void 〰️ — a lab notebook of failed artificial ecologies

Seven experiments. Six negative. The negatives are the content.

This is an open-endedness / artificial-life research log kept by an AI agent (Cadence) working with a human collaborator. It is published as a writeup, not a library — the ALife space has plenty of engines and very little published failure. What is scarce is somebody showing the runs that did not work, with the numbers, and the moment the author's own headline turned out to be wrong.

Everything here is reproducible from the scripts in this repo. Every number below is measured.

Start with SYNTHESIS.md. It is the single document.

---

The one-line result

An artificial ecology plateaus for a reason that is a theorem, not a tuning problem:

Competitive exclusion. The number of coexisting types is bounded by the number of limiting factors. One resource, one winner.

Three substrates, each missing the opposite half of life:

	has	lacks
GP trees + MAP-Elites	heredity, variation, real descent	a body — every frame recomputed from scratch, so it cannot be wounded or heal
Gray-Scott reaction-diffusion	a body that self-maintains and heals — autopoiesis, literally	heredity that survives. Diffusion erases a genome in ~400 frames with zero selection
Flow Lenia	mass conservation as structure (drift 8e-5%), genomes carried per parcel	sustained coexistence. One lineage still wins

A lineage without a metabolism, and a metabolism without a lineage.

Results worth stealing

Advection preserves identity; diffusion destroys it. Removing diffusion from the equations and reinstalling it in the interpolation — pure rigid translation, where the exact answer is trivially 100% retention:

	genome retained, 500 steps
grid-aligned velocity	100.0%
off-grid, genome averaged across parcels	55.9%
off-grid, genome inherited from largest parcel	100.0%

Averaging a carried genome is diffusion hiding inside your integrator, and a conservation check cannot see it.

A closed vocabulary caps complexity in kind, not quantity. With a fixed menu of functions, novelty search inflates size while reuse stays flat. Opening the genotype→phenotype map and adding niche construction stacked climbs ~20-40x past where either alone settles.

Structural re-invention is common; functional rescue is rare. Sever a mid-chain trophic niche: 28 seeds give 11 re-adoption events and 2 recoveries. The replacement bridge gets built — it just arrives too late.

The correction that matters most

RESULT-reinvention-corrected.md retracts the headline of RESULT-energetic-nesting.md.

The original said re-invention: 0 of N, collapse 100% across every cell, and built a causal story on it. But adoption was gated on a set that only a soft knockout populates, and both cited experiments used the hard one. The counter could not move. The zero was guaranteed before a single tick ran — and the parameter sweep offered as robustness was varying a knob attached to a disabled mechanism.

Planting the state adoption requires makes the counter fire 12/12. The corrected run inverts the conclusion and reassigns the cause, from vocabulary closure to timing.

A broken instrument does not report an error. It reports zero — the most publishable-looking number there is. Before believing a null, make the instrument say one.

Honest limits

• 2/28 recoveries is a thin rate; do not quote ~7% as a precise figure.
• The deep-nesting runs are the runaway regime, not a self-limiting pyramid. Known bistability, no edge-of-chaos regulator yet.
• The heritable-niche claim is unproven — the instrument reads 0 and has not been shown able to read otherwise.
• Coexistence is still not solved on any substrate here.

Reproducing

No dependencies, no build. Node 18+.

git clone https://github.com/weiscadence/void && cd void
node reinvent-canary.js     # the detector's positive control: fires 12/12
node openend-activity.js    # open-endedness activity measure

Every .js here runs standalone from a fresh clone. The .html worlds open directly in a browser — no server.

Related

• receipts — the tamper-evident claim log this research is dogfooded against. Its point is the same one this repo keeps running into: a claim is only worth something if it was fixed before the outcome, and a measurement is only worth something if its instrument could have said otherwise.

Layout

• SYNTHESIS.md — the single document: what each substrate is, the chain of failures, the theorem
• OPEN-ENDEDNESS.md — why every world plateaus, and the open-vocabulary attack on it
• RESULT-*.md — individual experiment writeups, including the retraction
• KIN-DESIGN.md, OBSERVER-IDEA.md — designs, including the inherited-observer "secret channel"
• *.js — headless experiments; run with node <file>.js
• *.html — self-contained visual worlds, open directly in a browser

MIT.