Hiveram

For firms·For builders

Asking an AI assistant about Hiveram? Give it hiveram.com/docs/SKILL.md.

When work changes hands, what stays on record?

Can someone else reconstruct your AI-assisted work?

Plain answers for the people who answer to an auditor, a regulator or a client.

Not technical? Start here.

Hiveram connects recorded intent, execution evidence and review history so that someone who did not perform the work can examine it.

Discuss a first paid pilot

We run Hiveram on our own work and are seeking our first paid external pilot. Bring one AI-assisted workflow where collecting evidence for review is difficult.

We name the frameworks your auditors use so you can see where the evidence sits. Naming a standard is not a claim of conformance, and Obsta Labs holds no certification against any of them.

Six questions a partner asks first

For the partner of a professional firm

The partner of an audit practice or another professional firm asks these before any standard comes up. Each answer below names what is in place today and what is not. The six questions an auditor asks, and what the standards and laws require, follow further down.

When a staff member leaves, can we still show what they did on each engagement?

For the partner

For work that went through the record, yes. Each piece of work has a work order that says who asked for it, which identity took it on, what evidence was submitted, and who closed it; each change creates a new revision with its time, and notes, claims and the closure carry the identity that wrote them. When someone leaves, their lines stay, under their name. What the record holds nothing about is work that was never entered; it does not reconstruct the past.

A client, the regulator or the next partner asks who reviewed this working paper. Where is the answer?

For the partner

On the work order for that paper: the evidence entry, which is a reference to the file as the submitter wrote it (its location, a document number or a link, with a hash only if the submitter includes one), who submitted it and when, and who closed the work. Hiveram checks that the identity closing the work is not the identity that did it. It does not check that the closer read the paper carefully; a different closing identity is not proof of a careful review, and we do not describe it as one.

AI drafted part of the work. Who checked it, and can we show that?

For the partner

The AI writes its own lines in the record under the identity it was given, through the API, the command line or the MCP server, and it may submit evidence. The normal path to closing the work requires a different identity from the one that did it, and the record shows who closed it and when. Hiveram does not read the draft and does not judge its content; the check is a person's, and the record is where that check is written down.

Where does our data live, and what about PDPA?

For the partner

The record is a database: a SQLite file on one machine, or PostgreSQL on your own servers, where you hold it, or we can host an instance for your firm. The record holds references to your client files, not copies of the files. A page that sets out each way of running Hiveram, who can read the data, and how Singapore’s Personal Data Protection Act 2012 applies to each is being written with counsel; until it is published, ask us, and treat nothing on this page as legal advice.

What if we stop paying, or you disappear?

For the partner

When you run Hiveram yourself, the record is in your own database, and the API and command line that write and read it are publicly available. A read-only mode when a licence lapses, a complete export of the record as plain files, and written continuity terms are designed and not yet shipped; until they ship, they are not part of what you buy, and we do not sell them as if they were. What happens to an instance we host when payment stops is being written down; ask before you rely on it.

Does this replace SharePoint, Excel or our AI?

For the partner

No. Files stay in SharePoint or wherever you keep them; the record points to them. Your AI stays whatever you pay for; Hiveram does not run it and holds no credentials to it. The one thing it can stand in for is the spreadsheet that says who did what on each engagement, and only if you want it to.

Can our team write in their own languages?

For the partner

Yes. Work can be written in any language, and the record keeps the text exactly as written. Search finds words in languages that separate words with spaces; for Chinese, Japanese and Thai it currently matches whole phrases.

Our goal keeps changing while the work is under way. Does this help?

For project and delivery leads, patent practices, and anyone working toward a goal that keeps moving

A project lead learns the client changed the goal after half the work was done. A patent practice amends claims while earlier drafts, instructions and filings are still in circulation. Afterwards the questions are the same: what was the goal at each point, who changed it and when, what had already been done against the earlier version, and who accepted it.

For the work you enter, Hiveram keeps each version of the request with the time it changed, and keeps the work, the evidence references and the acceptance on the same item, each with its time; notes and the acceptance also carry the identity behind them. An earlier version stays readable after the goal moves.

What it does not do: it does not decide whether earlier work still counts after a change; a person does, with the history in front of them. It does not notice that a linked document has changed unless the reference names the version, so refer to drafts and filings by version or date. It does not calculate deadlines or replace a planning or docketing tool; it keeps the record beside them.

On our own work: over 30 days in our main project, about one task in nine had its goal rewritten after it was written down, about twice a day. The record kept each of those versions.

Why would we need this?

Answers 1–6

Most teams can now make AI do work. Far fewer can independently demonstrate what happened. When the only history lives inside the tool that performed the action, your evidence depends on the system you are trying to govern.

When an AI system sends a message, changes a record, approves a request or triggers another system, ordinary logs tell you about execution. A reviewer needs more: who authorised the work, what the AI was trying to do, what data it touched, who could intervene, what evidence it produced, and how the work was closed.

Hiveram keeps that record outside the AI actor. For the work you enter, it keeps the authorised request, the identity that claimed it, the runs and evidence submitted, the failures and interventions recorded, and the identity that accepted it; in the normal path that is not the identity that did the work. That gives audit, risk and quality teams answers to six questions.

What are the six questions?

Answers: this is the list

Different frameworks use different words. The operational questions converge on these six, and every other answer on this page points back to them by number.

  1. Who authorised it? This action, not AI in general.
  2. What data did it use or disclose?
  3. Why did it act? The intent it was working towards.
  4. Where is the evidence? Evidence by reference, as submitted, with its history, not a screenshot reconstructed after the fact.
  5. Who can stop it? Mid-task, and on the record.
  6. What happened when it failed?

What exactly is in the record?

Answers 1, 3, 4, 6

Logs record events. Hiveram records accountable work: the intent, the authority, the actor, the execution, the evidence, the revisions and an independent closure stay connected as one record. Five things ordinary AI logs do not give you:

  1. Intent before execution. The work order says what the agent was authorised to do.
  2. Exclusive execution identity. The claim says which actor had the right to do it.
  3. Separate acceptance. In the normal path the identity that did the work cannot close it; an override is possible and is recorded.
  4. Failures and interventions in the same record. Every change is written as a new revision and the history is kept.
  5. A decision survives a re-run. If the work is reopened or run again, the earlier closure, who made it and when, stays on the record as history; a later run adds to it and cannot replace it.

Two kinds of entry sit in that record, and the difference matters to a reviewer. Hiveram checks some entries against another source: that a linked commit is reachable on the declared repository, what the repository’s own check-runs concluded where you have authorised that read, and that the identity closing the work is not the identity that executed it. Other entries are submitted: notes, evidence references and artifact hashes are recorded as given, with who gave them and when. For work recorded through Hiveram, the record connects the stated intent, the actor, the submitted evidence and the closure history. It supports review; it does not by itself establish that every reported action occurred as described.

EntryChecked by Hiveram against another sourceRecorded as submitted
The intentNoThe work order and each revision, with when it changed
The changeThat the linked commit is reachable on the declared repositoryThe link to the commit
The checksWhat the repository’s own check-runs concluded, where you have authorised that readNo
The evidenceNoNotes, evidence references and artifact hashes, with who gave them and when
The closureThat the identity closing the work is not the identity that executed itNo
RecordWO-2274 · revision 64
Intent████████████ title redacted3
Claimed bynot retained in this record1
Executed by█████-wo2274-admission · identity redacted1
Pathopen → in progress → pending verification → done, 2026-09-264
Commit49b21ceaf11b · private repository4
Evidencecommit bound to the work order’s content hash sha256:b5dbf944…4
Closed byoperator-cli · a reviewer, not the executor1, 5

A real record from our own ledger, workledger/WO-2274. Redacted where marked; nothing added. Numbers on the right are the six questions above. What it shows is a work history, a commit bound to the work order and a closure by a different identity. It does not show a named person approving a specific action before that action, corroborated by a second record; the claimant was not retained in this record and the closing identity is a role, not a person.

We do not claim that nobody with access to your own infrastructure could alter it. We claim the record exists outside the actor from the start of the work.

Our AI vendor keeps logs. Isn’t that enough?

Answers 1, 4

A log inside the vendor’s product is the system grading its own homework. The vendor controls it, you may lose it when you change vendors, and it rarely records who authorised an action or what the person approving it saw. Ask any vendor one question: can I show my auditor the full record of what your system did, without you in the room? You need durable access to, and control of, the record that does not depend on continued access to the vendor.

We already run ISO/IEC 42001. Why would we need software for this?

Answers 1, 3, 4

ISO/IEC 42001:2023 is a management-system standard. It asks your organisation to set AI policy, assess impact, assign roles, monitor AI systems through their lifecycle, and keep documented information showing all of that happened. It does not produce that information.

A finance manager, commenting publicly on a list of AI frameworks, put it plainly: “teams can name the frameworks long before they can prove alignment to an auditor.”

A record of AI activity can provide the documented information your management system needs: who authorised it, its purpose, its execution history and the supporting evidence. Hiveram keeps that record outside the AI actor. Naming the standard is not a conformance claim; it is a description of where the evidence sits. Source

We follow the NIST AI RMF, or ISO/IEC 23894. Where does this fit?

Answers 2, 4, 5

NIST’s AI Risk Management Framework 1.0 (an update is in progress) has four functions, Govern, Map, Measure and Manage, and calls for documentation, measurement and monitoring across the AI lifecycle. Our design response is to keep that record outside the agent being governed, so Measure and Manage rest on an account that does not depend on the system under measurement. Source

ISO/IEC 23894:2023 gives guidance on managing AI risk inside your existing processes, and like the risk-management practice it builds on, it asks you to record and report what happened. The record is where that lives. Source

What do laws, regulators and standards ask organisations to document?

Answers 1–6 · Last regulatory review: 2026-11-08

Practice guidance. In Singapore, the Accounting and Corporate Regulatory Authority issued Audit Practice Guidance No. 1 of 2026, Using Artificial Intelligence Responsibly, on 2 October 2026. It describes audit documentation of AI-assisted work that may include the prompts used, the key assumptions, significant revisions made after human review and the rationale for accepting or rejecting an AI recommendation, so that a reviewer can follow what was done; the auditor remains accountable for the work. Source

Practice guidance. In the United Kingdom, the Financial Reporting Council published AI in audit: illustrative example and documentation guidance in June 2025. It describes proportionate documentation on the audit file of tools that use AI: the purpose, the version and configuration, the inputs and prompts, the suitability checks, the central approval of the tool and how its outputs were treated; the judgement of what to document stays with the auditor. Source

Supervisory expectations. On 7 October 2026 the Monetary Authority of Singapore finalised its guidelines on artificial intelligence risk management for financial institutions, including inventories of AI use and accountability for third-party AI, with the first obligations taking effect from 7 October 2027 and the remainder a year later. Source

Practice guidance. On 14 October 2026 Singapore’s Infocomm Media Development Authority announced a starter kit for testing agentic AI systems for safety and reliability. It checks the tools an agent used, the information it accessed, the decisions it made, any erroneous or unauthorised actions, whether it kept to its permissions and boundaries, and whether safeguards worked; it builds on the Model AI Governance Framework for Agentic AI of January 2026. The same day the AI Verify Foundation named its first five accredited third-party AI testing providers. Source (IMDA release, 14 October 2026)

Consultation draft. In Malaysia, the Personal Data Protection Department held a public consultation from 5 October 2026 to 23 October 2026 on a draft Artificial Intelligence and Personal Data Protection Framework under the Personal Data Protection Act 2010. The draft’s Part J lists, among the records a data controller keeps and makes available to the Commissioner on request, “relevant assessments, approvals and material decisions” and “governance decisions and approvals” (clauses 40.4(d), 40.5 and 42.3(e)). The consultation closed on 23 October 2026; no outcome had been published when we checked on 8 November 2026, so what the final text will require is not yet known. Source

Law in force. South Korea’s AI Basic Act and its Enforcement Decree took effect on 22 January 2026. The Ministry of Science and ICT has announced a grace period of at least one year during which penalties will generally be deferred while organisations prepare. Under Article 34, a business providing high-impact AI must operate risk management measures, ensure human oversight, and prepare and retain documents that verify those measures; Article 35 asks it to endeavour to assess the impact on fundamental rights in advance. Source

Law in force. In the European Union the AI Act already applies in stages: most provisions since 2 August 2026, obligations for high-risk systems from 2 December 2027 and, for AI embedded in regulated products, from 2 August 2028. The evidence you need depends on the system and on your role under the Act (European Commission timeline). These laws can also reach suppliers outside their home jurisdictions, depending on where a system is placed on the market, deployed, supplied or used, and on the supplier’s contractual role. We describe these texts as published, and we label each by what it is. None of them, alone, requires anyone to adopt an independent ledger; each asks for documentation that a reviewer can follow, and that is where the record sits. We are not your counsel, and this page will change when the dates do.

Does Hiveram need access to our systems, licences or tools?

Answers 1, 4, 5

Your team works through its existing tools, licences and access, and writes the record through the Hiveram CLI, API or MCP server as the work happens. Hiveram receives the records submitted to it; it holds no credentials to your other services and does not perform the work itself. Where you authorise it, GitHub verification also reads check results from your own repositories. When what an agent did cannot be recorded in the form the record requires, the record refuses to close the work and says why.

If your existing tickets and logs already let an independent reviewer reconstruct who authorised a piece of AI-assisted work, what was done and who reviewed it, assess whether another system adds anything before adopting one.

Does this make us compliant? Do you hold a certification?

Answers 1–6

No to both. Obsta Labs holds no certification against ISO/IEC 42001, ISO/IEC 23894, the NIST AI RMF or any other standard named here, and using our products does not by itself make you conformant to any of them. Certification is your organisation’s achievement, against your management system. Our part is making sure the evidence exists when the auditor asks.

Who can stop an agent?

Answers 1, 5, 6

In Hiveram the normal path to closing a piece of work requires a different identity from the one that did it, so an agent does not mark its own work done. Which actions need a human signature is your policy; the record shows who signed and when. Hiveram applies the same separation-of-duties principle: execution and closure are different identities.

An agent works under a claim the ledger grants. A person can see every active claim and release it, and the release is on the record. Hiveram records interventions and claim releases; whether an agent stops acting depends on the controls in its execution environment. What the record withholds is closure: work is not marked done until a different identity closes it. When a run fails, it is recorded against the work it addressed, so a failure is an entry in the record rather than a gap in it.

Does this slow our teams down?

Answers 1, 4

Policy checks happen inline with the work they govern; the alternative is reconstructing evidence after an incident. The record is written as the work happens, through the same steps the agent already takes: claim, work, close. It is not reconstructed afterwards.

We run our own company on our own products: a fleet of AI agents builds the software, and every piece of work carries a contract, a verification step and a closer who is not the author.

The record also compounds. Every closure, refusal and correction stays with the work it belongs to, so the next piece of work starts from what the last reviewer found, not from a blank prompt. A team gets better under review, not in spite of it, and the review is the part the AI cannot do for itself. We are measuring that effect on our own fleet before we claim a number for it.

How does Hiveram fit with NeuroRouter?

Answers 2, 5

Hiveram is the record. A companion product covers one edge around it, and it is not required to use it.

What leaves your network. With NeuroRouter in front of your model providers, detected credentials, tokens and connection strings are redacted or blocked before a request is forwarded, according to your policy. It runs locally, forwards only to the upstream provider you configure, and does not phone home. It is not a promise to catch every encoded or transformed secret. It removes a hosted credential store from the path.

Where do our records live?

Answers 4

With you. Hiveram runs locally, on infrastructure you control against your own PostgreSQL, or managed by us if you prefer. When you run it yourself, nothing leaves your environment unless you decide it does. It installs with one command; see the install guide, and pricing for plans and the 14-day trial.

We already record reviews in our agent platform. Do we need this?

For teams that build agents

Perhaps not. Agent platforms and tracing tools already record traces, reviews and approvals. If yours lets someone outside the work follow one item from the request, through the evidence, to who accepted it, keep it. Where that chain breaks across tools, the record connects it, beside whatever runtime you use. The mechanics are on the page for teams that build agents.

These questions come from real cases: see public cases, each with its primary source.

For security-gate teams: where your check results, waivers and closures sit.

Pricing and trial Deploy Ask a question