Architecture/04

Ontic sits between your app and your model. Every claim is checked against your data. Evidence found: authorized. Evidence missing: blocked. Everything logged and signed.

How governance scales from shared SaaS to air-gapped hardware.

Without it

Every evaluation produces a signed record: what was asked, what was checked, what was authorized or blocked, and why. The record format is the same across all tiers — what changes is the enforcement locus and the attestation chain.

Ontic is

A gate that checks claims against real data
Deterministic — same input, same decision, every time
An audit trail you can hand to anyone

Ontic is not

A model (we don't generate anything)
A content filter (we don't judge tone or topic)
A monitoring tool (we prevent, not just detect)
A replacement for your data (we check against it)

Start where you are

The Workshop (Day 1): Add labeling and logging to your existing AI pipeline. No model changes. No data source setup.
The Refinery (Week 2+): Connect a data source. Define required fields. Turn on the evidence check.
The Clean Room (When you need it): Signed everything. Verified execution. For when a court or regulator might ask to see every step.

CFPO — prompt architecture, not prompt engineering

Compact models have no room for ambiguity. CFPO is the section-ordering convention that makes governance mechanically enforceable inside the prompt itself.

Deterministic structure — attention mechanisms weight position; rules arrive in a predictable order every time
Enforcement by contrast — paired ❌/✓ examples teach via demonstration, not instruction alone
Machine-parseable policy — YAML blocks can be programmatically audited, diffed, and versioned
Compiled, not authored — runtime templates assembled from fragments via the @ontic/prompts compiler (RFC-0004)
Every prompt is versioned in a typed registry with model, temperature, and change summary