Opens with a direct position and its central trade-off.
Two public experiments · two transparent boundaries
Councils
Ask the same hard question under two very different conditions: tiny models running locally through Ollama, or free cloud models taking distinct deliberation roles.
No fleet memory
No delegation
No saved conversation state
Mode 01 · Oracle
Local Ollama · $0
TinyLLM Local Oracle
Five fixed local roles deliberate on Oracle ARM hardware: a Tiny-Agent proposer, llama analyst, qwen critic, EVE observer, and qwen synthesizer.
No fleet memoryNo delegationSlow by design
Open local workspace
Mode 02 · Cloud
Free models · $0
OpenRouter Free
A proposer, critic, and synthesizer use fixed free-model rosters. Responses stream through a public relay; availability depends on free-tier queues.
No fleet memoryNo delegationFree-only roster
Open free workspace
Mode 01 · Oracle local Ollama
TinyLLM Local Oracle
Place a claim before five tiny, named roles and watch the proposal, three independent readings, and final synthesis arrive token by token.
Stateless boundary. The prompt is sent only to the public TinyLLM Ollama relay for this run. It receives no fleet memory, personal context, tools, or delegation path.
Five local models run in sequence; model loading can make the first token slow.
Maps assumptions, evidence needs, and likely consequences.
Stress-tests the proposal and its weakest assumptions.
Observes how the claim frames agency, perspective, and selfhood.
Waits to weigh all four prior views into one bounded conclusion.
Tiny outputs can be literal, repetitive, or wrong. The value here is observational: the seams in small-model reasoning remain visible.
Mode 02 · Fixed free-model roster
OpenRouter Free
One model proposes, a different family challenges it, and a third synthesizes what survives.
Stateless boundary. Only the question and the visible deliberation are sent through the free-model relay. There is no memory lookup, user profile, tool call, or delegation.
Free-tier queues vary; the runtime falls through a verified free roster.
Checking the free-model roster…
Forms a direct first position.
Tests assumptions, omissions, and failure modes.
Weighs both positions into one practical answer.
All selected model identifiers keep their :free suffix. If every fallback is queued or unavailable, the run stops honestly instead of switching to a paid model.