A Deterministic Architecture for Long-Horizon AI Systems
Fuente:
Zenodo
Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Recurso digital |
| Sprache: | Englisch |
| Veröffentlicht: |
Zenodo
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866901594653589504 |
|---|---|
| author | Maddry, Taylor |
| author_facet | Maddry, Taylor |
| contents | <p>This record contains Version 26 (frozen preprint) of <em>A Deterministic Architecture for Long-Horizon AI Systems</em>.<br><br>AI systems intended to operate as long-running autonomous or adaptive systems face a class of failure not captured <br>by traditional notions of error, robustness, or performance degradation. Even when locally correct and internally <br>consistent, such systems may become unsafe to rely upon as accumulated state and derived artifacts gradually <br>acquire de facto authority. This failure arises not from adversarial behavior or misalignment, but from implicit <br>authority drift driven by reuse, dependency, and ungoverned promotion of information over repeated autonomous <br>decision cycles.<br><br>We present a deterministic architecture for long-horizon autonomous and adaptive systems that prevents implicit <br>authority drift by construction. The architecture enforces a strict separation between immutable substrate law and <br>governed mutable state, treats authority as an explicit, versioned system property, and constrains all <br>authority-bearing change to discrete, inspectable, and reversible governance actions. A deterministic execution <br>pipeline with mandatory refusal semantics ensures that autonomous action selection and state transition occur only <br>when substrate-enforced authority conditions are satisfied.<br><br>The system employs a three-tier memory architecture that distinguishes ephemeral working context, persistent <br>non-authoritative artifacts, and authoritative epistemic basis. Research, analysis, and generative components may be <br>used freely within autonomous and adaptive workflows, but are treated as epistemically inert by default, producing <br>artifacts rather than authority. Provenance captures not only dependency relationships but epistemic admissibility, <br>enabling deterministic replay and post hoc validation of autonomous decisions under fixed authority configurations.<br><br>This work does not aim to produce correct answers, aligned behavior, or adaptive intelligence. Instead, it presents a <br>deterministic architectural framework that constrains how long-running autonomous systems may evolve, <br>accumulate state, and exercise authority over time. By enforcing explicit governance, reversibility, and determinism <br>under accumulation, the architecture supports autonomous operation in high-cost-of-failure domains where <br>auditability, bounded authority, and correction without implicit drift are required.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18857597 |
| institution | Zenodo |
| language | eng |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | A Deterministic Architecture for Long-Horizon AI Systems Maddry, Taylor deterministic architecture long-horizon AI authority drift epistemic authority governed promotion immutable enforcement provenance deterministic replay AI safety autonomous systems AI governance <p>This record contains Version 26 (frozen preprint) of <em>A Deterministic Architecture for Long-Horizon AI Systems</em>.<br><br>AI systems intended to operate as long-running autonomous or adaptive systems face a class of failure not captured <br>by traditional notions of error, robustness, or performance degradation. Even when locally correct and internally <br>consistent, such systems may become unsafe to rely upon as accumulated state and derived artifacts gradually <br>acquire de facto authority. This failure arises not from adversarial behavior or misalignment, but from implicit <br>authority drift driven by reuse, dependency, and ungoverned promotion of information over repeated autonomous <br>decision cycles.<br><br>We present a deterministic architecture for long-horizon autonomous and adaptive systems that prevents implicit <br>authority drift by construction. The architecture enforces a strict separation between immutable substrate law and <br>governed mutable state, treats authority as an explicit, versioned system property, and constrains all <br>authority-bearing change to discrete, inspectable, and reversible governance actions. A deterministic execution <br>pipeline with mandatory refusal semantics ensures that autonomous action selection and state transition occur only <br>when substrate-enforced authority conditions are satisfied.<br><br>The system employs a three-tier memory architecture that distinguishes ephemeral working context, persistent <br>non-authoritative artifacts, and authoritative epistemic basis. Research, analysis, and generative components may be <br>used freely within autonomous and adaptive workflows, but are treated as epistemically inert by default, producing <br>artifacts rather than authority. Provenance captures not only dependency relationships but epistemic admissibility, <br>enabling deterministic replay and post hoc validation of autonomous decisions under fixed authority configurations.<br><br>This work does not aim to produce correct answers, aligned behavior, or adaptive intelligence. Instead, it presents a <br>deterministic architectural framework that constrains how long-running autonomous systems may evolve, <br>accumulate state, and exercise authority over time. By enforcing explicit governance, reversibility, and determinism <br>under accumulation, the architecture supports autonomous operation in high-cost-of-failure domains where <br>auditability, bounded authority, and correction without implicit drift are required.</p> |
| title | A Deterministic Architecture for Long-Horizon AI Systems |
| topic | deterministic architecture long-horizon AI authority drift epistemic authority governed promotion immutable enforcement provenance deterministic replay AI safety autonomous systems AI governance |
| url | https://doi.org/10.5281/zenodo.18857597 |