# Experiments with the historical FUM variants

The central repository is intended to make cross-variant experiments easy without rewriting the historical sources.

Create experiment copies under a new top-level `experiments/` directory and record the source variant used.

## E1 - Reproduce the FUM Demo across seeds

**Source:** `variants/01_FUM_Demo/`

Goal: measure how stable the observed network evolution is across random initialization.

Record at minimum:

- seed;
- final edge count;
- degree distribution;
- spike-rate distribution;
- structural rewiring events;
- topology diagnostics;
- environment-switch recovery behavior.

A useful negative control is to disable one adaptive mechanism while keeping the same seed.

## E2 - SIE ablation

Compare:

1. fixed modulation / fixed learning rate;
2. Demo-era SIE;
3. the explicit SIE stability implementation in `02_early_FUM_tests/SIE_Analysis/`.

Keep the task, seed, initial graph, and simulation horizon fixed. Measure convergence, stability, weight norms, and adaptation after perturbation.

## E3 - Point later topology tools backward at the Demo

Export graph snapshots from `01_FUM_Demo` and convert them into the input representation expected by `02_early_FUM_tests/Knowledge_Graph_Analysis/`.

Question: do the later FUM topology metrics expose useful structure in the earlier integrated FUM simulation?

Do not change the original KG-TDA code merely to force a positive result. Put adapters in the experiment directory.

## E4 - Structural plasticity controls

Run matched trajectories with:

- all plasticity enabled;
- STDP disabled;
- structural rewiring disabled;
- SIE modulation disabled;
- homeostatic mechanisms disabled where available.

Track not only task score but graph health, fragmentation, sparsity, weight stability, and recovery from environment changes.

## E5 - Validate the topology assumptions

The early implementations use particular TDA/graph representations and health metrics. Test those assumptions directly.

Examples:

- disconnected graph versus connected graph with identical node count;
- random, small-world, and scale-free controls;
- rewired graphs preserving degree distribution;
- intentionally fragmented graphs;
- shuffled weights with topology held fixed.

The point is to determine which reported metric actually responds to topology versus representation artifacts.

## E6 - Environment-switch adaptation

The Demo changes environment state during the run. Instrument the change boundary and measure:

- pre-switch performance;
- immediate degradation;
- time to recovery;
- whether topology changes before, during, or after recovery;
- whether SIE signals predict subsequent structural changes.

## E7 - Habituation implementation

Where habituation exists as a state/term but is not meaningfully updated, implement a clearly isolated experimental variant and compare it against the untouched historical behavior.

## E8 - Replace placeholder structural signals

Some early structural-plasticity decisions use placeholder/random proxy signals. Replace one proxy at a time with a measurable endogenous quantity from the network and test whether the change improves stability or adaptation.

## E9 - Complete one Fully Unified Model training frontier

The third repository contains placeholder phase runners and concrete lower-level components.

Choose exactly one incomplete path, state the intended contract from the surrounding architecture docs, implement it **outside the historical source subtree**, and test it against the existing unit/integration surfaces.

This is a useful way to explore what the FUM architecture could support without rewriting provenance.

## E10 - Cross-generation mechanism ancestry

Trace one mechanism across the three FUM repositories, for example:

- SIE;
- graph topology;
- structural plasticity;
- STDP modulation;
- domain clustering;
- persistence/stability.

Document:

1. what exists in each state;
2. what changed;
3. what was removed;
4. what new burden the change addressed;
5. which behavior is actually demonstrated versus proposed.

Stop the comparison at the FUM boundary. FUVDM/VDM should be treated as successor generations, not silently imported as the "correct" version of FUM.