Toy World Scoring Functions

Claim Type: implementation_note
Scope: Toy world scoring functions
Depends On: IMPL-003
Status: stable
Claim ID: IMPL-012


Source: docs/processed/legacy_tree/examples/toy_world/scoring_functions.md

Toy world scoring functions

Current canonical framing does not require an explicit ethical cost term. The (M) function below is retained as a legacy evaluation proxy.

Reality cost (\mathcal{F})

Use a proxy for VFE (Variational Free Energy): prediction errors plus complexity.

  • exteroceptive prediction loss
  • interoceptive / homeostatic prediction loss
  • Kullback–Leibler divergence (complexity) between posterior and prior over latents

Legacy ethical cost (M)

Compute from predicted degradation:

  • degrade_self over hippocampal rollout horizon
  • degrade_other over hippocampal rollout horizon, weighted by coupling (\kappa)

Optionally include counterfactual regret:

  • avoidable_harm = degrade(chosen) - degrade(best_feasible_alternative)

Residue update

Increase (\phi(z)) along the executed latent path proportional to realized ethical degradation proxy (legacy (M)).

Open Questions

None noted in preserved sources.

  • IMPL-012
  • IMPL-003

References / Source Fragments

  • docs/processed/legacy_tree/examples/toy_world/scoring_functions.md

REE is developed by Daniel Golden (Latent Fields). Apache 2.0.