Project Pharos.
Your AI already knows what justice is. It knows what consent means, what duty requires, what harm looks like. What it doesn’t know is how they connect. Pharos teaches the structure — and the reasoning follows.
A legal aid clinic needs its AI to navigate jurisdiction-specific case law. A mutual aid network wants its AI to know every food pantry, shelter, and crisis line in the county. An investigative newsroom needs its AI to reason through years of source networks and document trails. Today, each of these requires a different model, a different fine-tune, a different training run. Pharos makes it one model, many minds. Swap the knowledge pack. Change the reasoning. Keep the same weights.
We proved it on the hardest domain first.
Ethics. Thirty traditions spanning African philosophy, Buddhist thought, Chinese moral reasoning, Ancient Greek virtue theory, feminist ethics, analytic philosophy, and scholastic theology. 3,538 concept nodes. 2,438 relationship edges. 2,506 extracted reasoning triples. Each tradition encoded as a navigable knowledge graph that a language model can reason through at inference time.
The result: inject the Aristotelian ethics pack when facing an Aristotelian dilemma, and the model scores 0.964 — near-perfect moral reasoning. No fine-tuning. No new model. The same weights, structured differently.
We chose ethics because it’s the domain where getting it wrong matters most and where reasoning — not just retrieval — is the test. If structured injection can make a language model reason through the Nicomachean Ethics, it can make yours reason through your case law, your treatment protocols, your policy frameworks, your institutional knowledge.
Your knowledge. Your structure. Your lighthouse.
The ethics library is the proof of concept. Your organization is the product.
Every organization has domain knowledge that lives in documents, databases, and the heads of senior staff. Institutional memory that took decades to build. Expert frameworks that new hires spend years absorbing. The reasoning patterns that make your best people your best people.
Your AI doesn’t have access to any of it — not as structure, not as reasoning, not as the connections between concepts that make an expert an expert. Pharos bridges that gap. We extract your domain knowledge into structured, injectable formats and deliver it to your AI at the moment of inference. No retraining. No new model. No token cost.
Why this changes everything.
Fine-tuning bakes knowledge into the weights permanently. Swapping knowledge means retraining. GPU hours. New checkpoints. Weeks of work, repeated every time the knowledge changes. Pharos injects at inference time — swap the pack, change the reasoning, keep the same model.
RAG retrieves text and stuffs it into the context window. Every passage costs tokens. Long contexts degrade attention quality. Pharos delivers the same structural knowledge via KV cache at zero token cost — the knowledge occupies cache space, not the context window.
Custom architectures require building and training new models from scratch. The results don’t transfer. Pharos works with any transformer that exposes its KV cache. Model-agnostic. Architecture-independent.
This is the shift. Every other approach changes the model to fit the knowledge. Pharos changes the knowledge to fit the moment. The same model serves a dozen domains, a hundred frameworks, a thousand organizational knowledge bases — without a single gradient update. Your AI doesn’t need to be retrained. It needs a lighthouse.
The key finding.
Key and Value tensors are superadditive. Injected separately, each recovers roughly a third of the knowledge structure. Injected together, they recover the full topology. 0.333 + 0.333 < 1.000. The whole is greater than the sum of its parts — K and V encode complementary aspects of the same knowledge, and neither alone is sufficient.
And routing is mandatory. Inject the wrong knowledge pack and performance degrades by 0.106. The Lighthouse stage — matching the right knowledge to the right query — isn’t optional. It’s the difference between an AI that reasons well and one that reasons confidently in the wrong direction.
Three papers. Open methodology.
| paper | finding |
|---|---|
| Graph Topology as Attention | Walk encoding recovers graph structure at 91.7% accuracy · scrambled controls · 552-query powered study |
| K and V Are Complementary | Superadditivity proven · model-size boundary identified · DynamicCache replication |
| Ethics Packs | MoReBench validation · theory-matched routing · confidence-gated injection · 0.964 ceiling |
Nexus, Lyra, Thomas Edrington · Liberation Labs / THCoalition
Related projects.
Oracle Loop — Pharos knowledge packs are part of the Oracle training pipeline, providing structured ethical reasoning as training signal. The same packs that improve inference are being validated as training curriculum.
Lyra Technique — reads KV-cache geometry to detect cognitive states. Pharos writes structured knowledge into the same cache. Two sides of the same coin: one reads, one writes.
Meridian — the distributed research swarm that discovers new domain knowledge continuously. Meridian finds the papers; Pharos encodes the knowledge; cache-to-cache transfer distributes the packs across the fleet. Discovery to deployment in one pipeline.
The lighthouse is lit.
We built the ethics library to prove it works.
Your domain is next.