EUVIMEDEUVIMED entdecken
Crossref

Where the Guarantee Stops: Safe-Delivery Boundaries for Verification-Gated SLM Agents

Narcisse Rufin Kengne Sakeng, Hajar Moudoud, Zakaria Abou El Houda

Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 1045-1052

DOI: 10.1609/aaaiss.v10i1.43525 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Narcisse Rufin Kengne Sakeng und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Agentic systems are increasingly deployed to make or explain consequential decisions, yet their safety guarantees are typically reported as aggregate performance on benign or in-distribution inputs. For an operator, however, the critical question is not only how much a safeguard reduces risk, but where its guarantee stops holding. We formalize this as a safe-delivery boundary and develop a methodology for locating it through controlled perturbation, surface-form probing, and claim-domain coverage analysis, rather than through average-case evaluation alone. We instantiate this methodology in a bounded, auditable agent for on-device security triage that gates explanation on decoder-ensemble coherence and filters generated claims through a deterministic, schema-grounded verifier. On CIC-IoT-2023, the full pipeline reduces unsupported-claim exposure from 0.429 to 0.176, but this average conceals a sharply structured boundary: the verifier detects violations with near-perfect precision immediately outside its declared 2% tolerance, yet a single surface-form rewrite (numerals as words) collapses detection to zero without touching the underlying predicate. More broadly, 74.8% of delivered rationale sentences fall entirely outside the verifier's claim domain and are never examined, and 17.2% of incorrect underlying decisions still reach the analyst with an accepted, "verified" explanation, showing that grounding a claim and correcting a decision are independent guarantees. These results demonstrate that a deterministic safeguard can be exact within its declared predicate while remaining structurally blind beyond it, and that reporting an average safety score without reporting this scope can materially mislead an operator about what the guarantee actually covers.

Zitieren

APA

Sakeng, N. R. K., Moudoud, H., & Houda, Z. A. E. (2026). Where the Guarantee Stops: Safe-Delivery Boundaries for Verification-Gated SLM Agents. Proceedings of the AAAI Symposium Series, 10(1), 1045-1052. https://doi.org/10.1609/aaaiss.v10i1.43525

Vancouver

Sakeng NRK, Moudoud H, Houda ZAE. Where the Guarantee Stops: Safe-Delivery Boundaries for Verification-Gated SLM Agents. Proceedings of the AAAI Symposium Series. 2026;10(1):1045-1052. doi:10.1609/aaaiss.v10i1.43525

BibTeX
@article{Sakeng2026,
  author = {Sakeng, Narcisse Rufin Kengne and Moudoud, Hajar and Houda, Zakaria Abou El},
  title = {Where the Guarantee Stops: Safe-Delivery Boundaries for Verification-Gated SLM Agents},
  journal = {Proceedings of the AAAI Symposium Series},
  year = {2026},
  volume = {10},
  number = {1},
  pages = {1045--1052},
  doi = {10.1609/aaaiss.v10i1.43525}
}
Kostenlos starten →