thoughtproof-verify

thoughtproof-verify

Tier: 2 (pre-trade verification — runs before any order execution)
Trigger: Wired into place_order dispatch in server.py
MCP Tool: run_algoclaw_skill("thoughtproof-verify", {"claim":"Place MNQ long 2 @ market","stake":"high"})
Source Pattern: ThoughtProof/thoughtproof-mcp (cross-model adversarial attestation)

What It Does

Before executing any order, routes the trade claim through 2–5 adversarial AI models for independent review. Each model acts as a critic and returns ALLOW / BLOCK / UNCERTAIN with a signed attestation.

Models used:

• Fast (2 models, ~$0.008): Claude Haiku + GPT-4o-mini
• Standard (3 models, ~$0.03): + Gemini Flash
• Deep (5 models, ~$0.08): + Grok + DeepSeek

Config

THOUGHTPROOF_ENABLED = os.environ.get("THOUGHTPROOF_ENABLED", "false") == "true"
THOUGHTPROOF_MIN_STAKE = os.environ.get("THOUGHTPROOF_MIN_STAKE", "high")  # low/medium/high/critical
THOUGHTPROOF_SPEED = os.environ.get("THOUGHTPROOF_SPEED", "fast")  # fast/standard/deep

Decision Logic

Verdict	Outcome
All ALLOW	Execute trade
Any BLOCK	Refuse trade, log reason, alert
2+ UNCERTAIN	Escalate to owner, hold trade

Output

{
  "claim": "Place MNQ long 2 contracts at market — AI signal 87.1%",
  "verdict": "ALLOW",
  "models_consulted": 2,
  "speed": "fast",
  "cost_usd": 0.008,
  "attestation_id": "tp_a1b2c3d4",
  "reasons": ["Signal consistent with regime", "Risk/reward ratio acceptable"],
  "blocked": false
}

Cost Justification

With $500/day max loss limit, $0.008 per order pre-verification costs < 0.01% of daily risk budget. Blocking even one bad $200 loss covers 25,000 order verifications.