# Obsidian: a living field

A small artificial ecosystem written in C89 for the IBM RS/6000 named
`obsidian`. The simulation runs natively on AIX 4.3.3 with IBM XL C 5.0.2 or
GCC 2.95.3. No graphics server or external libraries are required.

## Build and run

Unpack the source bundle and run these commands from its directory on AIX.

```sh
sh build.sh xlc
./bin/ecosystem-xl --seed 604 --ticks 4000 --snapshot-every 40 --output my-run
```

Use a new output directory for each run; existing directories are refused.
`sh build.sh gcc` builds `bin/ecosystem-gcc`. On a modern Unix machine,
`sh build.sh host` builds `bin/ecosystem-host`.

Output consists of raw P6 PPM images, `population.csv`, and `journal.txt`.
The program exits after the requested ticks; it installs no background service.
The default run produces 101 images (about 28.5 MiB) at 384 × 256 pixels.

## The experiment

160 organisms begin in a 96 × 64 world with wrapping edges. Each cell can
hold one organism and 0–255 food units. Two circular patches are especially
fertile. The left and right halves have opposite triangular seasons, which
repeat every 800 ticks. Food regrows before organisms act each tick.

Four inherited byte-valued traits affect behavior:

| Trait | Benefit | Cost or constraint |
| --- | --- | --- |
| Movement | Higher chance of attempting a move | Successful moves consume energy; random movement can leave fertile ground |
| Feeding | Harvests more food each tick | Higher ongoing metabolism |
| Longevity | Raises maximum age from 600 to 1,620 ticks | Raises the energy threshold for reproduction |
| Food sensing | More often chooses the richest vacant neighboring cell | Higher ongoing metabolism |

Organisms gain energy from food, die from starvation or age, and reproduce
when they have enough energy and an adjacent vacancy. A child receives 40%
of the parent's energy; reproduction also costs the parent 12 energy units.
Each child trait independently has a 12% chance of a mutation attempt,
adding −16 to +16, clamped to 0–255. The mutation counter counts actual
changes, rather than attempts. There are no predators or programmed species.

Each tick visits the fixed organism-slot array from a random starting point;
actions are sequential and can affect later actions. This is not a uniform
random shuffle. Newborns wait until the following tick. A dead original
participant's slot cannot be reused until that tick ends. These scheduling
choices are part of the model and may influence its outcomes.

Green background intensity shows food. Organism color encodes movement
(red), food sensing (green), and longevity (blue). Colors represent traits,
not species or ancestry. `max_generation` is the greatest ancestry depth
among current survivors, not the number of generations experienced by the
whole population. CSV trait columns are sums; divide by population for means.

## Reproducibility and checks

The simulation uses a masked 32-bit xorshift generator and integer arithmetic.
The same seed and arguments should produce identical frames and logs across
the supported compilers. It checks population counts, occupancy mappings,
food bounds, and living organisms' energy bounds every tick.

This is an exploratory toy model. Trait changes are selection within these
explicit rules, rather than evidence about biological evolution. A single
seed cannot establish a general result. Seasonal contrasts, fixed starting
traits, mutation-free controls, and many seeds would make useful next studies.

### First native run: seed 604

The 4,000-tick XL C run took 79.55 seconds elapsed (70.83 user CPU seconds).
It ended with 6,139 survivors, 32,160 births, 26,181 deaths, and 14,580
actual trait mutations. The population initially overshot its food supply,
declined, and eventually occupied nearly every cell.

| Mean trait | Initial | Final |
| --- | ---: | ---: |
| Movement | 124.33 | 100.88 |
| Feeding | 129.47 | 76.64 |
| Longevity | 139.93 | 147.56 |
| Food sensing | 132.04 | 78.13 |

Lower feeding and sensing means are consistent with selection for lower
metabolic costs in a crowded world, but this interpretation needs controls
and multiple seeds. Mean changes also reflect which initial lineages survive.
Nearly complete occupancy suppresses movement and sensing opportunities;
a next experiment could adjust food regrowth or add disturbances to preserve
spatial diversity.

Validation completed: all images and logs from the 400-tick seed-604 run
matched byte for byte across native XL C, native GCC, and the Linux compiler.
All 101 full-run images decoded correctly, and every recorded population
matched `160 + births - deaths`. A separate 200-tick seed-777 Linux run
passed Valgrind with zero errors and zero leaked bytes. Seven invalid-input
and existing-output checks passed. Linux sanitizer compilation was unavailable
because its installed runtime libraries were missing; Valgrind was used instead.

## Export on Linux

Copy the run directory back using SFTP/SCP and install Pillow and Matplotlib:

```sh
python3 export.py run-604 samples
```

The exporter makes an annotated GIF, still images, and a population/trait
plot. Simulation pixels and measurements come from the native run; captions,
charts, and GIF encoding are produced afterward on Linux. The animation's
playback speed is independent of the simulation's execution speed.