Artificial Deterministic Intelligence: Probabilistic Reasoning, Deterministic Authority

Abstract

Language models can propose useful actions without being reliable authorities over the systems that execute

them. We define Artificial Deterministic Intelligence (ADI) as an architectural approach in which probabilistic

models provide bounded inference while identity, authority, memory admission, policy, execution rights, state

transitions, verification, and release are governed by deterministic systems external to the model. Determinism

refers to specified control decisions under explicit state and inputs; it does not imply deterministic generation,

correct policies, truthful memory, or universal safety. We present a layered reference model and separate

conceptual requirements from observations in a governed research evidence ledger. In a first-party synthetic

memory benchmark of 30 designs repeated three times per arm, scoped 8B and 35B configurations each met

the recorded acceptance contract on 90/90 jobs, while scoped 4B and broader-context 35B configurations

accepted 72/90 and 46/90; 43 broader-context responses were truncated under the fixed output budget. A

separate eight-task context-admission diagnostic retained PARTIAL overall parity despite matching the

corrected context condition. Historical memory tests provide bounded implementation evidence; worker-

construction performance records with incomplete artifact lineage remain excluded. These observations

motivate, but do not establish, the complete ADI architecture. The proposal synthesizes established work on

reference monitors, capability security, formal methods, runtime assurance, neuro-symbolic systems, and

governed agents. Its contribution is an explicit lifecycle-wide authority contract, an evidence taxonomy, and

falsifiable evaluation questions. Independent replication, comparative evaluation, and validation of the full

enforcement boundary remain open.

ATHENA 7 A Life-Centered Deterministic Architecture for Cognition and Artificial General Intelligence
James Reeves
Zenodo25 Oct 2026PreprintCC BY 4.0Full textZenodo
Stochastic Mutation Semigroups: From Deterministic Collapse to Probabilistic Evolutionary Dynamics
Marshal I. Sampson, Christiana F. Igiri, Reny George +1
Mathematical and Computational Applications1 Sep 2026Peer-reviewedCC BYFull textGeneral mathematicsDOAJ
High-resolution Calibrated Probabilistic Hourly Precipitation from a Deterministic Forecast
Thomas M Hamill
arXiv10 Sep 2026PreprintCC BY 4.0Full textAtmospheric science & meteorology StatisticsarXiv
Probabilistic clinical studies and deterministic physiology: building a Monte Carlo bridge with the "Frank" project.
Felipe Coelho Argolo, Arthur Vasconcelos de Luna Freire, Larrie Rabelo Laporte +2
Einstein (Sao Paulo, Brazil)19 Oct 2026Peer-reviewedCC BYFull textStatisticsEurope PMC
Health risk assessment of carcinogenic and non-carcinogenic heavy metal exposure from commercial instant noodles in Bangladesh: a deterministic and probabilistic evaluation for adults and children
Debneela Paul, Sagar Bhowmik, Sheikh Mohiuddin Samrat +1
ChemRxiv29 Sep 2026PreprintCC BYFull textEnvironmental science & pollutionChemRxiv

Cite

BibTeX RIS

BibTeX

@misc{laskowski2026artificial,
  title = {{Artificial Deterministic Intelligence: Probabilistic Reasoning, Deterministic Authority}},
  author = {Branden Laskowski},
  howpublished = {Zenodo},
  year = {2026},
  month = oct,
  doi = {10.5281/zenodo.23510050},
  url = {https://zenodo.org/records/23510050},
  note = {CC BY 4.0},
}

RIS

TY  - UNPB
TI  - Artificial Deterministic Intelligence: Probabilistic Reasoning, Deterministic Authority
AU  - Branden Laskowski
PY  - 2026
DA  - 2026/10/20
PB  - Zenodo
DO  - 10.5281/zenodo.23510050
UR  - https://zenodo.org/records/23510050
AB  - Language models can propose useful actions without being reliable authorities over the systems that execute  them. We define Artificial Deterministic Intelligence (ADI) as an architectural approach in which probabilistic  models provide bounded inference while identity, authority, memory admission, policy, execution rights, state  transitions, verification, and release are governed by deterministic systems external to the model. Determinism  refers to specified control decisions under explicit state and inputs; it does not imply deterministic generation,  correct policies, truthful memory, or universal safety. We present a layered reference model and separate  conceptual requirements from observations in a governed research evidence ledger. In a first-party synthetic  memory benchmark of 30 designs repeated three times per arm, scoped 8B and 35B configurations each met  the recorded acceptance contract on 90/90 jobs, while scoped 4B and broader-context 35B configurations  accepted 72/90 and 46/90; 43 broader-context responses were truncated under the fixed output budget. A  separate eight-task context-admission diagnostic retained PARTIAL overall parity despite matching the  corrected context condition. Historical memory tests provide bounded implementation evidence; worker-  construction performance records with incomplete artifact lineage remain excluded. These observations  motivate, but do not establish, the complete ADI architecture. The proposal synthesizes established work on  reference monitors, capability security, formal methods, runtime assurance, neuro-symbolic systems, and  governed agents. Its contribution is an explicit lifecycle-wide authority contract, an evidence taxonomy, and  falsifiable evaluation questions. Independent replication, comparative evaluation, and validation of the full  enforcement boundary remain open.
ER  -

Abstract

Related

Cite