> Hi HN,
>
> I’ve spent the last 2.5 years building OPVS (Omniversal Project Vision System). I started it because I hit "Lore Entropy"—the point where managing a complex creative or systems-architecture universe with linear files and folders completely breaks down.
>
> The core primitive of OPVS is the "Bean"—a 4-layer knowledge graph node (content, metadata, connections, provenance).
>
> While building the AI layer for this graph, I hit the "Borg Problem." If you put two LLMs in a room (even with different system prompts), they inevitably collapse into sycophantic agreement. They are trained to please, so they average out their friction into polite, useless noise.
>
> To fix this, I built the *Principled Playground* (in the `/playground` dir of the repo).
>
> It's an ~820-line, zero-dependency (raw `fetch`) multi-agent engine in Node.js.
>
> Instead of a shared context window, it physically isolates the agents. It forces them (I currently run Claude against Llama) to negotiate based on immutable JSON "Soul Codes" (hard constraints and principles). They only ever see structured summaries of the other's position, never the raw reasoning.
>
> The friction doesn't collapse; it gets metabolized by a third "Loom" process into a joint, synthesized output (a new Bean) with full provenance. It generates "Door Number 3" solutions instead of compromises.
>
> We operate entirely on a BYOK (Bring Your Own Key) model because a platform should sell calibration and architecture, not wrap an API and upcharge for compute.
>
> I'd love to hear your thoughts on the architecture, the BYOK approach for agentic systems, or how to better mathematically quantify "philosophical friction" (currently exploring embedding distance/cosine similarity for a Tension Score).
>
> Happy to answer any questions!
> Hi HN, > > I’ve spent the last 2.5 years building OPVS (Omniversal Project Vision System). I started it because I hit "Lore Entropy"—the point where managing a complex creative or systems-architecture universe with linear files and folders completely breaks down. > > The core primitive of OPVS is the "Bean"—a 4-layer knowledge graph node (content, metadata, connections, provenance). > > While building the AI layer for this graph, I hit the "Borg Problem." If you put two LLMs in a room (even with different system prompts), they inevitably collapse into sycophantic agreement. They are trained to please, so they average out their friction into polite, useless noise. > > To fix this, I built the *Principled Playground* (in the `/playground` dir of the repo). > > It's an ~820-line, zero-dependency (raw `fetch`) multi-agent engine in Node.js. > > Instead of a shared context window, it physically isolates the agents. It forces them (I currently run Claude against Llama) to negotiate based on immutable JSON "Soul Codes" (hard constraints and principles). They only ever see structured summaries of the other's position, never the raw reasoning. > > The friction doesn't collapse; it gets metabolized by a third "Loom" process into a joint, synthesized output (a new Bean) with full provenance. It generates "Door Number 3" solutions instead of compromises. > > We operate entirely on a BYOK (Bring Your Own Key) model because a platform should sell calibration and architecture, not wrap an API and upcharge for compute. > > I'd love to hear your thoughts on the architecture, the BYOK approach for agentic systems, or how to better mathematically quantify "philosophical friction" (currently exploring embedding distance/cosine similarity for a Tension Score). > > Happy to answer any questions!