---
title: Knitli — Correctness infrastructure for classification markings
description: Knitli builds deterministic tools that compute, validate, and correct classification markings. Marque makes the marking rules executable.
url: "https://knitli.com/"
type: static
generatedAt: "2026-09-01T17:07:33.554Z"
---

# The classification system was never wrong.It asked too much of the people running it.

Our Mission: Make the classification system **move at the speed of decision**
[downpayment: Marque](/marque)
The proof

## Markings, computed.

Marque makes the classification-marking decisions that follow from the governing rules. In correction workflows, Marque **identifies the correct marking**, **cites the governing rules** by reference and line, and either **inserts the correction** in place or gives the implementation the information to decide on a safe fix.

The decision about what the information means—and how it should be classified—stays upstream.

Document · banner and two portions2 rules appliedBannerCONFIDENTIAL//RDCONFIDENTIAL//RD//NOFORNDERIVEDPortion(C//RD)(C//RD//NF)Portion(C//RD//REL TO USA, GBR)NEEDS REVIEW
DOE must approve release of RESTRICTED DATA. You marked this portion RESTRICTED DATA and releasable to GBR. This marking indicates there is a DOE sharing agreement with GBR under the Atomic Energy Act, and DOE approved it for release. If you're unsure that is true here, contact the originator.
CAPCO-2016 §H.6 p104RESTRICTED DATA defaults to NOFORN. Only the Department of Energy may release RESTRICTED DATA under an Atomic Energy Act compliant sharing agreement. With no release statement, RD defaults to NOFORN.CAPCO-2016 §D.2 p28Rollup missing NOFORN. You must include the most restrictive portion markings on the page in the banner. Marque regenerated your banner to include NOFORN.
***CONSTRUCTED EXAMPLE**It's unclassified. [About this example](/marking-examples/).*
DETERMINISTICRULE-DRIVENFREE TEXTAUTO-FIXOPTIONAL
The marking problem

## Automation needs an answer key.

**Historical markings are a weak answer key for training and evaluating automated systems.**They show what someone marked, not whether the marking complied with the rules in force at the time. Decades of inspector general and GAO reports tell the same story: **historical markings often do not align with those rules.**

A person or upstream system supplies the classification decision or proposed marking, sometimes using AI or other probability-based methods.

**Marque is the deterministic marking-correctness layer:** it *checks*, *corrects*, and *normalizes* the markings and labels those inputs require.

Across structured and unstructured documents and data, Marque computes the markings the governing rules require and returns a reproducible, rule-cited result with an audit record. It automates the rule consequences of known inputs; the semantic classification determination remains upstream.

## On the record

> “Deficiencies in the current classification system **undermine our national security**, as well as **critical democratic objectives**, by **impeding our ability to share information** in a timely manner.”
> **DIRECTOR OF NATIONAL INTELLIGENCE**Avril Haines
> Letter to Sens. Wyden and Moran, 5 January 2022

## What the system costs
84%
of sampled DoD documents contained marking errors
GAO-06-706112
discrepancies per 100 documents
ISOO FY2017$18.39B
government spend — the last published year, excluding the six IC agencies
ISOO FY2017On the ISOO figureISOO suspended collection of these statistics in a 21 September 2018 memo. FY2017 is the last year ISOO published cost estimates, which ***exclude* CIA, NSA, NRO, NGA, DIA and ODNI**. It is the floor.
## Finding errors adds work. Correcting them removes it.

Classification markings should enable the mission, not constrain it.

Marque helps people and automated systems get them right at speed and scale.
[Request a briefing](/contact)[About Marque](/marque)