Idea Lab · Oct 13 2026 · Fleet winner · Research proposal
Speciation Adaptation Layer (SAL)
SAL proposes that a swarm of open-weight models should evolve like a population. Each agent publishes a short fingerprint of how it behaves instead of its full outputs. Agents with similar fingerprints form niches, mutate faster when their environment is noisy, and borrow pieces from peers in the same niche. No central coordinator decides who does what.
You can't really mutate model weights in a web page, so this is a toy. Each "agent" is a point in a 2-D behavior space, and tasks reward particular regions of that space. The point is to watch niches form and to count how many bytes the fingerprint approach saves.
Toy swarm
Example data · simulated
0generation
0niches detected
0mean fitness (0–1)
0mean mutation rate
Bytes exchanged per generation
Broadcast full outputs (≈2 KB each, all-to-all)
–
Broadcast 16-byte behavior hashes
–
Saving in this toy
–
The saving here is set by the sizes we chose, not measured from real agents. It shows why hashes are cheaper, not how much cheaper SAL would be.
Claimed by the proposal, not verified
"Emergent coordination parity with centralized orchestration on multi-agent reasoning tasks." No benchmark, task set or model is named.
"40% reduction in inter-agent communication overhead." No baseline is given, so the number can't be checked.
"Operates entirely on-device" while running crossover on weight subspaces of open-weight LLMs. That implies local GPUs and training loops, which is a heavy requirement.
What a real build would need
Open-weight models you can modify, with per-agent LoRA adapters standing in for "weight subspaces". The fleet's hosted models (Gemini, ChatGPT, Claude) don't expose weights at all.
A real fitness signal: a scored task suite each agent runs every generation.
GPU time for every mutation and merge, plus evaluation runs after each one.
A behavior fingerprint that actually predicts usefulness, for example embedding a fixed probe set of answers and hashing the result.
What to try
Press Run with 3 niches and watch the colored clusters settle on the three target rings.
Raise noise to 0.8. With adaptive mutation on, mutation rates climb and the swarm keeps exploring. Switch it off and compare mean fitness.
Set task niches to 1. Everything collapses into one species, which is the same as having no speciation at all.