# Product overview — what TrustFortress / Cubie is and why it matters

## The problem

Production AI fleets waste enormous capacity on work that should never have run.
Measured on real public datacenter traces (Azure LLM, Alibaba GPU):

- A single NVLink cascade deadlock can strand a large fraction of a GPU fleet — in
  the studied event, **66.85%** of fleet capacity was frozen while a hypervisor saw
  "healthy" utilization.
- On the Azure-2024 trace, **~99.5%** of requests produced fewer output tokens than
  they consumed as input — work whose cost exceeded its product.

The root cause is that trust and safety decisions happen **too late** — after the
GPU cycle is committed, after the tool call fires, after the prompt is logged.
Traditional admission control (IAM/RBAC) answers "is this user allowed to run jobs?"
but not "is this hardware safe right now?" or "is this specific request authentic?"

## The idea: decide before you spend

**Cubie** is a **pre-inference admission gate**. It compiles the facts about a
request — who, what, where, when, why, how — into a compact 192-byte geometric
object (a "trust cube"), evaluates it in constant time, and returns a verdict —
**allow / allow-degraded / deny** — *before* the workload runs. Denials come with an
explainable code (which "face" failed), not a black-box score.

- **Geometric, not learned:** the decision is a deterministic evaluation over a
  54-cell topology (six faces × nine cells), with no floating point and no learned
  weights — so it is explainable and reproducible.
- **Formally specified:** the core predicates are stated in three independent proof
  kernels (Coq, Lean 4, Verus). Three separate checkers make a single soundness bug
  far less likely to slip through. (See [learnings.md](learnings.md) for the honest
  scope of what is proven vs. empirical.)
- **Silicon-rootable:** the gate can bind to hardware evidence (Intel TDX
  attestation, PUF device identity, NVLink/PCIe telemetry, eFuse) when available,
  and runs in a laptop-friendly mock mode when it isn't.

**TrustFortress** is the control plane around Cubie: identity and workload
attestation, a Merkle-chained audit trail, an MCP server for tool-call admission,
and EU AI Act evidence generation. It ships in three modes — local/on-prem,
Kubernetes, and Cloudflare Workers (edge).

## Who it's for

- **AI platform / inference teams** — stop paying for deadlocked or wasted GPU work; gate tool calls before they fire.
- **Regulated buyers (health, finance, public sector)** — an admission/attestation layer with EU AI Act Article-by-article evidence and a hardware-rooted audit trail.
- **Integrators / agents** — a simple admission API and a public MCP server to query and integrate.

## What makes it different

- **Pre-inference, not post-hoc:** the decision precedes the spend.
- **Explainable denials:** a small, named code table (not an opaque model) says *why*.
- **Formal backing:** predicates are machine-checked across three kernels.
- **Compliance as running code:** EU AI Act obligations map to concrete Rust crates that emit Annex IV technical documentation — not a static PDF promise (see the scope note below).

## Honest scope (read this)

We publish claims with their boundaries, not marketing absolutes:

- The **TEP 100/100/100 @ FAR=0** result is an **in-sample calibration identity**
  (the threshold is fit to the fault-free baseline); held-out generalization is not
  yet established.
- The **proof corpus** is a **mixed snapshot** — closed proofs *and* statement-form
  stubs; it is not a blanket zero-axiom certificate, and the specs are checked
  standalone (they do not yet formally verify the shipping runtime).
- The **agentic-waste** figures are re-measured from committed traces and scoped to
  the datasets named (the ~99.5% is the Azure-2024 trace).
- The **EU AI Act** crates are a compiling mapping/generation **scaffold** with some
  placeholder values — not a fully populated conformity pack.

Full evidence, pinned hashes, and reproduce commands: the **Grounded Claims
Registry** at <https://lib.trustfortress.ai/claims> and [../CLAIMS.md](../CLAIMS.md).

## Where to go next

- Questions → [FAQ.md](FAQ.md)
- Integrate / build → [agent-integrator-guide.md](agent-integrator-guide.md)
- Deploy / operate → [admin-onboarding.md](admin-onboarding.md)
- The evidence behind the claims → [../CLAIMS.md](../CLAIMS.md)