Skip to content

DarkEyes/RC-Arch

Repository files navigation

IDU -- Computational Demonstration

Reference implementation and demonstrations for the architecture in Interpretation, Learning, and Empathy as One Constraint: A Residual-Adequacy Architecture with Accountable Abstention.

This is the linear instance of the architecture: each regime reads content through orthogonal projection onto a private subspace, and the residual is the norm of what that projection fails to carry. The engine requires only a well-defined interpretation map yielding a residual, an MDL-gated commit, and deterministic finite-cost lookups (Theorem 1); linearity is the binding chosen here. Everything is deterministic and reproducible to the bit.

Files

  • idu.py -- the engine: regimes, misfit rho_i, activation, residual r(c), Regime--Act and Act-conflict graphs, the Decision loop (fixed priority HALT -> counter -> residual/action), focus-limited MDL basis expansion, the typed witnesses, and counterfactual simulation.
  • demo_abstention.py -- one engine, three requests, the three witnessed terminals (freeze_time, freeze_resid, freeze_halt).
  • demo_empathy.py -- two agents, shared label, private bases; agreement on clear-cut content, localized divergence on ambiguous content.
  • demo_prerequisites.py -- a target that is stuck until a prerequisite basis concentrates its residual within the focus window w; agent-relative.
  • make_figure.py -- regenerates the three-panel figure demo_figure.png.
  • run_all.py -- runs all three demos and the figure.
  • idu_demo.ipynb -- the same demonstrations as an inspectable notebook.

Run

python run_all.py        # all three demos + figure
python demo_abstention.py
python demo_empathy.py
python demo_prerequisites.py

Requires only numpy and matplotlib.

What each demonstration shows

Demo Phenomenon Result
abstention heterogeneity of not-knowing three structurally distinct witnessed terminals from one engine
empathy bounded mutual understanding forced, self-invisible, localized divergence between shared-label/private-basis agents
prerequisites learnability order prerequisite induced by residual concentration within w, agent-relative to focus capacity

About

Regime-Control Architecture: Cognitive architecture

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors