One corpus. Three endpoints. The same citation chain.

Stage 03 is how the layer leaves your tenant in a controlled way. The same corpus serves three endpoints at once: your internal dashboard, your brand-partner API, and your AI-agent retrieval feed. Each endpoint enforces its own scope at the data path. Each endpoint returns the same fact with the same citation. The configuration that turns a corpus into a business unit is all three endpoints active at once.

This page is the product view of the three endpoints: how you turn each one on, how scopes are minted and checked, and what your team can configure from the console. The security cascade that enforces tenant and segment isolation is documented under Trust & Security.

Your own strategy, merchandising, or operations team consumes the corpus through dashboards. Beauty: which categories to expand. Consulting: which capability gaps to staff. Restaurants: which shift patterns predict churn. Retail: which substitutions explain the drop in a category line. The dashboard groups fact rows by dimension, by segment, by time window. Every cell links back to the chunks that produced it.

This is the endpoint your executive team uses to make decisions. It is the lowest-controversy surface in the platform and the wedge that justifies the rest. Pulse-tier deployments stop here. Macro and Custom Enterprise tiers light up the next two endpoints on top of it.

The brand-partner API is the endpoint that converts the corpus from a cost line into a revenue line. Each brand partner gets an API key minted with a segment scope: the slice of your corpus they are allowed to read. At a beauty retailer, that scope is “facts whose segment vector intersects brand_partner = X.” The partner sees their own brand’s perception map, their substitution graph, and the customer-segment behavior in their category. They see nothing outside it.

Scope enforcement lives in the data access path. The session ticket minted for the partner carries the scope. That scope cascades into which graph nodes the partner can traverse, which Postgres rows they can read, and which verbatim chunks resolve when they cite. An unauthorized query is not denied with an error; the unauthorized data is invisible to the caller. The full cascade is described under the Security Cascade.

The console gives you the operator view: which partners are active, what each partner’s scope is, how many queries they ran this week, and which queries returned what. You revoke an API key from one place. The partner stops seeing the corpus the moment the key is revoked.

The third endpoint is the MCP surface that AI agents consume. Your advisor copilot, your merchandising planner, your customer-facing assistant: each gets a session ticket with its own scope, and each calls the same MCP tools to read the corpus. The tools are query_facts, traverse_relations, pivot_dimension, replay_as_of, and cite. The MCP hides the dual-store complexity. The agent does not have to choose between graph traversal and SQL.

Every retrieval an agent makes comes back as structured rows with citation back to the source chunks. An agent cannot ground a recommendation on a fact whose validity window does not include today. The bi-temporal model blocks it at the retrieval step. An advisor copilot recommending “skip the foundation, try the serum” is grounded in facts the platform can defend on the date the recommendation is made.

Brand partners can wire their own agents to this surface under their scope. A brand-partner planning agent calls the same MCP tools the dashboard calls, sees only that partner’s slice, and gets the same citation chain. Your team configures the partner’s agent scope in the console.

Every query against every endpoint writes a row to the audit trail: the caller, the scope, the tools invoked, the chunks cited, the timestamp. The trail is append-only. When a partner asks “what did your data say about our category six months ago, when we made the Q2 buying decision,” you can replay the exact query against the corpus as it stood on that date. The bi-temporal model makes the replay correct. The audit trail makes the replay defensible.

This property is what turns the corpus from a deliverable into infrastructure a brand partner can underwrite against. A research deliverable cannot answer the replay question. A static dashboard cannot answer the replay question. The Intelligence Layer can.

From the operator console, your team controls three things per endpoint. First, who has access: which internal users can read which dashboards, which partner keys are active, which agent service identities are allowed in. Second, what they see: the segment scope per consumer, the dimension allowlist, the temporal window. Third, what they used: the audit view of every query that ran this week, with chunk citations included.

The console does not require platform engineers to operate. The OAuth flows, the scope minting, and the audit pulls are all surfaced as operator-level controls. If your IT team wants to integrate the platform with your SSO, your SIEM, or your existing API gateway, the integration points are in the engineering reference, available on request.