Skip to content

KLEP 2.0.0-preview.4 - Structural Goals and Bounded Imagination

Pre-release
Pre-release

Choose a tag to compare

@Roll4d4 Roll4d4 released this 22 Jul 14:35

Preview 4 advances KLEP's lower, animal-behavior branch with two deliberately separate proofs: a Goal can use the system's structural self-map to solve for an existing capability, and a closed behavior proposal can be compiled and tested inside a bounded admission process.

This release does not connect a language model. It does not claim general planning, autonomous code generation, learned route invention, or general machine imagination.

Goal Structural Solution V1

The Observer can answer one narrow structural question for an explicitly targeted, otherwise empty root Goal: which already-registered Actions can guarantee the target Key?

The answer is immutable structural evidence, not a command. The Agent alone may adopt the Goal and lease the exact mapped root Actions. Every leased Action keeps its own Locks, lifecycle, and factual output identity. Catalog, tenure, fingerprint, or provider changes invalidate the cached solution.

Zombie 001 demonstrates this with a diagnostic one-step Goal. The Observer maps it to one existing proof Action; the Agent selects only the outer Goal, leases the Action once, and completes the Goal and its Intention only after the persistent proof Key becomes factual at the following boundary. This is a deliberately small proof of the machinery, not Cover/Route planning or model Imagination.

Bounded Imagination and Cover/Route V1

The new engine-free Imagination boundary separates two proposal channels:

  • A Strong Manifest may bind only a trusted project capability/version and closed, bounded, typed arguments.
  • A Weak Conjecture may describe unknown semantics, but it cannot compile, register, score, or execute.

The showcase feeds one hand-authored Cover/Route Strong Manifest through strict compilation, fresh materialization, exactly four disposable deterministic sandbox trials, an explicit accept/reject verdict, and a separate fresh live trial. Sandbox support, confidence, and fitness are admission evidence only. They are not world truth, Memory, learned expectations, attraction, or Agent authority.

Expanded Unity showcase

  • Four civilians and six zombies in a 28 x 18 arena with three physical cover walls.
  • Two persistent, movable ammunition nodes; nobody starts knowing their positions.
  • Timestamped ammo and human sightings can be copied between nearby agents. Freshness comes from the copied observedWorldTick payload, never delivery clocks.
  • Rifles carry six rounds and start with no reserve. A 36-health zombie takes eighteen clean two-damage hits: three complete magazines.
  • Autonomous equipment/ammo seeking, zombie avoidance, friendly-fire safety and Ethics evidence, reload/pickup behavior, and visible non-causal projectile trails.
  • Controls: Tab cycles civilians; W/A/S/D moves; mouse aims; left mouse fires; R reloads; E picks up; Backspace restarts; Exit to OS is included.

Release verification completed before publication: 25 portable suites passed with 4,988 assertion executions and zero build warnings/errors; 27 local runnable suites passed with 5,088 assertion executions, and three compile-only gates also passed; the Unity package imported and compiled in a clean project; and the packaged Windows player reproduced the structural Goal proof through completion.

Developer downloads and the Windows showcase are hosted on itch.io. The readable source, contracts, and headless regression surface remain here on GitHub.

KLEP V2.0 is licensed under MIT.