thoughtproof_skill

Overview

This skill provides epistemic verification for AI agent outputs by running multi-model checks and producing auditable consensus reports. It detects blind spots, ranks dissent, and produces a confidence-weighted synthesis so teams can make decisions with an evidence trail. The tool is BYOK and stores verification blocks locally for later review.

How this skill works

ThoughtProof normalizes the prompt, generates independent proposals from three or more different model families, runs adversarial critiques to surface contradictions and blind spots, then evaluates and synthesizes a consensus with confidence and a Model Diversity Index. Each run produces an Epistemic Block containing proposals, critiques, synthesis, and metadata that you can list or inspect. Commands like tp verify, tp deep, tp list, and tp show let you run the pipeline, run deeper multi-pass audits, and review archived blocks.

When to use it

• High-stakes decisions where an audit trail and independent reasoning are required (investment, compliance, legal).
• When you need a second opinion or suspect a single model is biased or overconfident.
• To build consensus across multiple model families before committing to a recommendation.
• When governance requires documented reasoning and dissent tracking for future audits.

Best practices

• Provide diverse API keys (OpenAI, Anthropic, xAI, Moonshot, etc.) to maximize model family diversity.
• Use the Standard or Deep tiers for business-critical or regulatory decisions to increase robustness.
• Chain context via tp verify --context last when building on prior verifications to preserve continuity.
• Store and lock Epistemic Blocks as part of your decision records for reproducibility and audits.

Example use cases

• Verifying an investment thesis with tp deep to rotate roles and run meta-synthesis across models.
• Auditing a compliance decision and exporting Epistemic Blocks for regulators or internal reviewers.
• Running quick sanity checks on architecture choices (microservices vs monolith) using tp verify.
• Detecting blind spots before product launches by comparing independent model proposals and critiques.