> ## Documentation Index
> Fetch the complete documentation index at: https://docs.inviolet.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# The 5 Principles

> The five durable promises Inviolet keeps — the doctrine behind every decision the platform makes.

The five durable promises Inviolet keeps. They come into force in order as you
climb the [maturity model](/concepts/maturity-model) — one principle per level —
and the first, intent, is the foundation the other four stand on.

## 1. Learn the intent behind every prompt.

<Note>
  **Status:** shipped  ·  **Tier:** violet  ·  **Fulfilled by** Phase 4  ·  **Maturity level:** Observe
</Note>

Every request an agent makes carries an intention — a labeled, human-owned
purpose the policy layer can evaluate. Inviolet captures the intent behind every
prompt: the signal identity and PAM can't see, and the foundation the other four
principles stand on. The more traffic it sees, the sharper it gets.

**Catches:** Agents taking an action nobody modeled — unauthorized data reads,
unapproved writes, purposes nobody described.

## 2. Test before enforcing.

<Note>
  **Status:** shipped  ·  **Tier:** violet  ·  **Fulfilled by** Phase 3  ·  **Maturity level:** Test
</Note>

Every rule can run in shadow first. The decision is logged as if enforced; the
actual request still goes through. Tune against real traffic, then flip the
switch.

**Catches:** Over-enforcement — policies that would block legitimate work if
deployed live without tuning.

## 3. Least access as the system learns.

<Note>
  **Status:** shipped  ·  **Tier:** deep violet  ·  **Fulfilled by** Phase 7 + Phase 8  ·  **Maturity level:** Enforce
</Note>

Grant the least access an intention needs — specific columns under specific
conditions, never a whole data source. Route the gray zone to a human, and turn
each decision into standing policy. Access tightens as the system learns.

**Catches:** Blanket connection-level grants that expose columns outside the
intended scope, and the "permit because we don't know" silent yes.

## 4. Agents never touch credentials.

<Note>
  **Status:** in flight  ·  **Tier:** ultraviolet  ·  **Fulfilled by** Phase 9 + Phase 16  ·  **Maturity level:** Isolate
</Note>

What reaches the agent is never the real key. Inviolet brokers a short-lived,
intention-scoped lease against your existing PAM — the credential stays in your
vault, the agent gets a one-shot token bound to a single declared intention.

**Catches:** Stolen, leaked, prompt-injected, or memorized credentials — the
entire class of attacks that depend on the agent ever holding the actual
secret.

## 5. Impossible to bypass.

<Note>
  **Status:** in flight  ·  **Tier:** ultraviolet  ·  **Fulfilled by** Phase 19  ·  **Maturity level:** Surround
</Note>

Every path an agent could take is covered — direct egress, every LLM host, every
runtime — and every one is proven: bypass-tested in CI, monitored in your SIEM,
attestable to an auditor. There is no way around the system.

**Catches:** The agent that skips the gateway entirely — a direct API call, an
unmediated host, a runtime nobody wrapped.

## Read next

For the full narrative on each principle, see the public
[principles page](https://inviolet.ai/principles). Each principle maps onto one
of the five levels of the [maturity model](/concepts/maturity-model).