Reactive
Single-task response without learning; ideal for gates, checks and deterministic classifiers.
RESOURCES · AI
What AI is, and where it stops.
AI performs tasks that once required human intelligence: learning, reasoning, problem-solving, perception and language understanding. Inside Mediator systems, each class carries an explicit authority boundary.
OVERVIEW
Four classes, one production reality.
Reactive machines respond without memory: given the same input they produce the same output, which makes them testable and safe for fixed tasks such as validation checks, classification gates and rules enforcement. Limited-memory systems learn from observed data and carry every current production workload worth naming: ranking, detection, forecasting, transcription, retrieval. Their behavior drifts with retraining, so Mediator pins model version, training window and evaluation receipt to every deployment. Theory-of-mind systems would model human intent and emotion; published results remain lab-bound with no auditable deployment path, so they are not presented as product capability. Self-aware systems would hold self-models and consciousness; no such system exists. Adoption moves on three rails: people who can state authority and falsifiers, platforms where data lineage survives contact with production, and processes that turn model output into receipted action instead of chat logs.
Limited memory
Learns from observed data; the only class carrying production workloads.Version, window and eval receipt pinned per deployment.
Training-window discipline
Behavior follows the data it saw; stale windows produce stale judgments.Retraining is a governed event with before/after evals.
Evaluation before promotion
Held-out sets, contradiction probes and regression packs gate every model change; failing evals hold the release.
Theory of mind
Would model human intent; lab-bound, no auditable deployment path, never sold as capability.
Self-aware
Would hold self-models; does not exist.
People
Operators who can state authority, falsifiers and rollback before the run starts.
Platforms
Data foundations where lineage, identity and timestamps survive production contact.
Processes
Workflows converting model output into approved, receipted action rather than chat history.
PROOF
Capability claims stay inside evidence.
Model class, training boundary and deployment state are stated separately.