NumProof

NumProof

The deterministic numeric truth layer for AI agents and spreadsheets.

Your agent writes *"gross margin improved from 42.1% to 44.8%"* or *"the workbook is internally consistent"* — NumProof tells you, deterministically, whether that number is VERIFY, REFUTE, or ABSTAIN, with a counterexample, cell/formula provenance, and a signed, machine-checkable audit bundle. It's exact arithmetic and symbolic math — not an LLM judging another LLM.

• ✅ Verify a single math/finance claim, or batch thousands in CI
• ✅ Audit xlsx/csv rows: footing, cross-footing, balance-sheet ties, margins, formula cells — with provenance
• ✅ Diff two report versions; evaluate covenant rule packs (DSCR, Debt/EBITDA, current ratio, …)
• ✅ Signed evidence bundle (JSON + HTML/PDF/ZIP) anyone can re-verify offline
• ✅ API · CLI · MCP server · optional x402 pay-per-call

This repo is the open-source client (SDK + MCP). The verification engine runs as a hosted

service — pip install numproof, point it at the API, done. (Same shape as stripe-python:

the SDK is open, the engine is the service.)

Live demo (no key): https://numproof.com · Docs: https://numproof.com/docs

---

30-second start

pip install numproof

from numproof import NumProof

np = NumProof.from_env() # NUMPROOF_API_KEY (get a free key: see below)

print(np.verify("120 + 90 + 340 + 15 == 565"))

{'verdict': 'VERIFY', 'certificate': 'EXACT_ARITHMETIC', ...}

print(np.verify("a 50% loss needs a 100% gain to break even")) # VERIFY: (1-0.5)*(1+1.0)==1 print(np.verify("two 10% raises equal a 21% total increase")) # VERIFY print(np.verify("operating margin is 18% when EBIT is 180 and revenue is 1000")) # VERIFY

No install? It's just HTTP:

curl -s https://numproof.com/demo -H 'Content-Type: application/json' \

-d '{"claim":"gross margin is 60% when gross profit is 600 and revenue is 1000"}'

Free API key:

curl -s https://numproof.com/signup -X POST -H 'Content-Type: application/json' -d '{}'

---

Audit a spreadsheet (with provenance)

rows = [["Revenue", 1000], ["COGS", 400], ["Gross Profit", 600], ["Gross Margin", "60%"]]

print(np.audit_rows(rows)["verdict"]) # PASS (600/1000 == 60%, footing, ties, ...)

covenant rule packs: DSCR, Debt/EBITDA, current ratio, custom thresholds

print(np.covenant_rows( [["EBITDA", 500], ["Debt Service", 300], ["Debt", 1200]], rule_pack="credit_covenants_basic", )["verdict"])

Every audit/diff/covenant result can be returned as a signed bundle + human-readable HTML/PDF report (format="zip"). Recipients verify it without trusting you or NumProof:

curl -s https://numproof.com/audit/verify -H 'Content-Type: application/json' -d @bundle.json

{"valid": true, "verdict": "PASS", "signer": "0x...", ...}

---

Use it from an AI agent (MCP)

NumProof ships an MCP server so Claude / OpenAI-style agents can call it as a tool — gate every numeric claim before it reaches a user, report, or auditor.

python -m numproof.mcp

{ "mcpServers": { "numproof": { "command": "python", "args": ["-m", "numproof.mcp"] } } }

Or point any MCP client at the hosted descriptor: https://numproof.com/mcp.json.

See examples/ for runnable scripts (verify, audit, covenants, agent-gate, MCP).

---

Drop-in guardrails for agent frameworks

Verify the numbers your agent emits before it acts, in the framework you already use (numproof/integrations/ — each lazily imports its framework, so the numproof client stays stdlib-only):

# OpenAI Agents SDK — output guardrail that trips on REFUTE

from agents import Agent from numproof.integrations.openai_agents import numproof_output_guardrail agent = Agent(name="...", instructions="...", output_guardrails=[numproof_output_guardrail()])

• OpenAI Agents SDK — numproof.integrations.openai_agents (output guardrail / tripwire)
• Pydantic AI — numproof.integrations.pydantic_ai (output validator; raises ModelRetry with the counterexample so the model self-corrects)
• LangChain — numproof.integrations.langchain (a NumProof Tool + an output checker)
• DeepEval (Confident AI) — numproof.integrations.deepeval (a deterministic NumProofMetric: VERIFY → score 1.0, REFUTE → score 0.0 with the counterexample on .reason)
• Guardrails AI — numproof.integrations.guardrails (a Hub Validator: REFUTE → FailResult with the counterexample, so your on_fail action — reask/fix/exception — fires)

VERIFY → pass · REFUTE → block/retry with the counterexample · ABSTAIN → pass-through (configurable). Runnable examples in examples/; install only the framework you use.

Independently re-checkable receipts (the part you can't fake)

Any verdict can be returned as a signed Verification Receipt — and you re-check it offline, trusting neither the transport nor NumProof:

pip install "numproof[verify]"

numproof-verify receipt.json --signer 0x

OK independently re-derived + signature valid

It recovers the EIP-191 signer (tamper-evident) and, for value/agg/identity/sequence claims, independently re-derives the verdict with stdlib Fraction + sympy. A tampered field, a wrong signer, or a verdict that doesn't actually hold all fail loudly — even a receipt NumProof itself mis-signed is caught by the re-derivation. An agent can recompute a number for itself; it cannot issue an independent, signed attestation a *second party* will accept. That independence — not the arithmetic — is the product. Format + spec: RECEIPT_FORMAT.md.

---

Why deterministic (and why it matters)

Generic "AI guardrails" use a model to grade a model — probabilistic, and itself can hallucinate. NumProof recomputes the math exactly (rational arithmetic + symbolic identity checking) and returns a reproducible verdict with a trace. When it can't prove something it says ABSTAIN rather than guess. For finance, regulated, and agent workflows, *"the number is provably right"* beats *"another model thinks it looks right."* Full table: comparison.md.

---

Pricing

---

What's in this repo

The numproof Python SDK (NumProof client), the MCP server, and runnable examples — all thin HTTP clients to the hosted API. MIT licensed. The verification engine, finance audit logic, formal (Lean) proof tier, and signing are the hosted service and are not in this repo.

Found a wrong verdict? Open an issue with the exact claim — correctness is the whole product.