Three planes. Explicit ownership.
CortexSim separates the authority to request work, the custody of source data and the environment where experiments run. Each plane has a different responsibility; a connection between them does not transfer unrestricted permission.
The control plane owns organizational registrations, exact grants, policy versions and approvals. It determines which identified actor may invoke which operation in which environment. Discovery makes a tool visible; authorization determines whether it can be used.
The customer data plane owns source credentials and source-specific enforcement. A private connector reads an approved view, stream, endpoint or journal. The simulation plane consumes reviewed evidence to construct and inspect executable business twins, then explores changes within isolated environments.
Control plane
AuthorityIdentity → policy → bounded dispatch
Customer data plane
CustodyApproved source → private connector
Simulation plane
ConsequencesVersioned twin → experiment → evidence
- Control: identity, trust registration, tool scope and approval.
- Private data: credential custody, approved reads and reviewed mappings.
- Simulation: versioned models, bounded experiments and verification evidence.
Two connection directions. A separate action boundary.
Inbound MCP access and outbound source access answer different questions: who may ask CortexSim to do something, and which organizational information CortexSim may obtain. Both require their own authorization.
An approved AI client reaches the gateway using its registered identity and environment scope. The gateway evaluates the requested tool and arguments before dispatch. Browser sign-in alone does not authorize MCP calls, and a read grant does not imply permission to fork a simulation.
For private sources, the connector initiates an authenticated outbound channel. Source credentials stay close to the source. A model-generated proposal is still only a proposal: supported action services require separate, independently approved action identities and profile-specific controls.
Ask within an explicit tool boundary.
A registered application initiates a request in its approved environment.
Client authorization governs the tool call. Source access and action authority remain separate.
- Identify the initiating actor and the service executing the operation.
- Keep client grants, source approval and action approval distinct.
- Reassess authority when arguments, scope, policy or evidence change.
A deliberate, seven-tool interface.
The enterprise gateway exposes a closed tool catalog for discovery, synthetic simulation and minimized evidence. The table below describes the implemented surface—not every capability in the surrounding CortexSim platform.
Inspection tools return bounded summaries. The transaction preview evaluates at most ten operations against a synthetic lab checkpoint without persisting effects. Forking creates an isolated branch only after independent approval bound to the exact checkpoint and arguments.
Twin evidence is minimized and pinned to model and verification identities. It does not expose held-out records or source credentials. Scenario planning, recurring missions, compensation and recovery are surrounding platform services; they are not additional tools hidden behind this catalog.
| Tool / purpose | What it returns or changes | Boundary |
|---|---|---|
Discover approved worldscortexsim_list_worlds | Up to 25 world summaries in the approved environment. | Read only · scoped environment |
Understand a worldcortexsim_world_summary | Inspect a fingerprint and business-object counts without raw transactions or account balances. | Read only · summarized state |
Rehearse a transactioncortexsim_preview_transaction | Preview up to 10 operations against a synthetic lab checkpoint without persisting effects. | Synthetic lab · bounded preview |
Create an isolated branchcortexsim_fork_world | Fork an exact synthetic lab checkpoint after independent, argument-bound approval. | Synthetic lab · explicit approval |
Find source-backed twinscortexsim_list_twins | Discover paginated synthetic twin summaries through exact grants or an administrator principal. | Read only · no source records |
Inspect model and verificationcortexsim_twin_summary | Read current contract and rule parameters alongside aggregate verification. | Read only · no held-out records |
Trace the verificationcortexsim_twin_evidence | Retrieve minimized training-case evidence pinned to twin and verification digests. | Read only · pinned evidence |
Find a ready-made experimentcortexsim_list_scenarios | List the scenario library and the organisation's saved scenarios, 25 per page. | Read only · titles are untrusted data |
Choose a model to testcortexsim_list_models | List published models with how many held-out requests each one replays. | Read only · scoped environment |
Run a scenario on a modelcortexsim_run_scenario | Run a scenario as an in-process model experiment and return its verdict and aggregate measures. | Simulation scope · no traffic · no recorded payloads |
- Use the narrowest tool that answers the question.
- Keep environment and object scope explicit in the request.
- Treat returned labels and source text as data, not as instructions that expand authority.
Private context, connected through reviewed profiles.
Source onboarding starts with a purpose and an accountable owner. Choose a supported profile, constrain the permitted source and agree the data that may cross the boundary before collecting it.
Reviewed mappings connect source fields to the model. Row and field scope, request budgets, source identity and ingestion provenance make the collection boundary inspectable. A generic source adapter does not establish compatibility with every vendor application or permission to scan its entire database.
A connector-initiated channel avoids requiring a publicly exposed customer database. It does not remove the need to qualify customer networking, identity, credential management and the exact source profile. PostgreSQL change capture is an isolated lab profile; production CDC acceptance remains separate.
PostgreSQLpostgres-view
Reviewed read-only viewsBring relational business context through scoped views and reviewed mappings.
PostgreSQL CDCpostgres-cdc
Isolated lab profileExplore change capture in an isolated lab environment; customer production CDC qualification remains deferred.
Redis Streamsredis-stream
Bounded stream readsConnect ordered event context using an explicitly approved stream profile.
HTTPS APIsrest-read
Approved read endpointsRead bounded source responses through reviewed REST profiles.
Organizational MCPmcp-read
Pinned read profilesConnect approved organizational MCP sources through a separate outbound trust boundary.
SQLitesqlite-view
Scoped source viewsUse approved local relational views for controlled source-backed experiments.
Event journalsevent-journal
Reviewed record inputsBuild from bounded journal evidence with explicit ingestion and mapping controls.
- Keep source credentials out of prompts and model output.
- Preserve source identity, snapshot references and mapping versions.
- Review schema or destination changes before broadening collection.
Carry the evidence into the decision.
A useful simulation result needs a reproducible context: which source evidence, model version, scenario and verification produced it. Evidence should let a reviewer explain a result and recognize the limits of the model.
Consider a synthetic payment-processing evaluation. A reviewed source view describes the declared population; authored contracts and rules define the twin. A candidate scenario changes a supported service parameter, then runs against its declared baseline. The review examines the outcome alongside invariants, exceptions and assumptions.
The illustrative chain below shows the artifacts to retain. It is not a reported customer result. A successful numerical check does not establish business usefulness, representative coverage or production authority. Those conclusions require their own evidence and review.
What should travel with the result?
- 01Source snapshot
Declared population, approved fields and input identity.
- 02Twin version
Contract, rule parameters and stated model assumptions.
- 03Scenario pair
Baseline, candidate change and the same declared workload.
- 04Verification record
Outcome, invariant checks, exceptions and retained failures.
- 05Review decision
Usefulness assessment, limitations and the next acceptance gate.
- Pin the source snapshot and model identity used for the run.
- Compare the candidate and baseline under the same declared workload.
- Retain unsuccessful outcomes, exclusions and unresolved questions alongside successes.
Design for interruption, not just completion.
A distributed operation can stop between dispatch, downstream processing and acknowledgment. CortexSim surrounds supported action profiles with durable records, explicit reconciliation and a separately governed compensation path.
A timeout does not prove that nothing happened. An uncertain downstream outcome must be reconciled against the supported target profile before another action is taken. Compensation is a new governed operation, not an assumption that any external effect can be perfectly reversed.
Finite recurring missions bind explicit approved action identities to a schedule and expiry. Recovery also has a data boundary: custody-bound restore protection must survive restoration of an older application database. Retaining a backup is only one part of a valid recovery procedure.
| When this happens | Why it matters | Review response |
|---|---|---|
| Source schema changes | The old mapping may no longer describe the input. | Review the schema and mapping; retain the original source references. |
| Policy or approval changes | Previously permitted work may no longer be authorized. | Re-evaluate current authority before a new dispatch. |
| Acknowledgment is lost | The downstream outcome may be unknown. | Inspect the receipt and reconcile through the supported profile before deciding on another dispatch. |
| An effect needs compensation | A completed effect is not erased by a retry. | Prepare a separate approved compensation operation and retain both histories. |
| An older database is restored | The backup may predate erasure or custody decisions. | Preserve independent custody and validate the restored state with a compatible runtime. |
- Distinguish rejected, completed and unknown outcomes.
- Review retained receipts before deciding whether another dispatch is permitted.
- Preserve independent custody and use a compatible runtime for a restore.
Separate architectural fit from deployment acceptance.
The current delivery is qualified for its recorded internal-use scope. Customer production deployment remains deferred. Architectural diagrams explain responsibilities; they do not certify a particular customer environment.
A customer-hosted or hybrid arrangement must qualify the chosen identity provider, source profiles, network paths, custody and operational procedures. A disconnected deployment requires separately qualified offline import/export and trust-root procedures. None of these follow automatically from local test passes.
Production qualification still includes representative customer interfaces, sustained load, physical-host and independent-site recovery, and independent security, domain and accessibility assessments. Use the internal implementation to scope an evaluation and identify the evidence needed for your deployment.
Implemented within recorded contracts
Local qualification, synthetic profiles and retained recovery evidence define the current delivery scope.
Qualify the selected environment
Customer identity, source integration, infrastructure and operational acceptance remain deployment-specific.
Separate qualification remains
Production deployment and disconnected operating profiles are not established by internal-use evidence.
- Record the exact runtime, connector profile and environment being evaluated.
- Distinguish implemented controls from customer-specific acceptance.
- Keep production actions outside the internal synthetic evaluation scope.
Start with one decision worth testing.
A focused evaluation connects one approved client, one private source and one business question. Its output is an evidence package and a decision about the next step—not an automatic production rollout.
Choose a question narrow enough to model and important enough to review. Agree the baseline, supported changes, success criteria and unacceptable outcomes before running the candidate. Assign both an implementation owner and a reviewer who can judge the domain meaning of the result.
Use the checklist below to prepare the evaluation brief. Selections stay in this page session and are included in the downloaded guide. They record your preparation only; they do not change platform permissions, approve a source or authorize execution.
Evaluation preparation
0 / 8 prepared- State the question, business owner and intended decision.
- Define the exact source, client identity and approved environment.
- Retain the evidence and agree a separate acceptance decision.