Technology Participation in Enterprise Intelligence — browse technologies by architectural contribution.
From a control decision to an evidenced outcome
Arqua architectural interpretation of selected public documentation, reviewed 25 September 2026. Not a benchmark, certification or deployment recommendation.
A useful enterprise control must be assessed against the operation it is intended to govern. What was proposed? Which revision was reviewed? What conditions applied when the operation could take effect? What did the implementing system actually permit, prevent or leave unresolved?
Arqua's architectural approach is to make these relationships explicit without assuming that another platform is required. Existing policy engines, applications, workflow systems and operating practices can provide substantial controls. The assessment starts by identifying those capabilities and the conditions under which they meet the enterprise requirement.
Different technologies contribute different responsibilities
Policy decisioning evaluates a request against declared rules and context. Workflow mechanisms coordinate participation and can apply controls to particular transitions. Durable execution supports recovery and continuity. Document processing prepares source material for use. Evaluation and tracing provide evidence for assessment.
These responsibilities can overlap within products, but are not interchangeable. A policy decision, a completed task, a recorded attempt and a target-system receipt may refer to different events. An implementation needs to preserve the relevant relationships rather than treat all of them as a single approval.
The five profiles below examine Cerbos, Camunda, Temporal, Unstructured and Arize Phoenix. They are selected documentary examples, not five competing equivalents or a recommended combined stack. The relationship between their mechanisms and enterprise responsibility is Arqua's architectural interpretation, not an observed implementation finding.
Cerbos: connect the decision to trusted inputs and enforcement
The Cerbos decision API, displaying PDP 0.54.0 in the reviewed source, evaluates specified principals, resources, actions and attributes. Requests can select policy versions and scopes; omitted versions use configured defaults. Responses provide decisions and optional evaluation metadata. This makes the policy question inspectable without establishing that the application enforced the answer.
Bind the supplied attributes to their trusted sources and the decision to the operation actually proposed. A permission result for one revision should not silently become permission for a changed request. A policy-version label also needs a known relationship to the policy content in force. These are implementation questions, not a finding that the decision engine is ineffective.
Cerbos's audit configuration supports joining its call identifier with application logs, with configurable capture, filtering and storage. Inspect those settings before promising complete evidence. The decision receipt, enforcing application's event and target outcome may be separate records. A governed application can use native policy decisions to exercise legitimate delegated authority; an extra policy engine is not automatically needed.
Camunda: recognise native blocking and its lifecycle boundary
The reviewed Camunda 8.9 user-task listener documentation describes blocking lifecycle transitions and denying assignment, update or completion. Creation and cancellation cannot be denied through that mechanism. Termination through process-instance modification deliberately skips the cancellation listener. These are specific control boundaries, not a blanket absence of preventive workflow controls.
Identify whether the protected operation is the task transition or a separate effect performed in another system. A denied transition preserves task state; it should not be interpreted as reversal of an external change. Administrative recovery paths need appropriate authority and evidence, rather than being assumed to have the same controls as ordinary completion.
Camunda's job-worker documentation describes at-least-once job delivery and timeout reassignment, potentially with two workers processing the same job. One accepted completion message does not establish that external work happened once. Worker and target behaviour must preserve the intended operation across duplicate attempts. Confirm the actual Camunda release, task interface and configured listeners before applying this selected account.
Temporal: distinguish recovery from permission to produce another effect
Temporal's Activity Definition documentation recommends idempotent activities. It distinguishes completed activities retained in workflow history from work whose completion was not reported and may be retried. Replaying completed history is not the same as executing the external operation again.
For an uncertain outcome, first establish whether the target already acted. Use an operation identity and receipt or deduplication mechanism appropriate to that target. An attempt identifier alone may not connect the same business operation across restarted workflows. Separately determine whether producing a new effect remains permitted under current conditions. Safe repetition and current permission answer different questions.
The Activity Execution documentation describes cooperative cancellation: heartbeat delivery, SDK behaviour and whether an activity accepts the request matter. Workflow code can choose whether to wait. A cancellation request is therefore not proof of interruption, and interruption is not proof that a committed effect was undone. Any compensating change has its own purpose, authority and result. Native durable execution and existing target safeguards may already provide the required solution.
Unstructured: retain enough lineage to check the source
Unstructured's document-element model includes element types, identifiers, text and metadata. Available metadata depends on the source and processing path. A generated identifier helps reference extracted information; it does not authenticate the source or establish that an applicable exception was captured.
Its open-source chunking documentation explains that consolidation can obscure element-level origin, page membership and coordinates. Retained orig_elements can preserve that detail. A chunk can span pages, so assigning it one convenient page number can misrepresent its origin.
Preserve the source revision, extraction configuration and relationship from original elements to consumed chunks. Check whether the decisive condition or exception survived, not merely whether a passage was retrieved. Retaining originals can support review, but may also retain sensitive image or table content; compressed or encoded material is not thereby protected. Choose appropriate access and retention rather than copy every original into general-purpose logs. This profile does not establish commercial-platform guarantees or an inspected extraction run.
Phoenix: evidence the checks that actually ran
Phoenix's dataset-evaluator documentation distinguishes automatic execution for UI experiments from programmatic experiments, which must receive evaluators explicitly. It also supports mapping evaluator inputs from dataset fields. Saving an evaluator is not evidence that it assessed a particular result.
Retain the dataset revision, required evaluator set, actual invocations, input mappings, configuration, results and failures. A missing, failed or skipped check must not be reported as passing. A successful evaluation can support a defined quality claim; it does not appoint a decision-maker or authorise an unrelated operational act. Expectations and thresholds should reflect the purpose, rather than simply reward agreement with a generated answer.
The retention documentation describes configurable time- or count-based trace purging from Phoenix 9.0 onwards, alongside deletion facilities. Plan what must remain reconstructable and who controls that custody. A missing trace does not prove an action never happened. This is a Phoenix-specific account, not an assertion about every Arize service or a requirement to retain sensitive content indefinitely.
Three possible implementation outcomes
A native application may already supply sufficient review, enforcement and evidence. A composed implementation may distribute responsibilities across policy, workflow and target systems. An evidence-assisted process may support an accountable human reviewer without introducing autonomous action.
The appropriate choice depends on the requirement and observed implementation, not on the number of products included. Another policy engine, orchestrator, graph or Arqua runtime is not a predetermined outcome.
Arqua's Enterprise Control Plane and Context Qualification definitions distinguish distributed governance from a single tool, and suitability for reliance from execution authority. Formation and qualification both belong within Runtime Context Assembly. Policy decisions, workflow controls and evaluations can contribute to these responsibilities without becoming the entire architecture.
For a proposed operation, connect the reviewed revision, applicable authority, current conditions, actual enforcing component, execution attempts and observed outcome. The evidence may remain distributed across existing systems. A single central evidence store is not required merely to make those relationships explicit.
Begin with one proposed service-record change
A bounded assessment can start with a source-supported change proposal, human review and a test target. Change the proposal after review, delay a decision, and simulate the target accepting work while its acknowledgement is lost. Check which operation was permitted, which result remains uncertain and whether a retry produces an unjustified additional effect.
Then vary the supporting evidence: omit a source exception during chunking, run an experiment without a required evaluator, or remove traces under the applicable retention policy. Determine which conclusions remain supported. These are proposed assessment cases, not executed product or Guide tests, and do not require all five products to participate.
The architectural objective is to preserve the relationship between the operation reviewed, the controls actually applied and the outcome evidenced—so that technical success is not mistaken for an authorised and understood result.