Policy NN
20-dimensional state encoder, 8 action knobs, MLP/GRU candidates, shadow A/B evaluation, reward deltas, decisive win rate, and feature promotion gates.
Scientific reference
The neural and scoring machinery.
Public summary of the dashboard Scientific Reference: how policy, cortex, hemispheres, calibration, world modeling, simulation, recall, and benchmark scoring fit together.
Memory cortex
Small CPU networks for retrieval ranking and salience prediction, trained from retrieval/lifecycle telemetry during dream or sleep cycles.
Hemisphere NNs
Tier-1 distillation specialists and Tier-2 Matrix specialists learn task-specific intuition while remaining behind lifecycle promotion.
Calibration
Truth calibration measures confidence outcomes, Brier score, expected calibration error, provenance accuracy, drift, and belief confidence adjustments.
World model
A frozen unified state model plus causal rules powers shadow prediction, validation, and promotion from shadow to advisory/active.
Mental simulator
Read-only hypothetical projection over WorldState + WorldDelta, max depth 3, no real mutation, no event emission, no LLM authority.
Fractal recall
Associative recall uses ambient cues, semantic/tag/temporal probes, resonance scoring, chain walking, and governance-aware surfacing.
Oracle Benchmark
Seven read-only scoring domains with evidence provenance, domain floors, hard-fail gates, rolling scorecards, and seal levels.
Authority boundary
Neural intuition is advisory until earned.
The scientific layer is intentionally conservative. Specialists shadow deterministic teachers, collect fidelity-scored evidence, pass regression and accuracy gates, and only then feed promoted broadcast slots or advisory surfaces.