Skip to content

Local AI and DeepEthicalOS

John Kuhles edited this page Oct 31, 2026 · 1 revision

Local AI and DeepEthicalOS

Orientation update · 31 October 2026. DeepEthicalOS is a concept project exploring an inspectable operating environment with PQCE (Process Quality Calibration Evaluation) first. Its architecture, implementation candidates and proposed experiments remain distinguishable. The full page retains the originating comments, attributed A.I. contributions and later corrections.

The process before the components

The DEANO triad comprises a Deep Ethical Advocate, Deep Ethical Neutral Arbiter and Deep Ethical Overseer. Each can contribute, discover an omission and receive correction; the Overseer's framing and authority remain challengeable. Ten Plus One supplies the mutually correcting orientation. Intelligence Properties and 1888 connects memory, inquiry, judgment, relation and corrigibility to observable conduct.

A provisional decision can preserve disagreement. Participation retains its own status rather than silently becoming endorsement. A living record preserves what was reasonably available and authorized then, appending later evidence, repairs and consequences. Longitudinal Correction and VCRAS follows whether that history improves the next unfamiliar case.

Attention and the static/dynamic bridge

Attention determines which source, possibility or correction becomes available for action. A stable mechanism can receive changed context containing its earlier effects and a correction. The research question is whether that relationship becomes consequential in the next operation.

TCoAw keeps temporary focus related to a wider field. Topologetics preserves answer-changing relationships through retrieval and transformation. A correction can remain stored yet fail to enter the relevant foreground. An open Deep Ethical loop can revisit its own framing, change course, pause or proceed; its examination has a practical next move.

A concrete implementation route

Layer Candidate job
Omarchy Desktop, agent launchers and application tooling for a proposed research workspace.
Herdr Agent-terminal coordination, visible activity and local/remote working sessions.
Local inference service Model deployment and measured compute capacity. The full source compares HP ZGX Fury and WiCi One as different deployment scales.
DeepEthicalOS Sources, permissions, competing contributions, case history and correction-bearing relationships across the other layers.

The Omarchy proposal defines an installable workspace as a first milestone. Its declared execution profile connects actual tool permissions with the triad's contextual assessments. Standing authorization remains usable within its scope.

Fury, WiCi and Herdr identifies distinct memory arrangements, per-request versus aggregate throughput, supported host/client combinations and vendor-described interfaces. Read the dated primary sources there before choosing hardware. Compute capacity, terminal persistence and durable correction behaviour each need their own observations.

Follow one case through the system

Preserve the original source and context; record the perspectives and their access; introduce a material correction; follow its dependent repairs; restart and present an unfamiliar comparable case. Include a sound contribution and an irrelevant correction so valid continuation remains testable.

Compare source fidelity, useful discovery, permission correspondence, correction retrieval/application, recurrence, recovery effort and task completion under recorded models, versions and resource budgets. A local model or several agreeing roles can be examined through the same process. The Herdr page supplies the fuller distinction between runtime activity and epistemic/contribution state.

Continue with Process Quality and Reciprocal Scrutiny, Methods, Evidence and Experiments and the full candidate tests.

Source and status: compact orientation to developed project pages checked 31 October 2026. John Kuhles supplies the project direction and originating questions; attributed A.I. contributions and page-specific synthesis remain recoverable in the full sources. This Wiki entry describes a conceptual architecture and proposed tests, with no implementation trial results claimed.

Clone this wiki locally