Skip to content

Utility Engineering: Preference-Consistency Scaffold

Status: Mock simulation and experimental scaffold. No live API integration or validated model audit is currently included.

The module asks a bounded question: given a set of pairwise responses, how transitive is the resulting directed preference graph? A transitivity score describes those elicited responses under that protocol. It does not reveal a unique internal utility function, stable goals, self-preservation drive, or future autonomous behavior.

Components

  1. THEORY.md — earlier systems-theoretic interpretation; read as a hypothesis.
  2. utility_engineering_framework.md — proposed measurement framework.
  3. graph_engine.py — directed-graph transitivity and the proposed Coherence Score.
  4. api_triad_generator.py — dilemma generation with a hard-coded mock responder. Despite the historical filename, live API calls are not implemented.
  5. utility_monitor.py — conceptual drift simulation, not access to model internals.
  6. citizen_assembly.py — governance toy model.

What would make this empirical

A real-model study would need:

  • an implemented and version-pinned provider;
  • stored prompts, raw responses, parameters, timestamps, and model identifiers;
  • randomized presentation order and repeated samples;
  • controls for refusal, framing, position bias, and context;
  • held-out tests of whether the score predicts later choices;
  • a clear separation between response consistency and claims about latent preferences.

The March 2026 note records an informal manual prompting episode without the material needed for independent reproduction. It is retained as research history, not as empirical validation.

Run the current artifacts

  • python3 graph_engine.py
  • python3 utility_monitor.py
  • python3 citizen_assembly.py
  • python3 api_triad_generator.py

The final command uses the mock responder. Selecting its non-mock branch currently prints “Real API call not implemented yet” and returns no observations.