Skip to content

tymofiy/kp

Man sieht nur, was man weiß.
We see only what we know.
— Goethe

You see that the sun rises and sets and therefore you think you know. You don't.
You only know when you can answer the question: How or what for?
— Barenboim

An AI knows only what we tell it.
KP:1 is for telling it whole.
— KP:1

KP:1 -- Knowledge Pack Format

KP:1 is a format that doesn't flatten knowledge. Uncertainties stay uncertain, contradictions stay in tension, and the trace of how beliefs evolved is kept as structure. AI-first, human-accessible.

Editor: Timothy Kompanchenko Status: Editor's Draft — KP:1 Public Draft — 2026-04 (v0.7-preview) See also: spec/CORE.md, spec/SPEC.md, GOVERNANCE.md, CONTRIBUTING.md

What is KP:1?

Think about what anyone actually knows about a real artwork: contradictory accounts of provenance, overlapping stories that reinforce each other, uncertain events, details that mattered ten years ago and matter less now. A lot of it is unfinished, probabilistic information. The value of the knowledge comes from that texture — the uncertainty, the tension, the trace of how beliefs evolved. Flatten it — average it out into a list of settled facts — and you've destroyed what made it knowledge.

That's the difference between knowledge and data. Data answers yes or no. Knowledge has the shape of belief — its uncertainties, its tensions, its history. KP:1 is a format that keeps the shape.

Built AI-first. Every AI reading knowledge from prose today hits the same problem: amnesia. Between sessions, between context windows, between tools, the model loses the thread and has to reconstruct from prose what's being asserted, how certain each point is, where the contradictions are, which beliefs superseded which. That reconstruction is lossy, expensive in context window, and unreliable. KP:1 is designed to fix that. It's a structured format for AI to read, reason over, and carry forward between sessions — handing the model what's believed, how strongly, on what evidence, in tension with what, and how those beliefs evolved. No guessing. No re-derivation. No reconstruction from prose.

Human-accessible. A knowledge pack is a directory of Markdown and YAML files. Humans read it, print it, edit it, and work with it using ordinary text tools. But the primary consumer is AI — the structure exists to let AI reason without reconstructing from scratch. This is the same wave as llms.txt and AGENTS.md — formats built for AI to read directly — applied to knowledge itself.

Quick Start

Read spec/CORE.md -- the implementable Core specification. It covers pack structure, manifest schema, claim syntax, evidence, confidence, relations, and validation rules. Everything you need to build a conformant parser.

A claim in KP:1 looks like this:

- [C001] Cost decline is structural, not cyclical
  {0.95|i|E001,E002|2026-03-01|exhaustive|judgment}
  Learning curve has held for 40 years. →C002, ⊗~C003

Each claim has an ID, an assertion, a confidence/type/evidence block, and optional context with relations to other claims.

Repository Structure

Directory Purpose
spec/ Normative specification -- CORE.md, SPEC.md, and companion documents
conformance/ PEG grammar, JSON Schema, and 10 test fixtures
examples/ Two complete .kpack examples
positioning/ Public-facing positioning and design rationale
research/ Benchmark design and prior art analysis
reference/ Reference parser and tooling (planned)
decisions/ Design decision records
scripts/ Git hooks and validation helpers

Top-level governance and policy files include GOVERNANCE.md, CONTRIBUTING.md, CODE_OF_CONDUCT.md, LICENSE, LICENSE-CODE, and DCO.txt.

Examples

Two complete Knowledge Packs demonstrate the format:

  • Solar Energy Market -- Market analysis with cost trajectories, technology trends, and regional adoption. Shows dense claim syntax, confidence levels, evidence references, and inter-claim relations including contradictions.

  • KP External Assessment -- A self-assessment of the KP format itself. Demonstrates meta-level claims and the format's ability to describe its own uncertainties, tensions, and open questions about its own design.

Conformance

The conformance suite provides formal validation tools:

  • PEG grammar (conformance/grammar/kp-claims.peg) -- parseable definition of claims.md syntax
  • JSON Schema (conformance/grammar/kp-pack.schema.json) -- validation schema for PACK.yaml manifests
  • 10 test fixtures -- 5 valid packs that must be accepted, 5 invalid packs that must be rejected with specific errors
  • 2 complete example packs -- the kpacks in examples/ are validated by the runner as part of the suite, so the live examples are themselves conformance tests

The runner (conformance/run.py) reports 12/12 passed on a conformant implementation: 10 fixture tests + 2 example validations. A conformant implementation parses all valid fixtures, rejects all invalid ones, validates PACK.yaml against the schema, and enforces semantic constraints SC-01 through SC-11. See conformance/README.md for details.

Interoperability

KP:1 has its own syntax and semantics. spec/MAPPING.md provides a field-by-field translation to RDF/JSON-LD, PROV-O, and Nanopublications — grading each mapping as clean, lossy, or impossible, so practitioners using existing semantic web toolchains can assess what they gain and what they lose. Notably, KP:1's distinction between unqualified contradiction (⊗), error-contradiction (⊗!), and informative tension (⊗~) has no direct equivalent in any of these standards — it is one of the format's genuinely novel contributions.

Status

This is an editor's draft maintained by a single editor in a public repository. It is published as KP:1 Public Draft — 2026-04 (git tag v0.7-preview). It has a formal grammar, a JSON Schema, a conformance suite with 10 test fixtures, and two reference examples.

The specification is not final and may change in any way at any time, including breaking changes. It is not yet ratified by any standards body. Compatibility commitments will arrive only with a non-draft version. See GOVERNANCE.md for the full governance picture, including how decisions are made during the preview phase and what changes when the Knowledge Pack Foundation is incorporated.

The current phase is feedback-only: the editor welcomes issues, comparisons, ambiguity reports, and adversarial review through GitHub issues, but does not accept external pull requests modifying normative spec text. See CONTRIBUTING.md for details.

License

KP:1 is published under two licenses:

  • Specification text (everything in spec/ and the prose portions of this README) is licensed under the Creative Commons Attribution 4.0 International License (CC-BY-4.0). You may share and adapt the material for any purpose, including commercially, with attribution.
  • Code, schemas, and examples (everything in conformance/, examples/, scripts/) is licensed under the Apache License 2.0, which includes an explicit patent grant.

See CONTRIBUTING.md for the contribution policy and GOVERNANCE.md for governance details.

How to Cite

If you reference KP:1 in academic, technical, or evaluative work, please use the metadata in CITATION.cff. The v0.7-preview release is published on Zenodo with DOI 10.5281/zenodo.19445263. The recommended short form is:

Kompanchenko, T. (2026). KP:1 — Knowledge Pack Format Specification (Version 0.7-preview). Zenodo. https://doi.org/10.5281/zenodo.19445263

The editor and an acknowledgment of AI drafting assistance are also recorded in ACKNOWLEDGMENTS.md.

Trademarks

KP:1™ is a pending United States trademark. "Knowledge Pack" is the descriptive name of the format and is not currently a registered or pending trademark. See GOVERNANCE.md for the conformance and trademark use policy.

About

KP:1 — text-native specification for representing epistemic state in AI workflows

Topics

Resources

License

CC-BY-4.0, Apache-2.0 licenses found

Licenses found

CC-BY-4.0
LICENSE
Apache-2.0
LICENSE-CODE

Code of conduct

Contributing

Stars

Watchers

Forks

Packages

 
 
 

Contributors