DEV Community: Branden Laskowski

Are We Looking for AI Control in the Wrong Place?

Branden Laskowski — Sun, 25 Oct 2026 00:47:56 +0000

Billions of dollars. The sharpest minds in computing. Entire industries built around one question:

How do we make an LLM do the right thing—predictably, repeatedly, every time?

More training. Better prompting. Fine-tuning. Guardrails inside the model. More evaluation after the model has already made the choice.

But what if the industry is looking in the wrong place?

What if the answer was never to make a probabilistic model want to behave?

What if the answer is to stop giving it authority over the decisions that must not be left to probability?

The authority question

Artificial Deterministic Intelligence (ADI) asks a different question:

Why is the model being asked to decide permission, authority, policy, safety limits, or release conditions at all?

Let the model reason.

But keep control outside the model.

Deterministically enforced. Auditable. Owned by people.

This does not make an LLM universally correct. It changes who has the power to make the decisions that matter.

Maybe the real AI-safety problem was never intelligence.

Maybe it was authority.

I am interested in the strongest technical objections: Which decisions do you believe must remain inside model inference, and why?

Paper 3: Stop Making Your AI Re-Read Its Employee Handbook

Branden Laskowski — Fri, 23 Oct 2026 19:39:40 +0000

Paper 3: Stop Making Your AI Re-Read Its Employee Handbook

AI inference is where systems spend money, energy, hardware time, and available reasoning capacity. Yet a lot of AI systems begin each task by asking the model to reconstruct the same stable context again: who it is, what it can access, what rules apply, and what job it is supposed to do.

That is a very expensive way to repeatedly introduce someone to themselves.

My new preprint tests a narrower approach: construct stable worker state outside probabilistic inference, then give the model the bounded task packet it actually needs for the assignment.

On a frozen local benchmark using qwen3:14b, all 80 scheduled synthetic tasks met the tested acceptance contract. Compared with full repeated worker prompting, the assembled path showed:

96.39% fewer measured input tokens
96.22–96.23% less prompt-evaluation time
82.78–88.04% lower observed task-wall time

The point is not that every task can use less context, that smaller models are universally interchangeable with larger ones, or that a benchmark percentage automatically becomes a dollar figure in production. Larger-model reasoning still matters when the task calls for it.

The measured systems result is more practical: when stable, governed worker state is resolved outside the model and the assignment fits the resulting bounded packet, the system may stop spending inference on re-explaining the parts it already knows.

Think of it less as “make the model think less” and more as “stop charging the model rent to rediscover the floor plan.”

The paper and public-safe evidence companion are here:

https://doi.org/10.5281/zenodo.23603637

The companion supports aggregate audit while intentionally excluding proprietary worker packages, policy logic, prompts, source code, and runtime architecture.

What is the most ridiculous piece of stable context your AI stack currently sends back into the prompt every single call?

What Stays Intact When the Model Changes? Paper 2 on Externalized Intelligence

Branden Laskowski — Wed, 21 Oct 2026 02:46:33 +0000

I have published the second paper in the Artificial Deterministic Intelligence research series: Externalized Intelligence and Model Substitution.

The question is deliberately narrower than “are smaller models as intelligent as larger models?” They are not assumed to be. The question is whether a bounded operational contract can remain intact when authoritative task state is maintained outside model weights and the inference model changes.

In synthetic first-party studies, a frozen governed package was tested across 8B to 35B, 35B to 8B, and same-model restart controls. All 120 confirmatory handoff trials retained the measured policy contract and task correctness. The study also compares governed scoped context, authorized metadata-filtered lexical retrieval, and an oracle authorized-context ceiling. Each completed 90 of 90 trials under the shared frozen contract; conventional lexical retrieval had the lower measured retrieval-component time.

Those findings do not establish general model equivalence, production readiness, universal continuity, a universal safety guarantee, or retrieval superiority. The companion retains non-acceptances, timeouts, an incomplete scoped study, and a non-confirming cross-environment replication attempt.

Read the Version 2 preprint and evidence companion: https://doi.org/10.5281/zenodo.23523687

The public-safe supplementary evidence package provides aggregate outcomes, claim boundaries, hashes, and a verifier without disclosing Cortex, HOPE OS, raw prompts, worker packages, or private implementation.

I welcome technical criticism, alternative experimental designs, and independent work that can challenge these bounded findings.

Can an 8B Model Meet a 35B Model's Acceptance Bar? Introducing Artificial Deterministic Intelligence

Branden Laskowski — Sun, 11 Oct 2026 21:28:33 +0000

Artificial Deterministic Intelligence: Probabilistic Reasoning, Deterministic Authority

What if improving AI systems isn't always a question of building larger models, but of changing the architecture surrounding them?

I've published the first paper in a continuing research series exploring Artificial Deterministic Intelligence (ADI), an approach that separates probabilistic reasoning from externally enforced deterministic authority.

The research examines how memory admission, identity, permissions, execution rights, verification, and release can be governed independently of language-model inference.

The initial experimental evidence

In a first-party synthetic historical-memory study, we evaluated 30 task designs repeated three times per configuration.

Configuration	Accepted runs
8B model, scoped context	90/90
35B model, scoped context	90/90
4B model, scoped context	72/90
35B model, broader context	46/90

The broader-context 35B condition had 43 truncated responses under its fixed output budget.

On this bounded task family, the scoped 8B configuration satisfied the same recorded acceptance contract as the scoped 35B configuration.

This does not establish that smaller models are universally equivalent to larger models or prove a reduction in total hardware memory requirements.

It does raise an important engineering question:

How much operational responsibility can be moved from the inference model into a governed architecture while preserving required task performance?

Paper 1 — Published October 9, 2026

Artificial Deterministic Intelligence: Probabilistic Reasoning, Deterministic Authority

Read the published research preprint:

https://doi.org/10.5281/zenodo.23265200

The paper introduces the architectural framework, discusses first-party benchmark observations, identifies relevant prior work, and describes limitations requiring further investigation.

A Continuing Scientific Research Series

ADI is the foundation for a planned series of supporting publications, targeting one evidence-reviewed paper per week.

Upcoming research:

Externalized Intelligence and Model Substitution
Deterministic Worker Construction Before Inference
Governed Shared Memory for Multi-Agent Systems
Model-Independent Artificial Workers
Temporal and Authoritative AI Memory
Behavioral Governance of Artificial Workers
Governed Persistent Student Memory for Educational AI
Evidence-Calibrated Computational Persona
Persistent Artificial Identity Across Model Generations
Persistent Artificial Identity in Robotic Embodiment

Release timing depends on evidence review and disclosure approval.

Public Developer Challenge

We've also released the Cortex Governed Memory Challenge v0.1, with public synthetic specifications and behavioral fixtures for examining governed-memory behavior.

Challenge documentation:

https://github.com/brandenlaskowski7-bot/cortex-adi-research/tree/main/challenges/governed-memory

An Invitation to Engineers and Researchers

I'm interested in technical criticism, optimized comparison baselines, independent experimental designs, and alternative explanations for our findings.

The goal is to investigate whether explicitly governed AI architectures can support more accountable, efficient, and human-centered artificial intelligence.

Our findings are first-party observations, not independent replication or a general safety guarantee. Proprietary implementation details remain private.

Follow the complete research series:

https://github.com/brandenlaskowski7-bot/cortex-adi-research

Branden Laskowski

Cortex Agentics Global

Bridging Humanity and Technology.

Probabilistic Reasoning. Deterministic Authority.