VerdifaX

Use cases

Built for AI governance.
Designed for any high-stakes computation.

Wherever a computational decision must be reproducible, tamper-evident, and independently verifiable, Verdifax is the audit infrastructure layer.

Live today

Demonstrated end-to-end today

Three use cases where the Verdifax runtime, audit bundle, and verifier are operational and observable , AI/ML model governance running live against the Anthropic Claude API, trade lifecycle batch attestation through a production batch endpoint, and reproducible research with shipped Python + R SDKs.

Live

Primary use case

AI / ML model governance

Every governed AI decision produces a sealed audit artifact anchored on a public log: from run 238 on (2026-10-16, 17:06 UTC) the complete record of a successful run, before then the run's identity and (from leaf v5) its 33 pre-anchor values, so a later change to what is anchored no longer matches the public log; from run 239 on, deleting the complete-record anchor also fails verification. Since verifier 0.12.0, a public-log entry counts only if it is signed with a published Verdifax log-signing key and the timestamp the log signs for it verifies, so a rewritten record that anyone else logs as a new entry fails. Verdifax holds that key. From run 240 on, each record seals a 600-second anchoring window, so a later re-log, even by Verdifax, fails or must carry a new sealed date; earlier records show their authenticated anchoring time but cannot fail on it. AIVP (AI Verification Protocol) runs live against the Anthropic Claude API: when a governed run includes AI output text, that text is sent to the Claude API for a governance review, and a hash of the text and the review's PIA hash are sealed into the manifest. Compliance teams can give the artifact to auditors and regulators as a record of what ran; legal teams can offer it as evidence, with admissibility decided by the court; engineering teams use the same record to debug and reproduce decisions.

NIST AI RMFISO/IEC 42001EU AI Act Art. 12EU AI Act Art. 13HIPAA
Live

Financial services

Trade lifecycle audit & reconciliation

Daily portfolio reconciliation, end-of-day risk calculations, derivatives mark-to-market, monthly hedge accounting, every slow-trading workflow that produces an auditable record. POST /execute/batch seals up to 1,000 rows per call end to end against production, with per-row error handling and the PEPG, CCV, MACC and AIVP halt receipts preserved per row. Real-time / microsecond HFT remains a specialist build, see the roadmap for in-process attestation.

MiFID IIBasel IVDodd-FrankCFTC
Live

Research integrity

Reproducible computational research

Verdifax records the declared research environment, runtime version, pinned dependencies, git SHA, declared random seeds, platform, in the audit bundle as Category 6, sealed by its section hash and, on successful runs sealed from run 238 on, covered by the complete-record anchor on the public log. POST /execute/verify-determinism runs the same payload twice under a pinned clock and a shared nonce, with the public-log stage replaced by an in-process ledger, and reports whether the two manifest hashes match; production records are checked by recomputation, not by replay. Python (pip install verdifax) and R (remotes::install_github("Verdifax/verdifax-sdk-r")) wrapper SDKs ship today, with a Jupyter notebook and an R Markdown template as reference workflows.

NIH data sharingFDA SaMDEU Open Science

Capability available, rollout pending

POC-ready today. Production customer rollout pending.

Verdifax's architecture supports these use cases on a defined POC scope today, but no production customer has rolled out yet. Each card states what we can deliver now and what's deferred to the V2 trajectory, see the roadmap for full enterprise build-out.

POC ready

Financial services

Lending & credit decisions

Adverse-action notices, model-risk reviews, and fair-lending audits all require explainable, reproducible decisioning. Verdifax's architecture can bind every credit decision to hashes of its inputs, the declared model version, and the policy in effect when it was made. Verdifax does not explain a model's decision; it records what was decided and under which policy. The customer keeps calling its existing credit model and submits each decision to /execute after the model call; Verdifax produces the audit-of-record without receiving model weights. POC scope deliverable today; production lending customer rollout pending.

ECOAReg BFair Housing ActSR 26-2
POC ready

Insurance

Insurance underwriting

Rate-setting, eligibility, and claims-decisioning models face the same model-risk-management scrutiny as bank credit models. Verdifax's architecture supports underwriting attestation artifacts that auditors, state insurance commissioners, and policyholders could independently verify, without exposing model weights or proprietary feature engineering. Same plug-in deployment shape as lending. POC scope deliverable today; production insurance customer rollout pending.

NAIC Model BulletinState DOI rulesGDPR
POC ready

Sustainability

ESG carbon attestation

Carbon emission calculations are deterministic when emission-factor tables are pinned. Verdifax can attest individual Scope 1 (stationary combustion, fugitive emissions) and Scope 2 (purchased electricity, location-based) calculations today when the caller submits the calculation with the pinned GHG Protocol factor table it used; Verdifax does not ship an emission-factor library. Full enterprise CSRD reporting across multi-tier supply chains requires LCA database integrations that are on the V2 roadmap.

GHG ProtocolTCFD

What this would have caught

Public-record incidents, and what a governed path would have recorded.

The cases below failed in different ways, and the honest reading matters: in some, the question was what an AI system actually said or decided, and a sealed record lets any later copy of the exchange be checked against hashes sealed at the time. In others, the answer itself was wrong, and what matters is a check that can halt it before anyone relies on it, with the sealed record then showing which controls ran and what they decided. Each card says which kind it is.

February 2024

Moffatt v. Air Canada

Civil Resolution Tribunal of British Columbia

Air Canada's chatbot told a grieving passenger he could claim a bereavement-fare refund retroactively. He bought the ticket, applied, was denied. The airline argued the chatbot was a separate legal entity responsible for its own words. What the chatbot said was not disputed. The tribunal disagreed with the airline and held Air Canada responsible for the chatbot's answer.

What changes with Verdifax · Prevention first

This is a prevention case, stated honestly: what the chatbot said was not disputed, so a record of the answer alone would not have changed the finding. What a governed path adds is a check before the answer is relied on: when a run includes AI output text, the AIVP stage sends it to a governance review, and a review decision other than release halts the run with a sealed halt receipt. The sealed record then binds a hash of the answer, the caller-declared model version and prompt hash, and the PEPG policy decision for that run, so it shows which controls ran and what they decided.

AIVPPEPGAER

June 2023

Mata v. Avianca

U.S. District Court, Southern District of New York

Attorney filed a brief with six fake case citations generated by ChatGPT. Sanctioned $5,000. Counsel claimed he had asked the model whether the cases were real and that it had confirmed. The record of what the model had said came from the lawyers' own later filings, and the conversation took place on a consumer chatbot outside any firm system.

What changes with Verdifax · Evidence first

Stated honestly: nothing helps a lawyer on a consumer chatbot. This case is what changes for firms that run AI inside a governed system instead. There, a Verdifax-attested run seals hashes of the prompt and the output, with the declared model version, at the time of drafting, so a copy of the exchange produced later can be checked against the sealed hashes: a copy matches only if it is the content that was hashed, apart from key order, whitespace, number spelling and string escaping (how a string's characters are escaped) in a JSON payload.

AIVPAER

November 2023

Estate of Lokken v. UnitedHealth

U.S. District Court, District of Minnesota

Class action alleging UnitedHealth used the nH Predict algorithm to deny Medicare Advantage post-acute care claims, with a 90% error rate alleged in the complaint. Which clinical inputs and which model version applied to each decision are the kind of facts such a case turns on.

What changes with Verdifax · Evidence first

Every recommendation under Verdifax would have sealed a hash of the clinical inputs submitted to it, the declared model version, a hash of the model's output, and the PEPG policy decision for that run. Each record could then be checked against its sealed hash; whether a decision followed medical guidelines would remain a question for the parties and the court.

DOGAIVPPEPG

Verdifax does not assert outcomes in these cases would have been different. The cited primitives, AIVP (AI Verification Protocol), DOG (Deterministic Oracle Gateway), PEPG (Pre-Execution Policy Gating), AER (Attestation Execution Record), all ship in v1 against the production orchestrator.

Beyond AI

Verdifax is infrastructure, not an AI tool.

The same cryptographic execution layer that audits an AI decision can audit any deterministic computation. If your workflow ends in a number, a label, a recommendation, or a ruling, and that output needs to survive a regulator, a court, or a peer-review challenge years later, Verdifax is built for it. The same primitive extends naturally to vehicle lifecycle attestation, industrial battery passports under EU Regulation 2023/1542, and any other VIN-anchored or asset-anchored provenance chain , specialist-roadmap territory but architecturally aligned.

Request access