Skip to main content

KLM — Knowledge Layers Model

An open reference standard for attestable knowledge in language-model systems.

A language model answers, and you believe it or you don't. KLM exists to make a third option possible: read the answer back — where its knowledge entered, what it did, how it was judged — with every signal carrying an honest epistemic status. The deliverable is no longer the text; it is the text and its attestation.

Show me where this came from. Everything else follows from taking that question seriously.

The documents

Document Role
KLM-Manifesto.md The thesis, in nine commitments
KLM-White-Paper.md The argument, the layer model, a worked example (v0.3)
KLM-Specification.md Normative conformance detail, RFC-2119 (v0.1)
docs/KLM-Gamma-Unification.md Normative γ schema decisions (klm-gamma/1.0)
docs/KLM-Amendment-Attested-Parametric.md Draft amendment: mechanical parametric attribution
docs/QUICKSTART.md Five minutes: validate, try to cheat, sign, chain, disclose
docs/FAQ.md What KLM is and deliberately is not
CONTRIBUTING.md Amendment process + implementation reports (the gap we most want filled)

The toolchain (schema/)

Deliberately stdlib-only Python — any third party verifies with nothing but a Python interpreter. No dependencies, no accounts, no trust in the emitter.

Tool What it does
klm-gamma.schema.json + validate_gamma.py The γ (epistemic vector) contract + validator
klm-attestation.schema.json + validate_attestation.py The attestation record + the KLM-0…5 conformance ladder
sign_attestation.py (+ ed25519_ref.py) Ed25519-signed envelopes; RFC 8032 fallback so verification needs zero deps
attestation_log.py Hash-chained append-only record log (tamper breaks the chain)
disclosure.py Selective disclosure views: user / auditor / operator (§12.1)
schema/examples/ Golden records — including adversarial ones that MUST fail

Quick start:

python3 schema/validate_attestation.py --level 3 schema/examples/attestation-worked-example.json
python3 schema/validate_attestation.py schema/examples/invalid/attestation-violations.json  # must FAIL

Conformance levels

KLM-0 Declared → KLM-1 Traceable → KLM-2 Grounded → KLM-3 Reflective → KLM-4 Attributable → KLM-5 Verifiable. Progress is not a feeling; it is passing the next level's test. See the Specification §8.

Status & invitation

KLM is a candidate standard: implemented end-to-end by a reference stack (SEDIM / Nage), with a cross-language second emitter passing the same validator — but its real test is a conforming implementation built on entirely different mechanisms, by someone else. Competing implementations, challenges to the conformance suite, and layers we got wrong are the point, not a threat.

License

Apache-2.0. Stewardship: Nage AI.

Metadata

Release files for klm-conformance 0.1.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for klm-conformance 0.1.0
File Size Uploaded
klm_conformance-0.1.0.tar.gz 76.2 kB Details

Built distribution (wheel)

Table of built distributions (wheels) for klm-conformance 0.1.0
File Interpreter ABI Platform
klm_conformance-0.1.0-py3-none-any.whl Python 3 none any Details

Total release size: 111.3 kB

Release files / klm_conformance-0.1.0.tar.gz

Download URL klm_conformance-0.1.0.tar.gz
Size 76.2 kB
Tags Source
SHA-256 checksum
How to use checksums
9d7a491a815a2e849a5a27afc64d894bb081817063d5c524bc39aba35d634fe8
BLAKE2b-256 checksum
How to use checksums
10b318e0cff0aa4d8e5c96d1984f131e3835ea08a175c701a093787e6bf5cc49
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.9.6

Release files / klm_conformance-0.1.0-py3-none-any.whl

Download URL klm_conformance-0.1.0-py3-none-any.whl
Size 35.0 kB
Tags Python 3
SHA-256 checksum
How to use checksums
166d97ce0c5a11eec9591f05b070af83f465f5545d6e48e314279566a7f16e25
BLAKE2b-256 checksum
How to use checksums
f151350dcb62ae3bb3d1e4d88574665d78f425417cfe7c8adde0aed7c160ca40
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/6.2.0 CPython/3.9.6

Release history Release notifications | RSS feed

0.3.0

2 release files

This release

0.1.0 This release

2 release files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page