The design-partner program

What you get

Who it's for

Teams shipping AI over financial data:

• Agents and copilots that verify a claim before the user sees the answer, including over MCP, hosted or local.
• Fintech applications embedding a verification layer behind their own product surface.
• Research platforms publishing work their readers can check.

Phase-1 coverage: SEC filings and financial-statement claims, 17 verifiable metrics (revenue, net income, EPS, shares outstanding, cash, gross profit, operating income, operating cash flow, R&D expense, total assets, stockholders' equity (parent-only by published definition), capex, calculated EBITDA (a published GAAP-component formula that catches non-GAAP adjusted-figure substitution, never a silent approximation of it), and its three components as metrics in their own right: interest expense, income tax expense, and depreciation and amortization; and calculated free cash flow, operating cash flow minus capex under a published formula, never the company's own reported figure), annual and quarterly, verified against EDGAR XBRL as filed. Every verdict carries deterministic, machine-readable context notes where the record warrants them: the analyst-grade caveats (fiscal frames, balance-sheet dates, metric definitions) an agent can act on.

What Artizar promises, and what it doesn't

Verifiability, not infallibility. Artizar does not certify claims as "true," and no system honestly can. What it delivers is a record: the same evidence in produces the same verdict out, byte for byte, and every verdict, including every refusal, ships with what makes it checkable:

• a deterministic outcome computed by a published rule table, never an LLM vote;
• the evidence chain to the primary source as filed, with URL, accession number, and content hash;
• version stamps that make re-verification the same computation;
• an Ed25519-sealed record committed to a public, append-only, witness-cosigned transparency log, re-verifiable offline with published keys only;
• a privacy model where the public log carries only a blinded commitment, so nobody can see what you verified, or how often.

When Artizar can't verify, it refuses (typed, sealed, and on the record) rather than guess. That refusal discipline is measured, not asserted: see the numbers below.

The proof, not the pitch

FilingsCheck v1.3, the frontier edition: eight models from Anthropic and OpenAI asked for figures from SEC filings across all 17 metrics, including Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and GPT-5.5 at their vendor's top reasoning effort, every answer verified through the production API at api.artizar.ai. 2530 sealed grades, each one witness-cosigned in the public log, with adversarial probes grading the engine itself (2032/2032 as expected; zero false greens, zero false alarms). Each run states the request settings it ran under, and its bundle proves them. Fiscal years are read the way each filer labels them, and where a filer's own filings disagree about which year a label names, Artizar refuses on the record rather than guess. Every number on that page regenerates from sealed bundles a third party can verify with one command. The v1.2, v1.1 and v1.0.1 editions remain published, append-only.

The public keys, the verdict and record schemas, and the verification method are published; the engine is not open source. Trust rides on the record, not on reading our code.

Become a design partner

Tell us what you're building and what you'd need verified. Onboarding is a conversation, a key, and a quota, usually same-week.

No signup form, no pricing page. The program is free while we build with partners, and the eventual pricing conversation will start from real usage data, not a rate card.