Skip to content
Free AI Benchmark — see where you are exposed before you deploy AI.Start the benchmark

Why MergeOn

Turn the knowledge you already have into governed context AI can use.

Policies, procedures, manuals, regulations and operating documents were written for people — not AI.

MergeOn Document Intelligence transforms them into structured Tier-3 Review Packs, preserving meaning, relationships, dependencies and source context. Accepted knowledge can then be governed in the Knowledge Center and supplied to AI at execution.

Better context. Less repeated processing. More efficient AI execution.

Explore Governed Knowledge

Build on MergeOn

The application asks for a business outcome. The Runtime governs how it is produced.

An integration calls a published Business Capability rather than a model endpoint, so knowledge, policy, protection, human authority and evidence stay part of the activity instead of becoming your problem to rebuild.

Everything below describes the contract you build against and the architecture behind it.

Explore the developer platform

Know where you stand

Most organizations do not have an AI problem. They have a clarity problem.

Before deciding what to build, it helps to establish what your organization already believes about ownership, governance and decision-making — and where those beliefs disagree with each other.

Start with an honest read of where you are. Everything else follows from it.

Start the free benchmark
Developers/API Reference

Integrate with the governed business contract — not the model behind it.

MergeOn separates the application contract from the AI participant that fulfils it. Requests target governed business activity in a published Runtime; models, agents, tools and systems participate behind that boundary under the controls attached to the operation.

The caller asks for business work. The Runtime determines how that work is governed and fulfilled.

Request contract
Caller
Runtime + environment
Published business capability
Structured input
Governed execution
Declared output
Execution identity / evidence
Conceptual contract — not an endpoint or SDK example.
The API mental model

Address the activity. Do not address the model.

In a direct model integration the application ends up knowing the provider, the model, the prompt, a model-specific request shape and how to handle its response. Every one of those couples the application to a probabilistic participant that will be replaced before the application is.

In MergeOn the durable integration target is the governed business activity. The caller supplies the declared input. The Runtime resolves which published configuration and which permitted participant fulfil it.

ApplicationBusiness capabilityPublished RuntimeGoverned execution
Fulfilled behind the boundary by any of
ModelAgentToolSystemProcess

Your application depends on the contract. MergeOn owns the execution boundary behind it.

Address the right thing

Runtime, environment, release and capability are different coordinates.

These are the coordinates that specify what a request actually targets. They are routinely flattened into a single idea of “which version”, and that is where most integration confusion starts.

Runtime

The governed operating boundary the activity belongs to.

Environment

The execution context — for example Sandbox, Staging or Production where applicable.

Release

The published executable version of the Runtime configuration.

Business Capability

The declared business activity and contract the caller intends to use.

Execution

The individual occurrence of that governed activity.

RuntimeEnvironmentPublished releaseBusiness CapabilityExecution

Do not collapse environment, release and execution into one generic version.

Execute what was published

Draft configuration is authoring state. Published configuration is execution state.

Developers need a stable target. A draft keeps changing while the activity is being designed — that is what drafts are for. Governed execution resolves against the published release appropriate to the target environment, rather than whatever the configuration happened to be at the moment of the call.

Draft

Authoring state. It can keep changing while the activity is being designed, which is exactly what makes it a poor execution target.

Review & validation →
Published release

Execution state. An approved version, resolved for the target environment. This is what a request actually runs against.

Execution →
Evidence

The record of what ran, including which published release governed it.

Author the draft. Publish the release. Execute the published truth.

Structured business contracts

Define what the activity accepts and what it promises to return.

A governed capability carries a declared business contract: its purpose, structured input, validation, declared output and execution context. The calling application should not have to read prose out of a completion to work out whether a business operation succeeded.

Structured inputGoverned activityDeclared output

Business integration needs a contract, not a conversation.

The caller does not bypass governance

A technically valid request can still be refused.

Admission to execution depends on conditions that sit outside the calling application’s logic. A request can satisfy the technical contract perfectly and still not be permitted to run — and that is the platform working, not the call going wrong.

RequestCommercial admissionPolicyProtectionHuman authority where requiredExecution

Valid does not mean authorized.

Handle the right failure

A system error and a governed refusal are not the same thing.

Three conceptually different outcomes, which naive clients collapse into one “it didn’t work” branch. That collapse is what produces retry storms against decisions that were made deliberately.

Request / system failure

The request cannot be processed because the technical contract was not satisfied, or a required service failed.

Fix the cause, then retry
Governed refusal

The request was understood, and a governing condition does not permit execution. Nothing malfunctioned.

Change something, do not retry
Authority required

The activity cannot continue until the required human authority has acted. It is waiting, not failing.

Wait and resume

Do not retry a governance decision as though it were a network error.

This page teaches the semantics. Exact status and code-level contracts belong in the technical reference when it is published, and are deliberately not guessed at here.

Trace the operation

Every execution needs an identity that survives beyond the request.

A single enterprise execution can involve the caller, a Runtime, an environment, a release, a business capability, a model or agent or tool, policy, human authority, downstream systems and an outcome. Durable execution identity is what keeps those facts connected once the HTTP request is long finished.

Execution ID

The identity of this occurrence of the governed activity.

Correlation

What ties related work together across the operation.

Root execution

Where a chain of governed work began, when one activity leads to another.

Trace the business operation, not just the HTTP request.

Retries are business events

Do not accidentally execute the same business activity twice.

Distributed systems retry. Timeouts happen. Clients lose responses they were waiting for. For a business operation, a blind retry is not a free action — it can become a second execution with real consequences.

Retry behaviour has to respect execution identity, and use the idempotency contract of the specific execution surface where that surface supports it. It is not safe to assume idempotency is universal across every surface.

Retry the request without duplicating the business effect.

Authority can pause the operation

Approval is not an error response.

Where Human Approval governs an activity, the operation can reach a state in which execution cannot continue until an authorized person decides. A caller that treats that as a model failure, or retries it, has misread governed workflow state as a fault.

RequestGoverned executionApproval requiredAuthorized decisionContinue / refuse

The API request can finish before the governed business operation does.

Minimize what participates

The model does not need every value the caller can see.

A calling application often holds sensitive information that is irrelevant to the AI participant’s task. Protection belongs inside the governed activity, so the execution can preserve the business context the task needs without automatically exposing every original value that happened to be in scope on the caller’s side.

Application access does not automatically become model access.

Supply accepted context

Do not make every caller rebuild retrieval logic.

When context is the caller’s responsibility, every integrating application ends up implementing its own retrieval, chunking and prompt assembly — each slightly differently, each drifting apart over time. Governed Knowledge lets accepted organizational context belong to the Runtime activity instead.

Application requestBusiness capabilityGoverned knowledgePermitted participant

Knowledge is part of the activity's operating context, not caller-side prompt plumbing.

The response is not the whole record

A business result and its execution evidence are different things.

The caller needs the declared business outcome. The organization also needs the governed record of how that outcome was produced — what was requested, which Runtime and environment applied, which published release executed, what governed knowledge participated, what policy and protection applied, who authorized where required, which participant executed and what outcome followed.

Those are two different consumers with two different needs, and conflating them makes both worse: the caller gets a response it has to parse around, and the organization gets evidence that only exists if someone kept the response.

Return the business result. Retain the operating evidence.

Enterprise boundaries

One business operation can involve more than one governed Runtime.

A Primary Runtime can coordinate enterprise activity while Connected Runtimes retain authority over capabilities belonging to their own operating area. For a caller this changes nothing structurally: you depend on the declared capability, never on what the other side does internally to fulfil it.

CallerPrimary RuntimeConnected Runtime capabilityReturn

Cross the boundary through a declared capability, not by reaching around it.

Version the right contract

The API, Runtime release and business capability are not the same version.

Four things version independently. Treating them as one is the fastest way to make a routine change look like a breaking one — or to miss a breaking one entirely.

Platform / API contract

How a caller communicates with MergeOn.

Runtime release

Which published configuration governs execution.

Business capability contract

What business input and output the caller depends on.

Model version

Which probabilistic participant may currently fulfil the activity.

Changing the model should not automatically become an API-breaking change.

The caller depends on the business capability contract. A qualified participant changing behind that contract is a governance event, not an integration rewrite.
Before you integrate

Know what your application is actually depending on.

Ten questions worth answering before the first line of integration code. Most production surprises trace back to one of them being assumed rather than established.

01

Which Runtime owns the activity?

02

Which environment is being targeted?

03

What published business capability is the caller using?

04

What structured input does it accept?

05

What declared output does it return?

06

What governance can refuse or pause execution?

07

Does the operation require Human Approval?

08

What retry and idempotency behaviour does the specific surface support?

09

What execution identity should the caller retain?

10

What evidence belongs to the operation?

Build to a business contract that survives the model underneath it.

Keep the caller stable while Runtime configuration, governed knowledge, controls and qualified AI participants evolve behind the execution boundary.