About the classification-marking examples on this site.

No. Nothing on this public site or in any Knitli public material is classified.

Most examples show a marking by itself. When an explanation needs document-like context, we construct that context from scratch and use fictitious topics. The North Pole Ministry of State Security is not a real spy agency.

The rules and syntax for classification markings are public. Some control names, compartment names, and special access program markings are classified or otherwise not publicly available. We do not use them. Every example here is supported by a public source or uses an obvious public stand-in, such as the imaginary BUTTER POPCORN special access program in the CAPCO Manual.

None of these examples comes from customer, operational, or classified material.

The simple answer is that Marque is a classification-marking tool. We cannot demonstrate whether it gets markings right with fake markings.

Using real markings makes the problems visible: how the rules interact, how a small change can alter a result, and why a plausible-looking marking can still be wrong. They let us show exactly what Marque changes and why.

Public marking guidance also makes independent improvement possible. Without it, small innovators like Knitli could never get started. We believe the CAPCO Register and Manual should be public by default. It currently requires a FOIA request.

The CAPCO Register and Manual is the primary source for Marque's marking rules and most examples on this site, but it is hardly the only public source. Public classification-marking guidance appears in regulations, manuals, training courses, forms, archived technical specifications, and congressional research.

Public sources we use include:

• The CAPCO Register and Manual, historical versions of the manual, and related guides.
• ODNI's Information Security Marking XML specifications and ICD 710.
• NARA's classification-marking FAQ and classification-marking booklet.
• DoD Manual 5200.01, Volume 1, and CDSE derivative-classification training.
• The State Department's Foreign Affairs Manual guidance and GSA classification cover sheets.
• 32 CFR Part 2001 and Executive Order 13526.
• The Congressional Research Service's overview of sensitive compartmented information.

It is easy to underestimate how much of this material is public when you are accustomed to accessing them on classified networks. But they are on the open internet, too.

No. Marque is deterministic software. It contains no models, trains on nothing, and has no training data.

That was a choice, not a limitation. Before Marque, we built CodeWeaver, a probabilistic NLP and hybrid-search system. We know how to build with models. We did not use one here because a model was the wrong tool for this job.

Marque's job requires reproducible results, complete explanations, and a traceable path from a marking to the governing rules and original classification authority. Marque is deterministic by design.

Knitli has no plans to operate Marque as a hosted service. Marque runs in the customer's environment. It can be integrated locally into an application, browser, or web app; run as a REST API within that environment; or used directly from the command line.

The data belongs to the customer and stays in the customer's environment, whether it is classified or unclassified.

Marque does not decide whether information should be classified. A person or authorized upstream process supplies the controlling facts. Marque computes the marking consequences those facts require: whether a marking follows the applicable rules, how multiple markings roll up, and what corrections those rules require.

That job is larger than it looks. The CAPCO Register and Manual is roughly 200 pages, and its rules interact across hundreds of markings. Marque applies those relationships in around a millisecond.

We know NLP is part of the larger solution. A probabilistic system can sit upstream of Marque, or a training or evaluation pipeline can use Marque's deterministic outputs as an answer key. In either case, Marque remains separate. It does not train models, and it does not learn from the records it checks.

A person or probabilistic system may supply or recommend the controlling classification inputs. Marque computes the marking consequences of the inputs it receives.

No. Marque will remain deterministic. Knitli may build or use probabilistic systems for other parts of the classification problem, but they will remain separate from Marque. Marque stays the transparent, verifiable grounding layer.