pot-sdk-crewai 0.2.0

ThoughtProof Protocol — CrewAI integration for multi-model adversarial verification

pot-sdk-crewai

ThoughtProof Protocol — CrewAI integration for multi-model adversarial claim verification.

Install

pip install pot-sdk-crewai                  # core only (no crewai)
pip install "pot-sdk-crewai[crewai]"        # with crewai

Quick start

Core function (no crewai required)

from pot_sdk_crewai import verify_claim

result = verify_claim(
    claim="The Earth is approximately 4.5 billion years old.",
    models=["openai/gpt-4o", "anthropic/claude-sonnet-4-5", "deepseek/deepseek-chat"],
    mode="adversarial",
)

print(result.verdict)      # "TRUE"
print(result.confidence)   # 0.91
print(result.convergence)  # 1.0
print(result.dissent)      # [] — all models agreed

Verification modes

Mode	Posture
adversarial	Actively challenges the claim, assumes false until proven
resistant	Requires strong evidence, resists plausible-sounding narratives
calibrative	Balanced probability assessment, calibrated confidence

ThoughtProofVerifyTool (CrewAI tool)

from crewai import Agent
from pot_sdk_crewai import ThoughtProofVerifyTool

verify_tool = ThoughtProofVerifyTool(
    models=["openai/gpt-4o", "anthropic/claude-sonnet-4-5"],
    mode="adversarial",
    threshold=0.7,
)

analyst = Agent(
    role="Research Analyst",
    goal="Produce verified, accurate research reports.",
    tools=[verify_tool],
    llm="openai/gpt-4o",
)

The tool's _run output:

Verdict: TRUE | Confidence: 0.91 | Convergence: 1.00

With dissent:

Verdict: FALSE | Confidence: 0.85 | Convergence: 0.67 Dissent from: anthropic/claude-sonnet-4-5.

ThoughtProofGuardrail (CrewAI task guardrail)

from crewai import Agent, Task
from pot_sdk_crewai import ThoughtProofGuardrail

guardrail = ThoughtProofGuardrail(
    models=["openai/gpt-4o", "anthropic/claude-sonnet-4-5"],
    mode="calibrative",
    min_confidence=0.7,
    allow_uncertain=False,
)

task = Task(
    description="Summarise the key findings of the 2024 IPCC report.",
    agent=analyst,
    guardrail=guardrail.validate,   # CrewAI guardrail interface
)

guardrail.validate(task_output) returns:

• (True, VerificationResult) — accepted
• (False, "Verification failed: ...") — rejected with reason

VerificationResult fields

Field	Type	Description
verdict	"TRUE" | "FALSE" | "UNCERTAIN"	Majority verdict
confidence	float	Mean confidence of majority models
convergence	float	Fraction of models that agreed with majority
model_verdicts	list[ModelVerdict]	Per-model breakdown
dissent	list[ModelVerdict]	Models that disagreed (DPR)
reasoning	str	Concatenated reasoning from all models

Dependencies

Runtime: litellm>=1.0.0, pydantic>=2.0.0 Optional: crewai>=0.60.0 (for ThoughtProofVerifyTool)

License

MIT