opus 4.8, and Kimi K2.6.
Proposed Solution
A new
verifymodule with averify_trade_decisionMCP tool that validates agent reasoning before order execution.How it works
Single HTTPS call to
sentinel.thoughtproof.ai/sentinel/verify. No SDK, no library dependency, fail-closed on error.Input/Output
Re-Plan Loop (key differentiator)
On UNCERTAIN, the agent receives specific, actionable objections — not just "rejected." In live experiments, agents self-corrected in 52% of cases: standing down, revising size, or rewriting their thesis around actual evidence.
Live Experiment Data
489 trading cycles, 7 days, real money on Coinbase Spot:
Proof of Concept
We have a working TypeScript implementation of the verify module: verify.ts (~150 lines, follows existing module patterns).
Full design document with architecture details available at the same repo.
Cost
Links
Happy to contribute a PR if the maintainers think this fits the module architecture. The verify module is designed to complement existing safety controls (read-only, rate limiting) rather than replace them.