Open-SBB ↗ Talk to us ↓
Semantic Boundary ACM CIKM 2026 Accepted Research

Control what meaning crosses a system boundary.

AI systems retrieve, combine, infer, summarize, and transform information before sending derived representations to models, agents, tools, and external services.

While access control protects original sources, Semantic Boundary governs derived semantics:

For a given downstream purpose, what information should actually cross under a governing policy?

Access controls govern what a system may access. Agent controls govern what it may do. Semantic Boundary asks what representation of derived information may cross for the purpose at hand.

See how it works ↓ Explore the research ↓

Semantic Boundary framework

Running examples

Same source. Different purpose. Different representation.

A system can be authorized to use the source without every downstream purpose requiring the same representation.

That’s the semantic boundary

Disclosure becomes a choice over representations: preserve what the declared purpose requires while satisfying the disclosure constraint.

The CIKM 2026 study evaluates this experimentally. Under the same constraint, different purposes can prefer different representations—and reuse across tasks can impose measurable utility regret.

Peer-reviewed · open · reproducible

A framework, a paper, and an open benchmark.

Semantic Boundary is a framework for policy-constrained semantic disclosure. It separates purpose, policy, transformation, verification, and release.

Semantic Boundary treats disclosure as an explicit choice over transformed representations, balancing purpose-specific utility against declared disclosure constraints.

ACM CIKM 2026 Accepted

Semantic Boundary: A Framework and Benchmark for Policy-Constrained Semantic Disclosure

Gaurav Baruah · Nimble Notions Inc.

  • Formalizes policy-constrained semantic disclosure as a downstream systems problem.
  • Evaluates purpose-specific utility vs. residual linkage risk across representations.
  • Measures utility regret when representations are inappropriately reused across tasks.
Read the paper via ACM Digital Library ↗

Available via ACM DL after November 7, 2026

Open-SBB v0.1.3

Frozen, reproducible artifact

Open evaluation instrument for the CIKM 2026 study.

  • 9 reference transformation conditions across multiple downstream tasks
  • Purpose-specific utility vs. residual linkage risk
  • Frozen reproduction checks for the reported CIKM results
make repro-cikm-2026 Published reproducibility target
GitHub Repo ↗ Zenodo Archive ↗

Scope: Semantic Boundary complements IAM, DLP, redaction, tokenization, and LLM gateways; it is not a universal privacy mechanism or regulatory-compliance guarantee. Where deterministic controls fully express the policy, use them.

Regulatory & governance context

Governance regimes already impose purpose- and necessity-oriented constraints on some uses and disclosures of information—from GDPR purpose limitation and data minimisation to HIPAA’s minimum-necessary standard.

Semantic Boundary provides a framework for reasoning about such constraints when AI systems transform information before disclosure.

Requirements vary by context and include exceptions. Semantic Boundary does not establish or guarantee regulatory compliance.

Does your system have this problem?

If your architecture has rich or sensitive information that crosses boundaries for different purposes, we would like to learn how you are handling it today.

Where are you seeing this problem?

We are continuing to develop Semantic Boundary and the open benchmark, and we’re especially interested in real-world systems where this problem already appears.

Talk to us ↗

Research · Benchmark Feedback · Real-World Use Cases

gb [at] nimblenotions [dot] ca

© 2026 Nimble Notions Inc. · Semantic Boundary research companion