A new control-layer document is now public:

"Bounded Capability Extraction Clause v0.2.1"

DOI: https://zenodo.org/records/20953796 GitHub release: https://github.com/Kot141078/advanced-global-intelligence/releases/tag/bcec-v0.2.1 Website page: https://ivankotov.eu/publications/bounded-capability-extraction-clause-v0-2-1/

This is not a new compliance claim.

It is a narrow clause for one specific problem:

"how a c-class system, agent mesh, local node, or anchor composition may extract capability from external oracle systems without silently turning distributed querying into reconstruction or laundering the result back into the corpus as clean native experience."

The core rule is simple:

"A6 may pool anchors. It must not pool extraction."

Modern AI systems no longer only ask one model one question.

They route tasks through local models, cloud judges, CLI agents, API tools, repositories, memory stores, release systems, and public documentation surfaces.

That creates a quiet risk:

one participant queries a little, another queries a little, a third summarizes, a fourth retains, a fifth turns the result into reusable policy, training material, or “experience”.

Individually, each step may look bounded.

Collectively, it can become an extraction fleet.

BCEC v0.2.1 defines a discipline for that boundary:

  • extraction source-classes;
  • egress metering;
  • composition-level caps;
  • anti-Sybil / principal accounting;
  • single-unit budget rules;
  • reconstructive-threshold handling;
  • witness-compatible extraction records;
  • handoff to Synthetic Laundering for re-admission control.

The purpose is not to ban external oracle use.

External models, cloud APIs, and judges are useful.

The purpose is to prevent three failures:

  1. treating lawful per-query use as lawful aggregate reconstruction;
  2. treating external capability as native capability after packaging;
  3. allowing distributed agents to hide the origin and extraction cost of what they retain.

The document has passed b-layer review for this textual version.

Status:

"final textual clause / b-layer PASS reviewed draft"

Non-claims:

  • not sealed integration;
  • not legal certification;
  • not product certification;
  • not conformance acceptance;
  • not a claim that distillation is impossible.

This clause belongs beside A6 Composition Layer, Synthetic Laundering, EA Value, ARQ, L4 Witness, ARL, Beacon, and CGAM.

In simpler terms:

when AI systems begin to use other AI systems as external capability sources, provenance is not enough.

You also need an extraction budget, a source-class, a witness trail, and a rule preventing “derived from outside” from becoming “ours by forgetting”.

That is the boundary BCEC is trying to make inspectable.