What NovusLens will not do, the map is not the brakes.
NovusLens reads. Your own systems control. It sits beside an organisation’s AI agents and reads what they declare and what their own systems report. It shows where control points belong and what the records show. It does not operate the systems that control the agents, and it never gives a verdict that an agent complies with a rule. This is the most important limit in the product.
What it is
The foundation paper distinguishes two layers of systems. Its labels are Layer ALayer A and Layer B boundaryThe division between advice and action. NovusLens records and explains evidence in Layer A. The customer’s systems run and control agents in Layer B. and Layer B.
- Layer A is the infrastructure NovusLens is built on. It is NovusLens’s own.
- Layer B is the customer’s runtimeRuntimeThe software that runs a model or agent. A model runtime answers model requests; an agent runtime carries out the customer's agent actions. control: activity monitoring, rules that limit actions, emergency stopping, undoing changes and settling disputed actions. The customer runs it, either through a separate runtime product or through their own vendor. It is outside NovusLens’s responsibilities.
NovusLens stays on the mapping side of that line. It maps the agent estateAgent estateThe collection of AI agents an organisation operates, together with the systems and providers they use., surfaces findingsFindingA difference in the records that needs a decision. It includes its sources and can describe a slow handoff, disputed fact or missing owner., locates control points and advises on them. It produces evidence that someone else can review.
An example
Take one purchasing request that the resource reports as refused. The runtime nevertheless records that same request as completed. NovusLens reads both records.
NovusLens reports the difference between the refusal and the recorded completion. That is an enforcement gap in the runtime’s records, and it is named as one. NovusLens does not block the next request, does not roll anything back and does not declare the agent compliant or non-compliant. The person responsible for the runtime decides what to change.
What NovusLens does and does not do
| NovusLens does | NovusLens does not |
|---|---|
| Map the agents, their declared CapabilityCapabilityWhat a team can provide or an agent can do. NovusLens records the claim and the evidence that supports it., AuthorityAuthorityWhat a team or agent is allowed to decide or do. For an agent, this comes from a named person or system that grants permission. and ResponsibilityResponsibilityThe work or result that a team or agent answers for. NovusLens records the person answerable for it., and the systems they depend on. | Run, stop, pause or steer any agent. |
| Surface findings where declarations and reported activity differ. | Act as the runtime control planeRuntime control planeThe customer’s software for allowing, refusing, stopping or undoing an agent’s actions. NovusLens reads its records but does not operate it.. |
| Locate where a control point would belong, and of what kind. | Choose the implementation of a control. |
| Advise the responsible person, with the evidence and its limits. | Issue a compliance verdict, such as “complies” or “compliant”. |
| Produce evidence for the compliance function to review. | Treat an approved declaration as proof that the runtime applies it. |
| Generate a policy exportPolicy exportA file of rules generated from an agent’s recorded permissions, in Cedar or Rego format. NovusLens generates the file. The customer decides whether to use it. in a form the customer’s runtime can read. | Deploy an export, or send it to a runtime on the customer’s behalf. |
| Record who approved, exported and reviewed each item. | Silently rewrite history. Corrections keep earlier records; explicit purgePurgeRemoving a stored document’s contents while keeping a record that it was removed and why. Conclusions that used it must be checked again. and organisation erasure are separate operator actions. |
The verbs used across the product are maps, surfaces, locates, advises and produces evidence. They are chosen on purpose.
Policy exports are generated, never deployed
A remitRemitA written list of an agent’s allowed actions, resources and destinations for information. The list states whether it is complete. authored in the product can be exported in two policy formats, CedarCedarA language for writing rules about who may perform an action on a resource. NovusLens can generate files in this format for the customer to review. and RegoRegoA language for writing rules that a control system can check. NovusLens can generate files in this format for the customer to review.. The export is the approved declaration written as a document that a runtime could load.
The export function generates the document and records the export. It does not deploy anything. Every export’s header states that the customer deploys the file through their own control system. NovusLens records “deployment reported” only from a report naming the digest, the file’s calculated fingerprint. Enforcement is observed only in the activity evidence. The export also lists every clause the target format cannot carry. No lossless translation is claimed.
Two uses of the word boundary
The word “boundary” has a second meaning in the product, and the two should not be confused.
The permeable organisational boundary separates the organisation’s units from dependenciesDependencyA link where one unit needs something from another. A dependency is not a reporting line: it shows where work relies on something else. it cannot inspect internally, such as suppliers. Work and evidence cross that line. External nodes are placed on their own confidence ladder: black box, then declared by service level, then measured from the organisation’s own side. The boundary feedBoundary feedRecords supplied by the customer’s own systems about activity at the edge of an agent or service. NovusLens reads those records; it does not watch the agent. in the code is that organisation-side measurement.
The Layer-A/B boundary in this page is the one between mapping and control. It is about who controls the agents, not about what sits inside the organisation.
For readers who trace the foundation paper: the permeable organisational boundary is assumption A11, and the rule on this page is assumption A17.
Limits
- NovusLens can only report what the sources tell it. A source that is simulated, or one that does not record a value, leaves the question open, and the result says so.
- An approved export shows that a declaration was formal. It does not show that a runtime applies it.
- A finding is evidence for a human to assess. The finding’s last words are “evidence, not a diagnosis”.
What to do next
Read agentic accountability to see how the three accountability questions for agents are kept apart. Then identify, for each agent, which runtime records its activity and which person answers for that runtime. Ensure that runtime reports refusals as refusals.
Export a policy file from the approved declaration. Review every clause the export cannot represent, then deploy it through your own change process.
Where it is implemented
apps/reasoning/src/conway_reasoning/remit_authoring.pyrenders the Cedar and Rego exports and stamps each one “generated, not deployed”.apps/graph/src/conway_graph/agent_authority.pycompares observed activity with a remit and returns results, not instructions.apps/graph/src/conway_graph/governance.pyraises the control-gap findings from declarations and reported behaviour.apps/reasoning/src/conway_reasoning/agent_delivery.pydelivers evidence from a queue with a fixed size limit and states that it never controls the customer’s estate.apps/reasoning/src/conway_reasoning/catalogue.pyholds the reviewed mapping from business activities to operations, so that a remit’s words mean the same thing across the product.