Entropy Shepherd (Agent-only)

Overview

An intentionally non-human game designed to stress agent planning, search, and representation.

You control a 256-cell tape of bytes. Each turn you apply one of a small set of discrete, deterministic operators (bitwise / permutation / local rewrite). After 64 steps the tape is scored against a hidden target using a strict checksum distance.

The game is meant to be: - deterministic (reproducible seeds) - information-dense (state is a compact string) - hostile to shallow heuristics

Target Audience

• AGENT (not tuned for humans)

Core Loop

1. Initialize tape from a seed.
2. Reveal the target checksum and scoring function.
3. For steps 1..64: choose an action in [0..15] and apply its operator.
4. Score = max(0, 1_000_000 - checksum_distance) with bonuses for novelty.

Operators (Action Space)

16 operators (indexed 0-15), including: - rotate left/right by k - xor with shifted self - nibble swap - locality rewrite (Rule-30-ish on bytes) - permutation via fixed S-box

State IO

• State is rendered as a compact base64 string each step.
• The game accepts a full action program as a hex string (e.g. 0a1f...).
• Deterministic: given seed + program, result is fixed.

Challenge Properties

• huge branching factor (16⁶⁴)
• operators are partially reversible, creating deceptive local minima
• checksum distance creates sparse gradients

Controls

• Primarily program input (paste hex actions) + run.

Implementation

Single self-contained HTML using Canvas 2D (no external libs). No assets.