Start with the claim you need examined. We agree the exact version, available materials, methods and limits before accepting the work. The conclusion may be supported, supported with qualifications, unsupported, contradicted within scope or inconclusive. Payment funds the audit, not a favourable verdict.
Discuss your claim and materials or review current research audit pricing.
If you are comparing approaches first, use our guide to choosing a research audit service.
What can be audited
AI may have helped produce a theorem, a simulation, a research manuscript or a numerical argument. Its involvement does not remove the need to check definitions, evidence and reasoning. It also does not make every artifact suitable for the same review.
- Mathematical proofs: theorem identity, definitions, assumptions, proof steps, counterexamples and formal artifacts where available.
- Computational results: the declared environment, inputs, executable workflow, output comparisons and numerical limitations within an agreed scope.
- Scientific claims: whether the stated model, analysis and available evidence support the defined conclusion, subject to domain fit and feasible checks.
- Research provenance and references: exact source versions, supporting citations, attribution and the boundary between an established result and a novelty claim.
A third-party audit for AI research should state what makes it third-party. For research produced outside Haivian's organizational control, the commissioning and review relationships still need disclosure. Work produced under the same control must not be marketed as externally independent merely because separate review steps were used.
Choose the method that answers the question
Reading a proof, running a calculation and checking a formal theorem are different activities. An engagement may combine them, but the report should identify each one rather than collapse them into a single validation label.
For a mathematical submission, conventional review examines the reasoning and its assumptions. If compatible formal artifacts are available and accepted into scope, formal checking also needs the exact theorem statement, definitions, permitted axioms, checker and correspondence with the manuscript. A successful kernel check of a sublemma does not establish the entire paper. Explore mathematical proof verification.
For research code, inspecting files for readiness does not show that the result reproduced. Executed reproducibility records what ran, with which inputs and environment, and how the outputs compared. Replication with newly collected data is a different undertaking. Explore computational reproducibility audits.
Scientific conclusions can also depend on observational uncertainty, experimental access or specialist knowledge. The fit assessment should identify those dependencies rather than promise universal verification from a manuscript alone.
What you receive
The agreed audit produces a bounded report that identifies the target version and claim, materials examined, methods actually performed, findings, material disagreements, limitations and remaining uncertainty. Where execution forms part of the accepted scope, the report distinguishes observed results from source inspection or proposed checks.
Haivian's Council method describes how competing assessments, evidence and dissent inform the recommendation. The report remains accountable to its evidence; agreement between AI agents is not itself a proof. Human responsibility for the engagement must not be confused with external human expert peer review.
If a calculation cannot run safely, a dependency is missing or the available data cannot settle the question, those limits belong in the outcome. They do not refute the research. Equally, a completed audit may identify an unsupported claim rather than validate it.
Public examples of exact scope
Haivian publishes original, scoped audit reports that readers can inspect before commissioning work. These examples concern results released in OpenAI/math; they do not imply OpenAI endorsement, shared ownership or a verified account of how a model discovered a result.
The Sidorenko counterexample audit reports formal checking of a finite-host statement. It distinguishes that endpoint from a forcing consequence outside the supplied formal comparison. Its execution history also separates retained work from newly built modules.
The Borsuk audit distinguishes the supplied formal target from a separately identified audit-authored Euclidean target and from conventional mathematical consequences. It does not claim a minimal dimension or a finite point-cloud witness.
The Seymour audit and Boolean-functions audit provide further examples with their own theorem boundaries, assumptions and trust disclosures. A result from one case is evidence about that case, not a guarantee that a new submission will pass.
What the service does not promise
An audit does not automatically establish whole-paper correctness, mathematical novelty, experimental replication, hardware advantage, regulatory compliance or journal acceptance. A new-looking result also does not prove that a model worked unaided, avoided training exposure or used no previous solution. Those are separate provenance questions requiring their own evidence.
Formal checking depends on the encoded statement and its trust base. Numerical checks depend on the stated model, arithmetic, data and environment. Reports should disclose retained tools or caches and any untested extensions rather than imply a clean-room rebuild or a second independent kernel when neither occurred.
Scope and pricing
Haivian's published initial Research Audit price is US$1,499 for one defined review. The accepted scope determines which methods apply. Additional compute, specialist work, remediation or material scope changes require a separate quote and explicit approval.
Publication is optional and not guaranteed. The currently described complete publication audit is a separate US$1,000 conditional stage after eligibility is established. Buying the first audit does not buy publication or a positive conclusion. Check current prices and conditions before commissioning work.
Preparing a submission
Begin with the manuscript or artifact version, the precise claim, your intended use and the evidence available to test it. Where relevant, include code and data descriptions, environment requirements, reference information and any access or licensing constraints.
Private material should use only the protected intake route and handling terms agreed for the engagement. Confirm authorized repository access before granting it; do not send passwords, API keys or repository tokens through a general enquiry. A public service page is not itself a private upload channel.
Discuss a scoped research audit. We will clarify whether the question, available evidence and requested methods fit before the engagement proceeds.
Review and publishing relationships
Haivian Technology provides the audit service and publishes its journals under common organizational control with Qonduit Limited, New Zealand. Richard Janutka is the named commissioning owner in the linked public examples. Haivian discloses material AI assistance, commissioning and review relationships in its records.
The linked mathematics audits used distinct-agent review from the same AI provider and disclose their retained toolchain and kernel limits. They are not presented as external human peer review, institutional accreditation or endorsement by the original authors. Publication requires the applicable scientific, editorial, rights, disclosure and exact-text approval; a customer may keep an audit private under the agreed terms.
For the broader process, see research audits, the Council method and current author guidance.