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:
- Declare — identify the model and capture its AI-BOM: provenance and licence. The model is now declared.
- Assess — KYE Protocol™ maps the obligations the intended use carries and reviews the evidence. The model is now assessed.
- 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.
- This article on the Hugging Face blog: hf.co/blog/masoudmarandi/kye-models-governance
- 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.
- Downloads last month
- 26