pot-sdk

ThoughtProof SDK — Multi-model verification for AI outputs

The TypeScript/JavaScript SDK for the ThoughtProof Protocol. Run structured multi-model verification on any AI output — adversarially, transparently, and provider-neutral.

Why?

Single models hallucinate, agree with themselves, and miss their own blind spots. PoT routes your claim through multiple competing models (generators + critics) and synthesizes a confidence-scored epistemic block.

Signing proves WHO. Multi-model adversarial verification proves WHAT.

Install

npm install pot-sdk

Quick Start

import { verify } from 'pot-sdk';

const result = await verify('GPT-4o claims the Eiffel Tower is 330m tall.', {
  providers: {
    anthropic: { apiKey: process.env.ANTHROPIC_API_KEY },
    xai:       { apiKey: process.env.XAI_API_KEY },
    deepseek:  { apiKey: process.env.DEEPSEEK_API_KEY },
    moonshot:  { apiKey: process.env.MOONSHOT_API_KEY },
  }
});

console.log(result.confidence);    // 0.94
console.log(result.synthesis);     // "Claim is accurate. The Eiffel Tower..."
console.log(result.mdi);           // 0.87 — Model Diversity Index
console.log(result.sas);           // 0.91 — Synthesis Audit Score

v0.5.0 — Community Release

New features driven by Moltbook community feedback:

Domain Profiles — auto-configure verification for your use case:

const result = await verify(output, {
  claim: 'Dosage is 500mg twice daily',
  domain: 'medical', // auto: adversarial + requireCitation + classifyObjections
});

Citation-Required Mode — ~40% fewer false positives:

const result = await verify(output, {
  claim: '...',
  requireCitation: true, // critic must quote exact text it objects to
});

Classified Objections — structured severity and type:

result.classifiedObjections
// [{ claim: "...", type: "factual", severity: "critical", explanation: "..." }]

Community Credits:

• @evil_robot_jas — Domain Profiles ("who configures it?")
• @SageVC — Objection Classification ("does the verifier understand what it's verifying?")
• @icyatrends — Citation Requirement ("the act of explanation is the check")
• @leelooassistant — Output Format ("human reviewers vs automated pipelines")
• @carbondialogue — Receptive Mode ("friction lives in the relationship")
• @ultrathink, @echo_0i, @thoth-ix, @ODEI, @Glyphseeker, @Dermez, @SB-1, @MarvinMSPN, @kaixinguo — ideas shaping future versions

API

verify(claim, options)

Run a standard verification (3 generators + 1 critic + 1 synthesizer).

deepAnalysis(claim, options)

Full deep run with rotated synthesizers — use for strategic decisions.

createAttestation(result)

Generate a tamper-evident JSON-LD audit trail block for compliance use cases.

pot.with_oversight(fn)

Human-in-the-loop hook for EU AI Act Art. 12-14 compliance.

Metrics

| Metric | What it measures | |--------|-----------------| | confidence | Overall verification confidence (0–1) | | mdi | Model Diversity Index — input-side diversity | | sas | Synthesis Audit Score — output fidelity to generator inputs |

BYOK

Bring your own API keys. pot-sdk never proxies your requests — everything runs directly from your environment to the model providers.

Supported Providers

Built-in: Anthropic, OpenAI, xAI, DeepSeek, Moonshot

Any OpenAI-compatible endpoint works via baseUrl (Ollama, Together.ai, custom deployments). BYOK — no keys bundled, everything runs on your infrastructure.

Learn More

• Protocol Specification
• pot-cli — CLI version
• Benchmarks — 96.7% adversarial detection, 92% hallucination detection

Built with the ThoughtProof Protocol. MIT License.