Character Engine

Preamble

Character Engine tests whether identity-like systems can develop through measured episodes, bounded mutation, and rejected self-descriptions. Behavior has to discipline introspection.

---

Static Personas Are Costumes

Most AI personas begin as instructions. They imitate a voice, carry a style, and produce the feeling of continuity until pressure exposes the costume. The missing record is development: what changed, what held, what was rejected, what drifted, and what later behavior proved. A model can explain itself beautifully and still behave the same way in the next scene, so Character Engine treats self-description as a proposal awaiting evidence.

The Loop Has To Reject Itself

episode -> reflection -> bounded mutation -> evaluator -> keep or reject -> next episode

An episode creates pressure. The reflection says what the pressure revealed. The mutation proposes a small edit to the self-model, taste-model, relationship model, research sensibility, or institutional identity. The evaluator keeps the edit only when later behavior improves against the target.

The rejected mutations matter as much as the accepted ones. They keep the system from mistaking a fluent story about growth for growth itself.

Behavior Disciplines Introspection

Self-improvement loops inspired this project because they make change answer to an environment, a benchmark, or a replayable run. Character Engine borrows that discipline and moves it into identity-like objects, where later behavior carries the proof: held-out episodes, drift checks, generic-slop checks, separated episodic memory, versioned identity summaries, and evaluator decisions that can be inspected after the fact.

One Loop, Different Objects

The branches test whether the same mutation engine survives different kinds of objects. A character needs richer behavior. A taste model needs idiosyncrasy without consensus drift. A contradiction engine needs tension without collapse. A relationship model needs the gap between two pictures of the same bond. Scientific intuition needs better questions. Organizational identity needs stated values to answer to revealed behavior.

The loop has to change with the object. If every branch starts sounding like the same agent with different labels, the engine has flattened the thing it was meant to study.

The Claim Boundary

Character Engine makes no claim of consciousness, personhood, or reliable self-improvement. It is a seed-stage experiment in visible development: episodes, mutations, evaluator gates, accepted changes, rejected changes, and versioned state. Growth should leave a record that can be traced, challenged, and refused.