Rome, 7 to 11 November 2026

Semantic Boundary @ CIKM ’26

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

Gaurav Baruah · Nimble Notions Inc.

At a glance

The question. Which information representation should cross a system boundary for a downstream purpose?

Why it matters. AI systems send traces of conversations, tool use and logs to monitoring, analytics and evaluation services. Removing sensitive words can strip task-relevant meaning, and it may still not stop sensitive information being inferred or linked across records.

What we did. We introduce Semantic Boundary, a framework for policy-constrained disclosure with provenance, and Open-SBB, a benchmark that compares nine alternative exports of the same synthetic medication-adherence events by task utility and residual linkage.

What we found. In our synthetic pilot, hiding sensitive words did not necessarily prevent linkage. Different tasks favoured different export conditions, and reusing one task’s preferred condition for another could reduce utility.

The takeaway. No representation was best for every task in the evaluated setting. The results motivate purpose-specific, policy-governed disclosure choices evaluated for both utility and residual linkage.

Explore Open-SBB

Code, the benchmark, and reproduction instructions. The cikm-2026 branch freezes the code, protocol, and reported results for the paper. The current release is v0.1.3.

Explore the benchmark 
uv venv
source .venv/bin/activate
uv pip install -e ".[dev]"
make repro-cikm-2026

Verifies the CIKM paper’s protocol and committed evaluation artifacts, including figure checks. It does not rerun model-based transformations or assessments.

Please share your perspective

We would really appreciate your help with a five-minute survey on information sharing in AI systems. It will help guide where we take this research next.

Take the survey

Contributions

  • Framework

    Semantic Boundary governs what meaning crosses to each downstream consumer: declare the consumer, its purpose and policy; cross into an export with a provenance record; verify before release.

  • Benchmark

    Open-SBB tests text-transformation and semantic disclosure strategies on synthetic medication-journal events released to an observability triage consumer and a pharmacologic analytics consumer.

    • Nine conditions transform each log: raw text, masking or replacing sensitive words, LLM substitution and rewriting, and typed semantic exports.
    • Triage tasks are failure mode and error stage. Analytics tasks are medication class, side-effect signal, adherence barrier, and behavioural cohort.
    • A closed-world adversary scores persona, attribute, and longitudinal linkage. The combined score orders the conditions and is not a re-identification probability. The reported choice is the most useful condition whose linkage stays under the limit. Token recovery is a separate check for whether sensitive words reappear.
  • Result
    • Hiding the sensitive words did not stop entries linking back to the same person.
    • No single text transformation or export was best across all registered tasks under the same linkage limit, and the best choice differed between triage and analytics.

© 2026 Nimble Notions Inc. · Semantic Boundary research companion

This page does not load Google Analytics. A theme choice stays on this device. The host may log the request. Privacy