BILL AND ONEIRO

Two local-first systems exploring what small models can do without a datacentre.

--------------------------------------------------------------------------
BILL -- a local-first conversational system

A local language model wrapped in retrieval, durable memory, planning, grounding
and verification, paired with a separate from-scratch predictor that can be read
line by line.

  REQUIREMENT -- YOU MUST HAVE OLLAMA INSTALLED
  Bill runs its language model locally through Ollama. Without Ollama installed
  and running, Bill will not start; there is no hosted fallback and no API key to
  supply, because nothing is sent off the machine by design. Install Ollama from
  ollama.com, then pull the model:

      ollama pull llama3.2:3b
      ollama serve

  The from-scratch core (bill2, reachable via /legacy) is the one part that runs
  without Ollama, since it uses no ML frameworks at all.

  Durable, user-confirmed memory that survives restarts
  Prioritised retrieval across a local corpus, a knowledge store, and web search
  A planning pass, an evidence contract, a factual verifier, a response gate
  bill2, the from-scratch core, reachable via /legacy

MODEL ATTRIBUTION, STATED PLAINLY
  The coherent prose comes from a local Llama 3.2 3B served through Ollama. The
  project does NOT claim that the from-scratch bill2 weights produce that prose.
  What Bill contributes is the orchestration around them.

  bill2 itself is written without ML frameworks -- a predictive concept graph
  with its own training, consolidation and calibration, developed across ten
  measured iterations each evaluated on held-out data, covering unit and neural
  recall, sleep-style replay consolidation, beam-traced reasoning, imitation
  learning, and calibration with measured expected calibration error.

Repository: github.com/jcooperkai-sys/Bill

--------------------------------------------------------------------------
ONEIRO -- a continuously running generative art system

  REQUIREMENT -- YOU MUST HAVE OLLAMA INSTALLED
  The teacher model runs locally through Ollama, the same as Bill. Without it the
  idea stream has no source and the system has nothing to learn from. The pixel-art
  generator itself is pure NumPy and needs nothing beyond Python.

A local model free-associates one word at a time. Each idea becomes a training
example for a pixel-art generator written from scratch in NumPy, which then
paints that idea from its own weights. Output streams to a live gallery. It runs
indefinitely and keeps learning while it runs.

  The teacher is independent of the student, so the generator never trains on its
  own samples -- the mechanism behind model collapse in self-training setups.
  Learning is online and persistent; state resumes between runs.
  Open vocabulary: unknown words are routed to an artist agent that designs a new
  sprite as drawing primitives, which the renderer draws.

Repository: github.com/jcooperkai-sys/Oneiro
