Project SynaΨ

Honest-by-architecture. An integrated cognitive system for personal assistants, published as a defensive disclosure.

Thesis: the source of truth is never the generator. The generator authors interpretation; decidable computation is routed to alternate architectures; an independently-trained critic audits; a continuous sentiment memory carries lived experience without interpretation.

Date of disclosure: 2026-08-20. Temporarily paused research 2026-09-04 in favor of research on the more immediate benefit of chain of answers (https://github.com/miltos22/Chain-of-answers)

---

The Architecture

1. Generator — produces the output stream. Demoted from source of truth; it is the author only of interpretation.
2. Router — an internal component of the cognitive process that activates routing authority for decidable spans. The activation mechanism is intentionally left unspecified; any means of granting the router authority over a span of the generator's process falls within this disclosure.
3. Alternate architectures — heterogeneous, non-LLM computation engines to which decidable spans are routed: simulators, solvers, checkers, calculators — and non-LLM neural networks that predict the world or physics (learned dynamics / world models in a non-LLM format). The result of a routed span is part of the same cognitive process as the generator's own output.
4. Critic expert(s) — one or more components specialized for critique, trained separately from the generator, so that their independence is structural rather than nominal. They are part of the thinking process itself: when the system self-criticizes, the critic expert(s) contribute their own thinking to the self-critique.
5. Knowledge library — a knowledge database: a pre-compiled, separately-trained expert dedicated to keeping track of the model's overall knowledge, requested regularly during generation.
6. Continuous sentiment memory — sentiment associations carried in the model's weights, not the context. It is structured like a list of ideas/subjects with sentiment values, but each association is a weight-level conditioning applied in real time when the reasoning process touches that subject — emulating lived experience rather than reading stored text. It stores no interpretations, only sentiment. The reasoning component interprets; the critic audits.
7. Value function — epistemic honesty: abstain > confidently-wrong, everywhere.

Governing Principles

• The source of truth is never the generator.
• Decidable is computed, not generated.
• The critic's independence is trained in, not assumed.
• Memory stores sentiment; reasoning interprets; the critic audits.

Claimed Novelty

The individual components have prior art (see PRIOR_ART.md). This disclosure claims:

1. The assembly — the components above combined into one coherent system for a personal assistant, with the generator structurally demoted from source of truth.
2. Routing to alternate architectures as a general principle — decidable computation routed to heterogeneous non-LLM architectures as an internal part of the cognitive process; the calculator is one instance, not the claim.

On the Name

Ψ is also used in the epistemic-honesty literature as the Confidence-Evidence Ratio (assertoric force proportional to evidential warrant). The name is an intentional fit: this system is built around honesty as its product.

Collaboration

The author is open to professional collaboration with any individual, lab, or organization taking this idea seriously. If this architecture resonates and you want to build it properly, reach out: miltiadiskd@gmail.com