INTELLIGENCE / MODELS, FUNCTIONS AND WORLDS
From a record to useful action.
Compact personal models, specialist teams and simulated environments develop through useful tasks, visible evaluation and repair.
A personal nanoLLM is a design ambition.
The documents propose a compact language model shaped by approved personal examples. Model size follows task evidence and device resources. The training contract states tasks, source categories, private or secret grants, vocabulary and fidelity criteria.
Training and evaluation.
Declaration/reflection pairs, contextual examples and corrections form candidate data. Training and held-out evaluation have separate roles. Tests examine meaning, retrieval, conditions, uncertainty and response to a correction. A version card records method, sources, licence, compute and limits.
Function inputs, outputs and permissions.
A function defines purpose, inputs, outputs and errors. A caller selects and orders tools within permission. An agent has a role and handover boundary. An Aura specialist team and neural mixture-of-experts are different forms of organisation.
Three observers open a simulation.
The environment observer describes the wider setting. The character observer describes what the agent notices. A narrator interprets action in context. Their disagreement is useful evidence about attention and assumptions. Simulated choices remain distinct from real-world authority.
Compute belongs to a workload.
Personal devices, session workstations and chosen shared nodes handle different jobs. Capture continuity has independent resources. The source’s RAM, latency, storage and power envelopes are planning examples to benchmark, rather than procurement or performance claims.