Governed Models: How KYE Protocol™ Brings Authority to AI Models

Your model is on Hugging Face. Its authority isn't.

The Hub answers "what can this model do?" brilliantly — weights,
benchmarks, model cards, dataset lineage. It leaves unanswered the
questions that block regulated adoption:

• May you use this model in this jurisdiction, for this purpose, as
  this entity?
• Which obligations apply when you deploy it — EU AI Act, GDPR, DORA,
  SOC 2?
• Can you prove all of that to an auditor months later, from public
  artifacts?

KYE Protocol™ (Know Your Entity™) exists to answer that second set of
questions — and to prove the answers.

Capability and authority are different questions

Capability is what a model can do. Authority is whether a specific
entity may use it, here, now, for this purpose, under these
obligations. The Hub owns the first question. KYE Protocol™ owns the
second. A regulated team needs both before it can ship.

This is not a criticism of model cards. A model card is a capability
disclosure written by the publisher. An authority record is a governed,
evidence-backed answer about a use — bound to the entity using the
model, the purpose it is used for, and the framework obligations that
use carries. One artifact cannot honestly be both.

The AI Bill of Materials: a Chain of Authority

KYE Protocol™ binds a model to an AI-BOM — the dataset, weights,
and model provenance — and maps its use to the obligations each
framework carries:

Dataset ──▶ Weights ──▶ Model ──▶ KYE™ binding ──▶ Obligation map ──▶ Evidence Pack™
provenance   training     Hub       AI-BOM ·         EU AI Act · GDPR    Replay-Proof™ ·
· licence    lineage      artifact  authority        DORA · SOC 2        public keys

• Chain of Authority — KYE Protocol™ records which datasets trained
  which weights composed which model, under which licence.
• Obligation mapping — the model's use is mapped to the
  obligations it carries, drawn from one canonical registry — never
  ad-hoc per deployment.
• Replay-Proof™ evidence — every governance decision seals into an
  evidence pack an auditor can re-verify from public artifacts.

Datasets are governed assets, not just files

The Chain of Authority starts at the dataset — so dataset governance is
not optional garnish. In KYE Protocol™ a dataset is a governed data
asset, declared with the open kye.data_asset.v1 contract:
classification, per-field PII inventory, provenance, licence, retention.
And every use of it by an agent — read, summarise, export,
train_model, share, disclose — is checked at decision-time against
an active kye.data_use_manifest.v1: does the purpose match, is the
classification supportable, is the residency allowed, is the retention
window still open? Default-deny on every axis, one signed evidence
row per access.

That is the KYE Data Governance Pack™,
and it closes the question model governance alone cannot: was this
model even allowed to be trained on that data? The datasets in the
KYE-Protocol org dogfood the
posture — each is declared as a public, Apache-2.0, provenance-pinned
governed data asset.

Declare → Assess → Certify

Model governance in KYE Protocol™ is a short path, climbed one posture
at a time:

1. Declare — identify the model and capture its AI-BOM: provenance
   and licence. The model is now declared.
2. Assess — KYE Protocol™ maps the obligations the intended use
   carries and reviews the evidence. The model is now assessed.
3. Certify — a signed evidence pack backs the framework coverage
   claim. The model is now certified — and adoptable by a regulated
   buyer.

Models inside agentic systems: the action boundary

Increasingly, models don't just answer — they act, inside agent
loops that delegate work across handoffs. Governance that stops at
deployment time misses the risk that matters at run time: the
authorised meaning of an instruction can drift between the principal
who gave it and the model call that executes it.

KYE Protocol™ therefore governs at the action boundary: before a
delegated action is admitted, it checks whether original intent,
constraints, context, memory, incentives, timing and state remain
aligned across handoffs. That check is KYE Meaning Continuity™ — you
can try an open reference implementation interactively in the
KYE Meaning Continuity Lab™ Space.

Honest scope

KYE Protocol™ certifies provenance and governance posture. It does
not certify capability, safety, or fairness — those remain the domain
of evals, red-teaming, and the model card. And a framework KYE
Protocol™ has not mapped stays honestly out of scope: coverage is
never inflated to 100%. Governance claims you cannot defend in front
of an auditor are worse than no claims at all.

Who this is for

• Model & dataset publishers — a governed signal that makes your
  artifact adoptable by regulated buyers.
• Regulated adopters in finance, health, and government — prove
  governance, don't assert it.
• Builders of agentic systems — an authority check at the action
  boundary, not a post-hoc audit.

Open contracts, governed runtime

The contracts are open (Apache-2.0) and live in this org:
schemas ·
reason-codes ·
examples.
The managed engines that operationalise them are commercial.

• Model governance on the Hub, end-to-end:
  kyeprotocol.com/huggingface.html
• Source of truth: github.com/KYE-Protocol
• Engage: kyeprotocol.com/engage.html

KYE™, KYE Protocol™, Know Your Entity™, KYE Meaning Continuity™,
Evidence Pack™ and Replay-Proof™ are trademarks of the KYE Protocol™
project. Text of this article: Apache-2.0.