# TrustFortress / Cubie — Knowledge Base

The single place to get questions answered about TrustFortress and Cubie —
for **coding agents / integrators**, **onboarding admins**, and **customers**.
Everything here is public. Ask it in natural language:

- **Semantic search:** `GET https://lib.trustfortress.ai/api/ai-search?q=<question>`
- **Grounded answer:** MCP `ask_library` at `https://lib.trustfortress.ai/mcp` (see [MCP.md](../MCP.md))
- **Browse:** <https://lib.trustfortress.ai/browse/cubie-tf%2Fknowledge-base%2F2026-07-31%2F>

## Start here by role

| You are a… | Read |
|---|---|
| **Customer / evaluator** | [product-overview.md](product-overview.md) → [FAQ.md](FAQ.md) → [learnings.md](learnings.md) (honest scope) → [../CLAIMS.md](../CLAIMS.md) (evidence) |
| **Coding agent / integrator** | [agent-integrator-guide.md](agent-integrator-guide.md) → [workload-admin-evidence.md](workload-admin-evidence.md) → [../MCP.md](../MCP.md) → [FAQ.md](FAQ.md) |
| **Onboarding admin / operator** | [admin-onboarding.md](admin-onboarding.md) → [workload-admin-evidence.md](workload-admin-evidence.md) → [../AI_GATEWAY.md](../AI_GATEWAY.md) + [../AI_SEARCH.md](../AI_SEARCH.md) |

## What TrustFortress / Cubie is (one line)

A **pre-inference admission gate** for AI workloads: a formally-specified,
constant-time geometric trust decision (Cubie) that decides *allow / degrade / deny*
**before** a GPU cycle or tool call is spent — wrapped by the TrustFortress control
plane (identity, audit, MCP, EU AI Act evidence). See [product-overview.md](product-overview.md).

## The four headline claims (with honest scope)

1. **Triple-kernel proof corpus** (Coq + Lean 4 + Verus) — a *mixed-snapshot* corpus, not a blanket zero-axiom certificate.
2. **TEP fault detector** — 100/100/100 FDR at FAR=0 as an **in-sample calibration identity** (held-out transfer not yet established).
3. **Agentic waste** — 66.85% blast radius / ~99.5% output<input, **re-measured** from committed traces (scoped to Azure-2024).
4. **EU AI Act coverage** — a compiling Annex IV mapping *scaffold*, not a fully populated conformity pack.

Full scope, evidence, and reproduce steps: **[../CLAIMS.md](../CLAIMS.md)** and <https://lib.trustfortress.ai/claims>.

## Documents in this knowledge base

- **[product-overview.md](product-overview.md)** — what it is, the problem, how it works, why it matters (customer).
- **[FAQ.md](FAQ.md)** — the most common questions, answered.
- **[learnings.md](learnings.md)** — honest scope boundaries, proven-vs-in-sample, what to reproduce.
- **[agent-integrator-guide.md](agent-integrator-guide.md)** — repo/crate map, public APIs, build/run, how to connect.
- **[admin-onboarding.md](admin-onboarding.md)** — deploy, identity, provision AI Gateway / AI Search, go-live.
- **[workload-admin-evidence.md](workload-admin-evidence.md)** — how to interpret ontology/compiler replay, CUB read-back, embedded receipts, runtime observations, exact-head certification, and owner authorization without collapsing them into one claim.
- Also indexed here: the Grounded Claims Registry ([../CLAIMS.md](../CLAIMS.md)), [MCP](../MCP.md), [AI Gateway](../AI_GATEWAY.md), [AI Search](../AI_SEARCH.md) guides, and the getting-started + demo docs.

> **Scope of this KB:** public product, usage, integration, compliance posture, and honest learnings. Deep repo internals, hot-path engineering constants, patent-detailed mechanics, and partner-confidential (Intel CNDA) material are **not** public — they live behind an authenticated tier (planned).
> **Patents:** US 11,956,363 B2, US 12,073,439 B2, US 2024/0195628 A1. TrustFortress/Cubie is proprietary (Centillion / TrustFortress.ai).