INNER I
Minimal Invariant Observer
An observation layer for intelligent systems. MIO wraps AI models with an explicit layer for observation, grounding, uncertainty, coherence, evidence, and action.
The model generates.
MIO observes.
Inner I evaluates.
OBSERVER
MIO is a model-independent AI observation and grounding harness. It does not replace the model. It sits around it: input, observation, interpretation, evaluation, action. The underlying model may change. The observer remains.
INDEPENDENCE
Generative models hide their epistemic state. MIO makes known, unknown, evidence, contradiction, and action visible — so confidence is never presented as truth.
HOW IT WORKS
Watch a request travel: input is observed, a swappable model generates, claims are grounded, Inner I evaluates, the action gate decides. Pause, or click a node.
ARCHITECTURE TOUR
REQUEST
The user asks. MIO records the input as a first-class observation — what was received, not what should be generated. The observer relation starts here: a distinction between the system and the changing informational state it is about to process.
MODEL
generates
OBSERVER
records
INNER I
evaluates
ACTION
decides
ADAPTERS
Grok, OpenAI, Anthropic, Gemini, Llama, Qwen, Mistral, Ollama, or a custom OpenAI-compatible endpoint. MIO talks to models through adapters. The observer core does not change.
GROUNDING
Groundedness is not model confidence. It is evidence support divided by claims that require support. Claims are first-class: supported, partial, unsupported, contradicted, or unverified.
ABSTENTION
Unknown is information. When grounding falls below threshold, MIO prefers ASK, SEARCH, VERIFY, or ABSTAIN over fabricated certainty. Silence before fabrication.
OPEN
A minimum observable interface: INPUT, OBSERVE, MODEL, EVIDENCE, EVALUATE, DECIDE, OUTPUT. Events are structured JSON. The same protocol can wrap any generator.
MPE ∩ MIO
MPE is the phenomenological limit case of experience. MIO is the proposed computational limit case of the observer relation that can register, preserve, or evaluate experience. Phenomenological claims, computational claims, empirical results, and MIO hypotheses are kept distinct. None of this is presented as established scientific fact.
PAPER
Read the 2026 MPE–MIO concept paper, then run the next research from inside MIO. Discoveries can be proposed into observer memory. The model does not silently retrain.
BUILD WITH MIO
POST /api/mio/run streams observation events. GET /api/health reports adapter status. See the protocol page for the full interface.
Enter MIO