thoughtproof

# ThoughtProof — Epistemic Verification Skill Multi-agent verification protocol for AI decisions. Like a TÜV for AI reasoning. ## How It Works ThoughtProof runs your question through multiple independent AI agents (different model families), then a critic layer identifies blind spots, and a synthesizer produces a consensus with confidence scores. **Pipeline:** Normalize → Generate (3+ models) → Critique (adversarial) → Evaluate → Synthesize ## Prerequisites - `pot-cli` installed: `npm install -g pot-cli` - At least one API key (Anthropic, OpenAI, xAI, or Moonshot) - More keys = more model diversity = better verification ## Quick Start ### Verify a claim or decision ```bash tp verify "Should we use microservices or monolith for our MVP?" ``` ### Chain context from previous verifications ```bash tp verify --context last "What about scaling considerations?" ``` ### Deep analysis with rotated roles ```bash tp deep "Is this investment thesis sound?" ``` ## Configuration pot-cli reads config from `~/.potrc.json`: ```json { "generators": [ { "provider": "xai", "model": "grok-4-1-fast" }, { "provider": "moonshot", "model": "kimi-k2.5" }, { "provider": "anthropic", "model": "claude-sonnet-4-6" } ], "critic": { "provider": "anthropic", "model": "claude-opus-4-6" }, "synthesizer": { "provider": "anthropic", "model": "claude-opus-4-6" } } ``` Show current config: `tp config` ### Model Diversity Requirement ThoughtProof enforces ≥3 different model families for generators. This is core to the protocol — no single provider can verify itself. ## Output Each verification produces an **Epistemic Block**: - **Proposals** from each generator (independent reasoning) - **Critique** identifying blind spots, contradictions, and risks - **Synthesis** with consensus score, confidence level, and dissent - **MDI** (Model Diversity Index) — measures independence of reasoning Blocks are stored locally as JSON and can be reviewed with `tp list` / `tp show <n>`. ## Commands | Command | Description | |---------|-------------| | `tp verify <question>` | Run full verification pipeline | | `tp verify --context last` | Chain from previous block | | `tp deep <question>` | Deep verify: multiple runs, rotated roles, meta-synthesis | | `tp list` | Show block history | | `tp show <n>` | Show a specific block | | `tp config` | Show current configuration | ## Tiers | Tier | Agents | Time | Best For | |------|--------|------|----------| | Light | 3 | ~30s | Quick sanity checks | | Standard | 5-7 | ~3min | Business decisions | | Deep | 7-12 | ~5min | High-stakes, regulatory | ## When to Use ThoughtProof - **High-stakes decisions** — investment, legal, medical, compliance - **Audit trail needed** — regulatory, governance, due diligence - **Blind spot detection** — when you suspect a single model is biased - **Cross-domain questions** — where no single model is expert ## When NOT to Use - Simple factual lookups (Google it) - Creative writing (subjective, no "correct" answer) - Time-sensitive queries under 30 seconds - Questions with trivially verifiable answers ## Architecture Note ThoughtProof is BYOK (Bring Your Own Key). Your API keys, your data, your models. Nothing routes through ThoughtProof servers. The skill is MIT-licensed; the consensus protocol is BSL-licensed. ## References - `references/block-format.md` — Epistemic Block JSON schema - `references/consensus-protocol.md` — How consensus is calculated