The intelligence layer

With the governance stack proven, the intelligence layer was built on top of it. The full pipeline:

Note where it starts. Detection and classification run before any intelligence does. The system knows how stressed it is before it decides how ambitious to be.

Anarchy — bounded simulation

Event-triggered rather than continuous. Anarchy wakes on specific triggers — traffic spike, error surge, confidence collapse, novelty detection, strategy failure loop — and spawns a lightweight micro-simulator to evaluate three candidate strategies: baseline (what the system was going to do), conservative (stability-focused) and aggressive (performance-focused).

Each is scored:

Score = (Accuracy × 0.4)
      + (Stability × 0.3)
      - (Latency × 0.15)
      - (Resource Cost × 0.15)

An override fires only if the best simulated score beats baseline by at least 10%, and the whole thing runs inside a hard 30 ms budget.

Both numbers are doing real work. The 10% margin means marginal improvements do not get to disturb a working system — the default has to be beaten decisively, not narrowly. The 30 ms cap means the deliberation cannot become the emergency.

ORACLE — the world model

An RSSM-based world model that predicts future states in latent space rather than running full simulations. Short-path conservative mode, confidence scaling, and memory-assisted candidate ordering.

The system imagines futures instead of simulating them, which is the only version that fits inside the time budget. A simulator that is accurate and too slow produces nothing; the prediction has to arrive while the decision is still open.

MORL — competing objectives

Multi-objective reinforcement learning across goals that genuinely conflict: minimise travel time, minimise emissions, prioritise emergency vehicles, reduce fuel consumption. Every one of those trades against at least one other.

MORL is automatically gated off under low confidence or high risk. The most sophisticated component is the first one switched off when conditions degrade, which is the correct ordering and an uncomfortable one to implement.

PULSAR — small, reversible corrections

A deterministic micro-adjustment engine. Reversible tuning under hard pulse-count and latency budgets. Small, precise, bounded.

PULSAR exists because not every correction should go through the full decision path. Some adjustments are small enough that the deliberation costs more than the error, and reversibility is what makes it safe to skip the deliberation.

Fusion — the coordinator

Coordinates ORACLE, MORL and PULSAR under safety gates: trust recalibration, emergency oracle down-weighting, a safety fallback gate, and VIHAAN-aware clamps.

Fusion is where "how much do I currently believe each of my own components" is decided, and it is answered continuously rather than configured once.

Meta — slow adaptation

A background controller that monitors runtime signals and applies smoothed control updates — but only when stable conditions justify adapting at all. Gradual, conservative, never aggressive.

Meta is the layer that could most easily destroy the system, because a controller that adapts the controller during instability amplifies exactly what it is trying to damp. So it is deliberately the slowest thing in the architecture, and it refuses to act while the system is stressed.