Haivian Technology publishes this guide and sells research audits. Haivian is included transparently in the examples below; this is not an independent ranking or a claim that one provider is best. Provider descriptions were checked on 17 October 2026 and should be confirmed before purchase.
Define the result before comparing providers
Write down the claim, its version and the decision that depends on it. “Audit my paper” may mean checking a theorem, reproducing a figure, tracing references or assessing whether the supplied evidence supports a scientific conclusion. Those tasks need different materials and may reach different outcomes.
Ask the provider to identify what is included and excluded, the accepted inputs, the decisive checks and what counts as a meaningful difference from the submitted result. Agree how it will report missing data, execution failure, disagreement and inconclusive evidence. A service that promises validation before seeing the materials has not yet established what can be tested.
If AI helped create the work, make that provenance visible where relevant. Auditing AI-generated research outputs is not the same as evaluating an AI model for safety, security, bias or regulatory compliance.
Compare the kind of check you need
| Type of work | What to ask | What the outcome can show | Important boundary |
|---|---|---|---|
| Static code and artifact readiness | Will you inspect files, dependencies and paper/code correspondence without executing them? | Documented availability and readiness issues | Readiness is not observed reproduction. |
| Computational reproducibility | Which workflow, inputs and outputs will be run and compared? | Observed results in the recorded environment | Rerunning supplied data is not a new experimental replication. |
| Mathematical or scientific reasoning review | Which assumptions and load-bearing arguments will be challenged? | Scoped support, gaps, counterexamples and uncertainty | Reasoning review is not automatically formal proof checking. |
| Formal proof checking | Which exact theorem, definitions, axioms and checker will be used? | Acceptance of the encoded statement within its trust base | A checked statement or sublemma may not capture the entire manuscript. |
| Journal peer review or specialist consultation | Who reviews the work, and what checks are commissioned? | The assessment the venue or expert actually performs | Peer review does not necessarily include code execution or formal verification. |
An engagement can combine these methods. Its report should still state which conclusions came from which checks.
Ask for a public example and read its limits
A sample report is more useful than a general promise of rigour. Look for an exact target, source versions, the method actually performed, evidence of execution where claimed and an explicit conclusion. Check whether limitations appear beside the result or only in a distant disclaimer.
Haivian's Sidorenko audit separates a checked finite-host statement from a further consequence outside the supplied formal comparison. The Borsuk audit distinguishes supplied and audit-authored formal targets. Those boundaries help a buyer understand what a formal result does—and does not—cover.
The Seymour and Boolean-functions records likewise have specific theorem scope and trust limits. They are examples of completed work, not a guarantee that another claim will be supported.
For any provider, distinguish a published method from a completed case. A promise to execute code needs an observed execution record for the engagement in question. A hash identifies bytes; it does not establish that a scientific conclusion is true.
Examples of different approaches
The following examples are not interchangeable offers, and their inclusion is not an endorsement. The linked organizations describe their own services; this guide has not audited their delivery.
CODECHECK describes a process for independent execution of research computations, with auditable workflows and certificates of executable computation. Its public site and register let a buyer inspect the process and examples. Ask which computations a proposed check covers and how any modifications are recorded. A standard commercial price and turnaround were not verified for this guide.
RepoReady describes checks of material availability, rerun readiness and paper/code agreement. Its current explanation states that those checks do not run the code yet. Individual access uses the researcher's OpenRouter key with model usage billed separately; institutional plans differ. Compare that readiness service with executed reproduction only after accounting for the difference in method and total cost.
Verified Research Lab describes bounded testing of computer-systems and software research, including reconstruction, reruns and versioned reports. Its service explanation states the scope and engagement process. Confirm the particular reviewer, methods, price, timing and publication terms rather than infer them from the category name.
Lean is a proof assistant, not a commissioned research audit service. It can be part of formal checking, but somebody must still identify the right statement, inspect assumptions and explain how the formal result corresponds to the paper. Software availability does not price that work or provide a publication decision.
Haivian, the publisher of this guide, offers defined research audits and publishes scoped records such as the examples above. Its current pricing describes a US$1,499 initial Research Audit and a separate US$1,000 publication-audit stage after eligibility. Scope changes, additional compute and specialist work require separate agreement. These are not prices for unlimited checking, a favourable finding or guaranteed publication. See the Council method and service boundaries.
A missing detail in these examples means it was not verified, not that the provider lacks the capability. Recheck current offers, service versions and terms when requesting a proposal.
Identify who is accountable and what independent means
Ask who contracts for the work, who performs each review step and who is responsible for the delivered report. Human accountability, subject expertise and external human peer review are separate questions. The presence of a responsible person does not establish that a human specialist personally checked every proof or calculation.
A third-party audit for AI research, science or mathematics should disclose the research producer, commissioning party, financial interests and organizational relationships. Separate review contexts may provide useful procedural separation, but they do not make common ownership disappear. If AI agents contribute, ask whether they use the same provider, what they saw before their assessments and how disagreements were preserved.
Haivian's linked mathematics reports disclose same-provider AI-agent review and retained toolchain trust. Haivian's audit service and journals share organizational control. These relationships matter when comparing commissioned scrutiny with an external journal's review process; neither should be described as the other.
Compare scope and price together
Request the written deliverables and conditions before comparing numbers. A static artifact check and an executed workflow are different purchases. A short conventional proof review and a new formalization can require very different work.
Clarify whether the price includes environment reconstruction, data acquisition, specialist review, compute, correction rounds and final presentation. Ask how a missing dependency or infeasible calculation affects the deliverable and fee. Confirm taxes, currency and cancellation terms. Do not treat an unspecified turnaround as immediate delivery.
For Haivian, the pricing page is the source of current offers. Additional work requires agreement; a purchase does not expand the accepted scope silently. Use the supported enquiry and intake process rather than assume a general purchase link proves every private-material workflow is available.
Agree confidentiality and publication separately
Before sharing unpublished work, confirm the access route, permitted handling, retention terms, rights and who can see the materials. Do not send passwords or private repository tokens through a general enquiry. Repository invitations must name an agreed recipient, not an assumed account.
Ask whether the report can remain private and what approval is needed before any text, figure, code or evidence is published. An optional journal pathway is not a publication guarantee. Where audit and publishing share control, that relationship should be explicit, along with scientific and editorial acceptance criteria.
Haivian's author guidance describes rights, disclosures, exact-text approval and correction/version expectations. It is not a private upload channel, and the journals do not advertise guaranteed indexing. Read the actual engagement terms before relying on any confidentiality or publication arrangement.
A checklist for your shortlist
- What exact claim and version will be tested?
- Which materials and permissions are required?
- Will code be executed, reconstructed or only inspected?
- Which mathematical reasoning and formal endpoints are included?
- What results would challenge the claim, and how will uncertainty be reported?
- Who is accountable, who performs the checks and where is AI used?
- What commissioning, financial or ownership relationships affect independence?
- What public example demonstrates the method within an honest scope?
- What is included in the price, and what needs a separate quote?
- What confidentiality, publication and correction terms apply?
Use the answers to choose the appropriate method, not merely the most confident marketing language. A credible audit can support a result, qualify it or find it unsupported. Its value lies in showing what the evidence warrants.
To compare that checklist with Haivian's offer, review the research audit service or discuss a defined claim.
Authorship and update policy
Original guide author and publisher: Haivian Technology. Haivian sells the service described here and publishes its audit records under common control with Qonduit Limited, New Zealand. This is AI-assisted company editorial work, not an external provider ranking or independent human expert assessment of the named services.
The public primary sources above support the specific service descriptions, not a comparative quality score. Provider facts and prices are dated observations. Corrections should identify the affected claim and version; substantive updates should carry a real update date rather than silently replacing a prior description.