A god game by Creative Edge Software, rebuilt by AI coding agents I directed from public specifications and its decoded level files. A generated-level edition runs on this page.
The live game is unavailable right now.
The live game needs a mouse. On a computer it runs right here.
About 14 MB. Measured October 14, 2026; on a laptop or desktop it starts loading as soon as you arrive and takes the screen when it is ready, unless your browser asks to save data. Sound is off in the game on this page.
Your browser asks to save data, so the game waits until you press Run the game; Wi-Fi recommended.
Workers breed inside houses. Pick a Baldie up with the Hand, drop it into a house and the count climbs.
Keyboard: Tab or F6 leaves the game.
Project-made art. The screen shows stills from the procedural edition until the game runs; the game uses generated levels and art made for this project.
Limit Playtested only by me. The rules are the new engine's own tuning, not checked against the original game.
Baldies is a god game from 1995 by Creative Edge Software, first on the Atari Jaguar CD and then on PC in 1996. You breed a tribe of bald workers, put them to work as builders, soldiers and scientists, research inventions and outwit the Hairies. My remake had no source code to start from: one earlier prototype worked from public specifications, and the other decoded the level files and used a disassembler and source-linked Consumption Engine context to read data tables. The original isn't sold on GOG or Steam today. By , 341 people had asked GOG for it.
The Windows desktop remake implements the Hand, four jobs, building, the Shovel, Wings, inventions, the Hairy opponents, saving anywhere, and an attract mode. The browser edition on simoneresearch.com runs the same Python simulation as the desktop remake, on generated levels with project-made art, plus the live rules controls. The desktop build can load the original levels (20 per world). The rules are the remake's own tuning.
The whole world is drawn from 1,280 tiles of 16 by 16 pixels. Trees, rocks and houses are painted into the tiles, not placed on top of them.
A level file is a grid of tile numbers plus who starts where and what may be researched. Everything that made those grids play, the breeding, the jobs, the fights, lived only inside the program.
There are 31 inventions. Each level carries its own settings for the Hairy AI.
The original simulation ran at ten steps a second. The remake runs at twenty.
In the remake, a house holds 10 Baldies and grows to 20, 30, 40 and 50.
How I rebuilt it with AI agents
I directed AI coding agents across two prototypes and combined them into this remake. One prototype was written from public specifications without decompiling the executable; the other decoded the level files and used a disassembler with Consumption Engine context to read data tables such as house capacities. The earliest preserved file is dated October 7, 2026, two playable prototypes were reviewed October 8, and the combined remake completed at commit c37f1b5 on October 9 (calendar time). The desktop build passes 334 automated tests and reads the original levels from a copy you own. The browser edition contains no original game files.
Why it matters
Baldies was made by Creative Edge Software in 1995. I rebuilt it without access to its original source code. The remake's rules now exist as code that can be read, tested and changed, and the sliders below do that while the game runs. The broader question is whether the same approach can help recover file formats and data from other old programs, and connect them with new ones.
Change the rules of the remake while it runs
This is the remake's own engine, running its procedural edition in your browser. Like an arcade cabinet, it starts in attract mode, cycling through its five worlds by itself; take over whenever you like. The sliders change the game's rules while it runs, and Reset puts each one back to the remake's default.
The live game and these rule controls need a computer with a mouse. On this device, the stills above remain available.
Text summary of the 5 live rules (for screen readers and phones)
Walking speed (walk_speed, default 1.2 tiles/s, range 0.3 to 6.0): sets baldies.sim.units.WALK_SPEED so Baldies and Hairies cross tiles faster or slower.
Breeding speed (breed_rate, default 0.02 per worker/s, range 0.0 to 0.5): sets baldies.sim.buildings.BREED_BASE_RATE for workers inside a house.
House capacity (house_capacity, default 1.0x, range 0.25x to 4.0x): scales every building's capacity from the [10, 20, 30, 40, 50] table extracted from the original level data.
Hairy aggression (hairy_difficulty, per-level difficulty, range 0.0 to 1.0): scales Hairy patrol and attack pressure against Baldy settlements.
Cost of using the Hand (grab_cost, default 2.0, range 0.0 to 20.0): sets baldies.sim.units.GRAB_COST in red energy when lifting non-Baldy objects.
How the two prototypes combined, what was extracted, and what I am testing next
What Baldies knew, and where it lives now
Each piece of design data I found in the 1995 program is now a named part of the remake. The design decisions in Baldies, such as house sizes, existed only inside the original program. House sizes are now a table in ordinary Python, and the rules panel on the homepage changes them while the game runs, along with breeding and the cost of the Hand. I call that fluid software. The original pixel art is part of what is worth keeping.
Kept only inside the original program
Where it lives now
The world format: 1,280 tiles of 16 by 16 pixels
the desktop edition
What each tile is: flat, slope, water, house footprint, blocked object
The simulation step: ten per second in the original, twenty in the remake
the fixed-tick game loop
Combining programs: this rebuild is already a merge
One prototype was written from public specifications without decompiling the executable, contributing the Python simulation, interface and procedural art. The other prototype decoded the original level files and used a disassembler with Consumption Engine context to read data tables such as house capacities; its separate decompilation-explored simulation was not used in the combined remake. Recovering formats and rules into readable modules makes it practical to combine working parts from separate efforts.
What I plan to test next
Next I plan to test this two-prototype method on three game designs: a colony simulator, a creature-breeding roguelike and an open-world role-playing game. These are planned explorations, not completed builds.
Baldies was made by Creative Edge Software of Edinburgh, founded by David Wightman, who was its lead designer and programmer; the art was by Alan Duncan, David Brown and Paul Docherty. It began in 1993 as an Amiga game and shipped first on the Atari Jaguar CD in December 1995, published by Atari Corporation, then on PC in November 1996 from Panasonic Interactive Media, with later PC, Saturn, PlayStation and Macintosh releases in Europe and Japan (as Baldy Land) between 1998 and 2003. Media notice. If the rights holders ever re-release Baldies, I'd welcome it and step aside.
Also on this desk.
Reviewed
I direct AI coding agents (the team) and check what they build. Each row gives the result and its main limit; open one for the evidence. All ten projects.
I am testing whether AI agents can keep their work tied to its sources and refuse stale results. Source-blind rebuilds from code bricks passed in 3 of 3 runs on ApprovalBench, a single-package target (October 18).
Limit These are my own runs, with 0 independent reproductions logged. The runtime is private; only a synthetic fixture is public. All limits
Design decision, evidence and linksHide the evidence
Two earlier findings set up this work. Across 37 real-repository tasks, a compiled prose guide scored 48.8% and library retrieval tools 49.0%, against 53.8% for an ordinary workflow. In a separate September 30 study, prepared prose context used 89% more tokens (21,103 vs 11,151) on policy generation.
In one set of test edits on ApprovalBench, every edit that changed behaviour was flagged and no cosmetic edit was.
The Decision Ledger (7df8429) holds 34 decisions: 11 confirmed, 14 missed their bar, 5 inconclusive and 4 open. Each pass bar was recorded before its check, with timestamps kept by the agents themselves.
A starter kit that frames agent builds with module contracts, an approval gate that ties each decision to one hashed action, and an app that checks claims across model providers.
Limit Each stands on its own. None is integrated with the Consumption Engine runtime.
Design decision, evidence and linksHide the evidence
FluidSoftware (v0.1.1) gives coding agents module contracts, scoped context and SHA-256 delivery receipts. Its one matched comparison shows the workflow is feasible, not that it is faster.
BigBoss (v0.1.0) passed 399 tests on Ubuntu and Windows. Keeping the agent isolated from the approval API is the operator's job.
The Council: one recorded run shows a claim check catching an error the critique missed. That is an example, not a measured quality gain.
The Voynich Evidence Atlas, an 8,400-game Pokémon TCG evaluation, a 1,979-commit study of agent work and a 3D city simulator built in Godot.
Limit Each has its own scope: the manuscript is undeciphered, the card-game gains held only against Alakazam, and the commit study is author-run in one repository.
Design decision, evidence and linksHide the evidence
I design the systems and the evaluation protocols. The fastest way to check a claim here is to run the public fixture or follow the non-author review protocol, then send me what you found.
