| _version_ | 1866901097657925632 |
|---|---|
| author | Bogdanov, Maryan |
| author_facet | Bogdanov, Maryan |
| contents | <p>This release marks the first complete and validated public version of ASE-SDK.</p> <p>ASE-SDK implements the Admissible Step Engine (ASE): a deterministic, stateless safety layer that gates proposed state-update steps (ΔS) such that the resulting next state remains within a formally defined admissible domain.</p> <h3>What ASE-SDK is</h3> <ul> <li>A generic safety envelope over state-update dynamics</li> <li>Applicable to control loops, optimization pipelines, and learning systems</li> <li>Deterministic, fail-closed, and explicitly bounded by construction</li> </ul> <h3>What ASE-SDK is not</h3> <ul> <li>Not an optimizer</li> <li>Not a learning algorithm</li> <li>Not a controller or policy</li> </ul> <h3>Validation scope</h3> <p>This release includes executable validation evidence demonstrating the use of ASE as a safety envelope for ML-style training loops:</p> <ul> <li>Parameter-level envelopes (θ)</li> <li>Extended training-state envelopes (θ + EMA + step)</li> <li>Full optimizer-state envelopes (θ + m + v + EMA + step, Adam-like)</li> </ul> <p>In all cases, admissibility is enforced on the <em>next system state</em>, rather than on raw gradients or internal optimizer signals.</p> <h3>Reproducibility</h3> <p>All validation results are reproducible via deterministic builds and test runs. Internal simulations are provided for validation purposes and are not part of the public API.</p> <h3>Documentation</h3> <ul> <li><code>README.md</code> — API contract, design intent, and integration boundaries</li> <li><code>VALIDATION.md</code> — executable validation evidence and test references</li> </ul> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18480297 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | ASE-SDK v1.1.1 — Reference Validation Release Bogdanov, Maryan <p>This release marks the first complete and validated public version of ASE-SDK.</p> <p>ASE-SDK implements the Admissible Step Engine (ASE): a deterministic, stateless safety layer that gates proposed state-update steps (ΔS) such that the resulting next state remains within a formally defined admissible domain.</p> <h3>What ASE-SDK is</h3> <ul> <li>A generic safety envelope over state-update dynamics</li> <li>Applicable to control loops, optimization pipelines, and learning systems</li> <li>Deterministic, fail-closed, and explicitly bounded by construction</li> </ul> <h3>What ASE-SDK is not</h3> <ul> <li>Not an optimizer</li> <li>Not a learning algorithm</li> <li>Not a controller or policy</li> </ul> <h3>Validation scope</h3> <p>This release includes executable validation evidence demonstrating the use of ASE as a safety envelope for ML-style training loops:</p> <ul> <li>Parameter-level envelopes (θ)</li> <li>Extended training-state envelopes (θ + EMA + step)</li> <li>Full optimizer-state envelopes (θ + m + v + EMA + step, Adam-like)</li> </ul> <p>In all cases, admissibility is enforced on the <em>next system state</em>, rather than on raw gradients or internal optimizer signals.</p> <h3>Reproducibility</h3> <p>All validation results are reproducible via deterministic builds and test runs. Internal simulations are provided for validation purposes and are not part of the public API.</p> <h3>Documentation</h3> <ul> <li><code>README.md</code> — API contract, design intent, and integration boundaries</li> <li><code>VALIDATION.md</code> — executable validation evidence and test references</li> </ul> |
| title | ASE-SDK v1.1.1 — Reference Validation Release |
| url | https://doi.org/10.5281/zenodo.18480297 |