Managed inference, with a conscience about what it retrieves.

1 · The inference gateway

Every model call in the platform flows through a single gateway that exposes capability tiers — fast, standard, frontier, utility, image — never vendor model names. Swapping a backing model is a configuration change gated by an executable CI evaluation contract: required tiers must exist, fallback chains must be legal, embedding dimensions must match. Fallback is tier-degrade only (frontier to standard to fast) — never a silent hop to a different vendor.

Today the tiers are backed by Azure AI Foundry serverless deployments. Because applications only know tier names, the same platform can be re-pointed at on-premise serving as our edge hardware program matures — with zero application changes.

2 · Truth-value retrieval

Retrieval-augmented generation is only as good as what it retrieves. Zimax classifies every document at ingestion — Authoritative (human-asserted, audited), Working (current operational material), or Observed (unvetted input) — and retrieval scoring multiplies raw relevance by truth class and a time-decay factor. Supersession detection demotes outdated versions on a half-life rather than deleting them, and the machine can never promote a document to Authoritative: that takes a human, a written reason, and an append-only audit trail.

Search itself is hybrid: dense vector similarity fused with native BM25 keyword scoring, plus an entity graph for relationship queries — with truth values inherited along graph paths, so a chain of reasoning is only as trustworthy as its weakest link.

This layer exists because we lived the failure it prevents: our own platform once served its founders' superseded design documents as confidently as current ones. We built the fix into the core.

3 · Isolation enforced by policy

Every tool call crosses a policy engine that fails closed. The tenant in the verified token must match the tenant in the request. Vector search filters are injected server-side — never trusted from the client. Clearance is computed as the maximum of caller and document, so a permissive caller cannot unlock a restricted document. Every allow and deny is written to a decision log.

4 · Two regions, one platform

The platform runs live in a US region and a Mexico region — same images (pinned by content digest), same declarative configuration, region differences expressed as reviewed overlay patches. Data residency is a first-class construct: each region has its own identity provider, its own key vault, and its own backup chain.

5 · The edge inference roadmap

Some inference belongs close to the data. Our Tier 0 edge program runs speech-to-text, reranking, and embedding workloads on compact on-premise hardware — keeping raw audio and document content inside the customer boundary while cloud tiers handle frontier reasoning. One complete unit is in validation today; it is a proving ground, stated as exactly that.