Open Standard · Spec v1.0.0-draft.1

Knowledge your agents can actually trust.

DKP is an open standard for packaging curated domain knowledge into validated, structured bundles - so both Human and AI agents stop hallucinating and start knowing, while human agents can explore knowledge in various ways.

Get the Spec Get the CLI Browse the Directory
0 Bundle Layers
0 Quality Gates
0 Trust Tiers
The Problem

AI agents are only as good as their knowledge.

Most knowledge fed to AI today is unstructured - raw documents, loose notes, ad-hoc context dumps. The result: hallucinations, missed facts, and outputs no one can trust or audit.

Without DKP

  • Unstructured blobs with no schema or versioning
  • No provenance - where did this come from?
  • Unknown quality, no way to measure improvement
  • No built-in eval sets to catch hallucinations
  • Every team reinvents the same ingestion pipeline

With DKP

  • Validated 6-layer bundle with machine-readable schema
  • Full provenance chain - sources, rights, editorial sign-off
  • 8-gate quality certification with deterministic checks
  • Built-in eval sets with measurable quality delta
  • One standard format any MCP client or RAG pipeline loads
How It Works

From raw knowledge to trusted AI context - in three steps.

  1. Create

    Run dkp init @my-organization/my-pack --domain <category> to scaffold the complete 6-layer bundle structure. Every required file is generated from validated templates - schema-correct from day one. Or run dkp new to go further: it scaffolds the bundle and then uses an LLM to generate the machine and human layers, export the OKF layer, validate, and package - all in one command.

  2. Validate

    Run dkp validate to pass all 8 quality gates automatically. Then dkp eval runs your built-in test suite and reports a measurable quality delta. If eval cases fail, dkp fix rewrites the failing retrieval chunks - grounded on the failure reasons - and regenerates the eval set against the fresh corpus.

  3. Deploy

    Run dkp build && dkp sign && dkp publish to produce a cryptographically signed, checksummed archive consumable by any MCP client, RAG pipeline, or LLM application.

Features

Everything a production knowledge pipeline needs.

DKP handles the structure, quality, and delivery layers so you can focus on the knowledge itself.

6-Layer Bundle Structure

Every concern in its own layer: machine/ for AI-ready assets, okf/ for OKF compatibility, human/ for docs, evidence/ for provenance, skills/ for SKILL.md skill files, and l10n/ for localization.

8-Gate Quality Standard

Three certification tiers - Conformant, Evaluated, and Reviewed - with deterministic automated checks and optional human editorial attestation backed by dated sign-off in evidence/review_notes.md.

Knowledge Graph

Typed edges between concepts - requires, contradicts, elaborates, supersedes, depends-on - enable GraphRAG-style multi-hop retrieval that chunks alone can't achieve.

Built-in Eval Sets

Ship Q&A test cases inside the bundle. dkp eval runs them against any model and returns a quality delta - a measurable, reproducible score that travels with the pack.

Supply-Chain Integrity

SHA-256 checksums on every file, bundle-level Ed25519 signing, and a manifest-level SBOM. Processors verify signatures against the publisher's PGP fingerprint before loading.

MCP First-Class

Any DKP bundle exposes standard Model Context Protocol resources and tools - inject, search, chunk, get - so any MCP-compatible agent can mount it with zero custom integration.

Audience Scoping

Tag content for named audience profiles - consumer vs. clinician, analyst vs. executive. The --audience flag filters every output layer so each agent or reader sees only the content meant for them.

WASM Procedures

Bundle executable logic alongside declarative knowledge in machine/procedures/. Agents invoke sandboxed WASM/WASI modules for deterministic computation - no external calls, no side effects.

OKF Compatible

DKP is a strict superset of the Open Knowledge Format - every bundle contains a fully conformant OKF bundle in its okf/ layer. Any OKF-native agent framework can load and traverse that layer without modification, and without any lock-in to DKP tooling.

Bundle Anatomy

Every layer has a job. Nothing is mixed.

A DKP bundle is a directory (or a compressed archive) with a strict layout that every producer writes and every processor reads the same way.

Seven asset types in machine/

  • DomainTerm Glossary entry with aliases and relationships
  • DomainRule Operational must-do / must-avoid statement
  • Constraint Edge case, anti-pattern, or hard limit
  • DecisionProcedure Traversable decision tree in JSON
  • KnowledgeChunk Self-contained fact optimized for RAG
  • EntityType Domain entity with typed attributes
  • EvalCase Q&A pair for measuring pack quality
Quality Certification

Three tiers. Eight gates. No ambiguity.

The 8-gate quality standard defines exactly what "good" means for a DKP bundle - with automated structural checks and optional human editorial attestation.

Fully Automated
DKP-Conformant

The baseline. Your bundle passes structural validation and is ready for machine consumption.

  • Schema-valid JSON/JSONL assets
  • All required layers present
  • Manifest complete and parseable
  • Knowledge graph references resolve
  • OKF concept frontmatter valid
  • Checksums verified
dkp validate
Automated + Eval
DKP-Evaluated

Conformant plus a measurable quality delta. Your bundle ships with proof it actually helps.

  • All Conformant gates pass
  • Eval set present in machine/
  • Quality delta > configured threshold
  • Retrieval chunks tested
  • No hallucination markers
  • Eval report saved to evidence/
dkp validate && dkp eval
Human Sign-Off Required
DKP-Reviewed

The gold standard. Every gate passed, every claim traced, every layer reviewed and attested.

  • All Evaluated gates pass
  • Relevance and originality verified
  • Human usability confirmed
  • Cross-layer consistency checked
  • Dated editorial sign-off
  • Ed25519 bundle signature applied
dkp release-check
Command-Line Interface

One tool. The entire lifecycle.

The dkp CLI covers everything from scaffolding your first bundle to signing and publishing to a registry - the full knowledge engineering workflow.

$ dkp init @my-organization/medical-protocols --domain Health
Scaffolding 6-layer bundle... ✓
Generated machine/rules.json ✓
Generated machine/graph.json ✓
Created machine/eval_set.jsonl ✓
$ dkp validate
[1/8] Relevance pass
[4/8] Machine Usability pass
[8/8] OKF Conformance pass
Tier: DKP-Conformant
$ dkp eval
Questions: 42 Passed: 41 Delta: +0.23
Tier: DKP-Evaluated
$ dkp sign --key ./signing.key
Ed25519 signature applied ✓
$ dkp publish
Published medical-protocols@1.0.0

Authoring

init new generate

Inspection

info list get search chunk inject

Validation & Eval

validate eval diff fix review

Build & Sign

build sign keygen release-check

Registry

publish install update yank registry

Serve & Explore

serve tui webui prompt graph okf
Install the CLI
Explore

Four ways to get inside a bundle.

Whether you want a quick terminal view, a browser UI, or an LLM conversation grounded on the pack, dkp has a mode for it.

dkp tui

Terminal UI

A full-screen terminal explorer with five panels - Assets, Search, Chunks, Eval, and MCP - so you can inspect every layer of a bundle without leaving your shell.

dkp tui ./my-pack
dkp webui

Browser UI

Spins up a local web server and opens a browser SPA with full-text search, asset browsing, and an interactive knowledge graph visualizer - no install beyond the CLI.

dkp webui ./my-pack
dkp prompt

Grounded Chat

Ask questions against a bundle in single-shot or interactive REPL mode. The pack's system prompt, glossary, and retrieval chunks are injected as context - answers are grounded on the pack, not on general model knowledge.

dkp prompt ./my-pack "What are the key rules?"
dkp serve

MCP Server

Mount a bundle as a live MCP server over stdio or HTTP. Exposes inject, search, chunk, and get as standard MCP tools so any MCP-compatible agent or IDE can query the pack directly. When the bundle includes WASM procedures, list_procedures and run_procedure are exposed as additional tools - letting agents invoke sandboxed deterministic logic that ships inside the pack alongside the knowledge it operates on.

dkp serve ./my-pack --transport http
Who It's For

Built for every role in the knowledge pipeline.

Knowledge Engineer

Package your expertise for AI.

You've spent years building domain knowledge. DKP gives you a structured authoring format with validation, versioning, and quality certification - so your expertise survives hand-off to AI systems and remains traceable, auditable, and updateable over time.

dkp init dkp validate dkp review
AI Developer

Skip the preprocessing. Start with structure.

Building a RAG pipeline or LLM app? Mount a DKP bundle via MCP and get structured, validated, citable knowledge with zero ingestion work. Retrieval chunks, system prompts, and eval sets are already formatted for your pipeline.

dkp serve dkp inject dkp eval
Enterprise Team

Deploy AI with audit trails and access control.

Deploying AI at scale means provenance, compliance, and security. DKP's signed bundles, SBOM-style manifest, audience scoping, and access control fields give your security and legal teams exactly what they need for sign-off.

dkp sign dkp publish dkp registry

Ready to structure your knowledge?

Read the specification, install the CLI, or explore - everything you need to start packaging knowledge that AI agents actually trust.

Browse the Directory Install the CLI Read the Spec View on GitHub