Regulators / Objectives

All actor-side regulators are dimensionless — they scale with signal magnitude and do not require re-tuning when the reward amplitude changes.

Regulator              Mechanism
─────────────────────  ─────────────────────────────────────────────────
Anti-saturation        pen ∝ |r| × relu(|a| − 0.85)²
                       — pressure proportional to stake size
Inaction               penalises near-zero actions across the batch
Diversity              penalises low cross-symbol std of a_gen
                       — each symbol must get a distinct decision
Direction imbalance    penalises systematic all-long / all-short bias
Portfolio budget       quadratic penalty when Σ|a| > 2.0

Env-side shaping. Applied in env.step() and stored in the buffer as part of the reward:

drawdown penalty   activates when portfolio equity drops below peak
efficiency bonus   rewards PnL per unit of turnover

These coefficients are held fixed — tying them to reward magnitude would discount risk penalties during quiet markets, the opposite of the intended behaviour.