Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | English |
| Published: |
Zenodo
2026
|
| Subjects: | |
| Online Access: | https://doi.org/10.5281/zenodo.19869287 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- <p>Large language models produce fluent, confident output across a wide distributional surface but fail systematically at the boundaries of domains governed by external validation mechanisms — law, medicine, engineering, empirical science. This paper develops a theoretical foundation for that failure mode. It introduces the concept of thin context as a formally defined epistemic condition: the state in which a representation lacks the domain-specific constraints necessary for a validation mechanism to constitute an authoritative answer. It identifies semantic compression — the statistical compression of meaning performed by distributional language representation — as the generative mechanism of thin context, and it argues that thin context is the epistemic substrate from which the structural failure modes of agentic systems are generated.</p> <p>The analysis proceeds in three movements. The first (§§2–5) develops the theoretical apparatus: language as lossy compression (Shannon, Harris, the distributional hypothesis); epistemic domains and their validation mechanisms (law, medicine, engineering, science); the formal definition of thin context; the mechanism of semantic compression that produces it. The second (§§6–7) draws the bridge to human cognition and to the HGC³AE² framework: human inference under thin context succeeds where distributional inference fails because humans carry domain-specific validation authority that distributional models structurally cannot; HGC³AE² is not a governance preference layered on top of a capable system — it is what the epistemic analysis structurally requires. The third (§§8–11) develops the operational implications: architectural interventions (retrieval-augmented generation, constrained decoding, explainability) reduce but do not eliminate thin context; human epistemic authority at domain boundaries is the supply side of the validation mechanism that no architecture replaces; evaluation regimes calibrated to distributional objectives cannot detect thin context at domain boundaries and must be reformed.</p> <p>This paper is the epistemic companion to *Mitigating Confident Misalignment in Agentic Systems: The HGC³AE² Framework* (Kuiper 2026). Paper One identified confident misalignment as the dominant failure mode and proposed HGC³AE² as a governance-first architecture for addressing it. Paper Two provides the epistemic account of why that architecture is structurally necessary rather than merely prudent.</p> <p><strong>Rights envelope:</strong> Citation permitted with full attribution. No reproduction, redistribution, or derivative works without written permission. AI/ML training use disallowed. See the citation policy at https://nonsequitur.tech/pubs/citation-policy/ for the full rights envelope.</p> <p>Canonical site URL: <a href="https://nonsequitur.tech/white-papers/epistemic-constraints/">https://nonsequitur.tech/white-papers/epistemic-constraints/</a></p> <p>Public archive: <a href="https://github.com/LittleYeti-Dev/yks-pubs/blob/main/papers/epistemic-constraints-v1-preprint.pdf">yks-pubs/papers/epistemic-constraints-v1-preprint.pdf</a></p>