SEESWM

SEESWM is a research prototype for experimenting with small neural agents that exchange messages in a simulated grid world. It includes multiple agent specializations, graph topologies, world-model components, training scripts, and exploratory evaluation utilities.

Status: proof of concept. This repository does not contain a trained model checkpoint, peer-reviewed result, safety evaluation, or evidence of a production-ready agent system. The checked-in results are small exploratory runs and should not be read as validation of emergent intelligence.

Research question

Can a graph of specialized agents develop useful coordination or division of labor that a carefully matched monolithic model does not?

The codebase provides infrastructure for investigating that question. The tracked evidence does not answer it yet: the committed comparison uses untrained networks and its nominal baseline is not parameter matched.

What is implemented

• perception, reasoning, memory, and planning agent variants;
• message passing over small-world, modular, hierarchical, and other graphs;
• a configurable resource-and-hazard grid environment;
• experimental world-model, metacognition, neuromorphic, and evolutionary modules;
• ablation, scaling, generalization, statistics, and interpretability helpers;
• training and exploratory evaluation scripts under experiments/.

These modules are experimental scaffolding. Their presence does not establish that every proposed mechanism has been trained, validated, or integrated into a single end-to-end system.

Tracked evidence

Two JSON artifacts are committed so the current claims can be audited directly.

Fresh-network comparison

results/hypothesis_validation.json was recorded on 2025-12-19. Its generator constructed new networks and evaluated them without a training step. The artifact does not record a source revision or seed. The current validate_hypothesis.py retains the fresh-network scope but corrects the baseline sizing and role-count invariants, so it is an extension of that experiment rather than a bit-for-bit reproducer.

Recorded experiment	Scope	Recorded result
Random regression comparison	10 trials	swarm MSE 0.4486 vs baseline MSE 0.4522; swarm lower in 7/10 trials
Internal aggregation-score sweep	5 trials per size	highest mean score at 20 agents (0.0761)
Random topology sweep	10 trials per topology	modular graph had the highest internal score (0.0705)
Grid-world rollouts	10 episodes	reward 0.509 ± 2.235; survival 94 ± 18 steps; exploration 1.56% ± 0.72%

The comparison is not equivalent-parameter: the artifact records 563,520 swarm parameters and 44,752 baseline parameters. The MSE difference is small, and no confidence interval or significance test is recorded for it.

Untrained scaling sweep

results/emergence/emergence_scaling_results.json was recorded on 2026-01-03 for 4, 10, 20, 50, 100, and 150 agents. The artifact does not record a checkpoint identifier or invocation arguments, and no model checkpoint is tracked in this repository. It should therefore be treated as an untrained/default scaling sweep unless independent provenance is supplied.

The recorded mean reward peaks at 10 agents (1.236) and falls to 0.800 at 100 agents, while both parameter count and message count increase with swarm size. The historical artifact's “phase transition” labels came from a local slope-change heuristic; they are descriptive flags, not statistical evidence of a physical or learned phase transition. The old plot is retained beside the artifact for provenance, but is not presented here as a research result.

What these artifacts do not show

• a trained swarm outperforming a trained, compute-matched baseline;
• 100% transfer or generalization to unseen environments;
• statistically validated emergent specialization;
• robustness, graceful degradation, alignment, or safety properties;
• results outside one simulated grid-world family;
• reproducibility across hardware, dependency versions, or independent teams.

Reproduce and extend

Python 3.10 or 3.11 is recommended.

python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e '.[dev,analysis,viz]'

# Unit tests
python -m pytest

# Re-run the fresh-network experiments
python experiments/validate_hypothesis.py --device cpu --seed 0 \
  --output results/hypothesis_validation.local.json

# Quick fresh-random behavioral/scaling smoke test
python experiments/emergence_scaling_analysis.py --device cpu --seed 0 --quick \
  --output results/behavioral-local

# Train a schema-v2 candidate checkpoint, then evaluate that exact policy
python experiments/train_for_validation.py --device cpu \
  --save results/models/candidate.pt
python experiments/validate_rigorously.py --device cpu --seeds 10 --episodes 10 \
  --max-steps 100 --model results/models/candidate.pt \
  --output results/validation-local

Write new outputs to a separate path so the committed evidence remains intact. The candidate writers train_for_validation.py, train_specialized.py, and train_specialization.py write restricted-loader-compatible schema-v2 checkpoints containing the swarm configuration, policy head, environment, and native-type metrics. Other historical training scripts have not been migrated and their outputs are not accepted by the schema-v2 evaluator. The evaluator requires PyTorch 2.10 or newer, uses the restricted loader, bounds checkpoint size, and rejects older schemas or any configuration, topology, agent, edge, or policy mismatch instead of falling back to random weights. Its diagnostic report records the checkpoint SHA-256. These controls do not make arbitrary third-party .pt files safe; use only a checkpoint with trusted provenance and a verified digest. See SECURITY.md. The diagnostic is still not a substitute for trained controls or independent validation.

Minimum credible next experiment

1. Train the swarm and a parameter- and compute-matched monolithic baseline.
2. Pre-register primary metrics, seeds, stopping rules, and exclusion criteria.
3. Evaluate on tasks that require information sharing, plus held-out world layouts and at least one established multi-agent benchmark.
4. Report per-seed results, confidence intervals, effect sizes, and failures.
5. Publish checkpoints, raw trajectories, environment versions, and an exact dependency lock.

Until then, SEESWM is best understood as an experimental framework and source of testable hypotheses, not evidence that collective intelligence has emerged.

License

MIT. This project is not affiliated with or endorsed by any cited researcher, institution, benchmark, or model provider.