← All Experiments

Experiment: Correlated Anima Beliefs

Source: exp_anima_correlated.sx — Validates Conjecture 7.1

Hypothesis

When two belief agents observe correlated evidence streams (e.g., rain and wet ground), learned interaction should transfer information between them, reducing joint prediction error. When streams are uncorrelated, interaction should correctly decouple and add no spurious signal. Additionally, sweeping the desire coupling parameter \(c\) should reveal an optimal regularisation strength.

Method

Two sub-experiments. First: paired belief agents observe correlated Bernoulli streams (\(P(\text{wet}|\text{rain}) = 0.9\), \(P(\text{wet}|\neg\text{rain}) = 0.1\)) versus independent streams, with and without learned interaction. Second: desire coupling sweep from \(c = -1.0\) (fully misaligned) to \(c = +1.0\) (fully aligned).

Experiment 1: Correlated vs Independent Streams

Agent A observes "rain" (\(p = 0.6\)). Agent B observes "wet ground" (conditionally dependent on rain). Interaction is learned via meta-gradient on the \(\alpha_{ij}\) matrix (Theorem 4).

Condition Joint Loss Improvement
Independent (no interaction) 0.000203 —
Learned interaction (correlated) 0.0000921 55% reduction
Learned interaction (uncorrelated) 0.000201 <1% (correct decoupling)

Result 1

Learned interaction on correlated streams achieves a 55% reduction in joint prediction loss (0.000203 → 0.0000921). Critically, when streams are uncorrelated, the learned interaction strength \(\alpha_{ij}\) converges to near zero, correctly decoupling the agents. This validates the information-theoretic prediction: the interaction matrix learns the conditional dependence structure.

Experiment 2: Desire Coupling Sweep

Single agent with desire coupling \(c\) swept from \(-1.0\) to \(+1.0\) in increments of 0.25. Each configuration runs 500 steps on a stationary stream (\(p = 0.7\)).

Coupling \(c\) Final Loss Interpretation
-1.0 (fully opposed) 0.000510 Over-regularised
-0.5 0.000430 Optimal regularisation
0.0 (neutral) 0.000625 No regularisation
+0.5 0.000590 Mild confirmation bias
+1.0 (fully aligned) 0.000602 Confirmation bias

Result 2

Misaligned desire at \(c = -0.5\) achieves the lowest loss (0.000430), a 31% improvement over neutral. Full misalignment (\(c = -1.0\)) over-regularises, increasing loss to 0.000510. At \(c = +1.0\) (aligned), the loss is 0.000602 — worse than neutral because confirmation bias amplifies observation noise. The U-shaped curve around \(c \approx -0.5\) confirms the bias-variance tradeoff predicted by Theorem 7.

Analysis

The correlated-stream result is the key finding. The interaction matrix \(\alpha_{ij}\) learns the mutual information between streams. When \(I(X_A; X_B) > 0\), the meta-gradient drives \(\alpha_{AB} > 0\), allowing information transfer. When \(I(X_A; X_B) = 0\), \(\alpha_{AB} \to 0\). This is exactly the behaviour predicted by Conjecture 7.1.

The desire sweep extends the single-agent result from exp-anima-deep by mapping the full coupling landscape. The optimal coupling is not at the extremes but at a moderate misalignment, consistent with the regularisation interpretation: too much opposition introduces excessive bias; too little allows unchecked variance.

Conclusion

Conjecture 7.1 validated. Correlated streams benefit from 55% loss reduction via learned interaction. Uncorrelated streams correctly decouple. The desire coupling optimum at \(c = -0.5\) confirms the Bayesian regularisation mechanism of Theorem 7. The interaction matrix learns conditional dependence structure without supervision.

Reproducibility

# Clone and build
git clone https://github.com/senuamedia/lab.git
cd simplex && ./build.sh && cd ..

# Clone theorem-proof
git clone https://github.com/senuamedia/theorem-proof.git
cd theorem-proof

# Compile
../simplex/build/sxc exp_anima_correlated.sx -o build/exp_anima_correlated.ll

# Link with runtime
OPENSSL_PREFIX=$(brew --prefix openssl)
clang -O2 build/exp_anima_correlated.ll \
  ../simplex/runtime/standalone_runtime.c \
  -I"$OPENSSL_PREFIX/include" \
  -L"$OPENSSL_PREFIX/lib" \
  -lssl -lcrypto -lm \
  -o build/exp_anima_correlated

# Run
./build/exp_anima_correlated

Related Theorems