A Deterministic Linguistic Entropy Gate for Large Language Model Pipelines
Fuente:
Zenodo
Guardado en:
| Autor principal: | |
|---|---|
| Formato: | Recurso digital |
| Lenguaje: | inglés |
| Publicado: |
Zenodo
2026
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866901496353783808 |
|---|---|
| author | ROSATI BERISTAIN, ERNESTO |
| author_facet | ROSATI BERISTAIN, ERNESTO |
| contents | <p><strong>llm-entropy-filter</strong> is a deterministic linguistic entropy gate designed for use in large language model (LLM) pipelines. The system analyzes incoming text using rule-based heuristics to estimate linguistic entropy and structural signals commonly associated with spam, prompt injection attempts, or anomalous inputs.</p> <p>Unlike probabilistic classifiers or machine-learning filters, the engine operates through a deterministic scoring mechanism defined by configurable rulesets. This design ensures predictable behavior, reproducibility, and low computational cost, making the filter suitable as a lightweight pre-processing or middleware layer in AI systems.</p> <p>The engine evaluates textual signals such as structural irregularities, excessive repetition, shouting patterns, spam indicators, and other linguistic anomalies. These signals are combined into an entropy score that can trigger configurable actions such as:</p> <ul> <li> <p>pass (allow input)</p> </li> <li> <p>review (flag for inspection)</p> </li> <li> <p>block (reject input)</p> </li> </ul> <p>Because the evaluation process is deterministic and configuration-driven, outcomes are fully reproducible across deployments. The filter can therefore be used as a transparent and auditable gate before invoking computationally expensive LLM inference.</p> <p>The project is implemented in JavaScript/TypeScript and distributed under the Apache-2.0 license. It includes example integrations, configurable rulesets, benchmarking scripts, and documentation describing the architectural design and limitations of the approach.</p> <p>The system does not attempt to infer semantic truth, user intent, or ethical meaning. Instead, it focuses strictly on structural linguistic signals that can be evaluated deterministically. As such, the entropy gate is intended to complement, not replace, higher-level safety or reasoning mechanisms in AI systems.</p> <p>This release accompanies the technical document:</p> <p><strong>“A Deterministic Linguistic Entropy Gate for Large Language Model Pipelines.”</strong></p> <p>github.com/rosatisoft/llm-entropy-filter</p> <p><a href="https://github.com/rosatisoft/llm-entropy-filter">rosatisoft/llm-entropy-filter: entropy_analysis, intention_evaluation, triad_response, plain_text_summary</a></p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18857125 |
| institution | Zenodo |
| language | eng |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | A Deterministic Linguistic Entropy Gate for Large Language Model Pipelines ROSATI BERISTAIN, ERNESTO LLM safety linguistic entropy deterministic filtering prompt injection AI middleware rule-based filtering AI governance <p><strong>llm-entropy-filter</strong> is a deterministic linguistic entropy gate designed for use in large language model (LLM) pipelines. The system analyzes incoming text using rule-based heuristics to estimate linguistic entropy and structural signals commonly associated with spam, prompt injection attempts, or anomalous inputs.</p> <p>Unlike probabilistic classifiers or machine-learning filters, the engine operates through a deterministic scoring mechanism defined by configurable rulesets. This design ensures predictable behavior, reproducibility, and low computational cost, making the filter suitable as a lightweight pre-processing or middleware layer in AI systems.</p> <p>The engine evaluates textual signals such as structural irregularities, excessive repetition, shouting patterns, spam indicators, and other linguistic anomalies. These signals are combined into an entropy score that can trigger configurable actions such as:</p> <ul> <li> <p>pass (allow input)</p> </li> <li> <p>review (flag for inspection)</p> </li> <li> <p>block (reject input)</p> </li> </ul> <p>Because the evaluation process is deterministic and configuration-driven, outcomes are fully reproducible across deployments. The filter can therefore be used as a transparent and auditable gate before invoking computationally expensive LLM inference.</p> <p>The project is implemented in JavaScript/TypeScript and distributed under the Apache-2.0 license. It includes example integrations, configurable rulesets, benchmarking scripts, and documentation describing the architectural design and limitations of the approach.</p> <p>The system does not attempt to infer semantic truth, user intent, or ethical meaning. Instead, it focuses strictly on structural linguistic signals that can be evaluated deterministically. As such, the entropy gate is intended to complement, not replace, higher-level safety or reasoning mechanisms in AI systems.</p> <p>This release accompanies the technical document:</p> <p><strong>“A Deterministic Linguistic Entropy Gate for Large Language Model Pipelines.”</strong></p> <p>github.com/rosatisoft/llm-entropy-filter</p> <p><a href="https://github.com/rosatisoft/llm-entropy-filter">rosatisoft/llm-entropy-filter: entropy_analysis, intention_evaluation, triad_response, plain_text_summary</a></p> |
| title | A Deterministic Linguistic Entropy Gate for Large Language Model Pipelines |
| topic | LLM safety linguistic entropy deterministic filtering prompt injection AI middleware rule-based filtering AI governance |
| url | https://doi.org/10.5281/zenodo.18857125 |