Saturation Dynamics: All 128 Neurons Are Volatile

Experiment: q4_saturation — 2026-02-11
"The '112 settled + 16 active' picture from a single time point is a snapshot, not the dynamics."

What This Experiment Shows

Looking at a single time step of the sat-rnn, one might see 112 neurons deeply saturated (|h| ≈ 1.0) and only 16 near the decision threshold. This suggests a static picture: most neurons are "settled" and carry persistent features, while a few "active" neurons process new information.

This picture is completely wrong.

When we track the actual dynamics across all 1024 positions, every single neuron — all 128 — is volatile. They all flip their sign repeatedly, with a mean dwell time of just 3.3 steps. The static snapshot was misleading because saturation is transient: a neuron may be deeply saturated at time t but flip at t+1.

"All 128 neurons are volatile. Mean dwell time is 3.3 steps — a rapidly mixing Boolean state."
— q234-results.tex

Key Numbers

128/128
neurons classified as "volatile" (>50 sign flips)
0
frozen neurons (0 flips)
0
settled neurons (1-5 flips)
3.3
mean dwell time (steps)

Sign Flips Per Position

At each time step, many neurons flip their sign from the previous step. The number varies from ~20 to ~80 flips per step, with a mean around 50.

Number of Sign Flips at Each Position

The number of flips is NOT constant — it varies with the input. Some bytes trigger coordinated state changes (many flips), while others change fewer neurons. This is consistent with the input byte sensitivity analysis: rare bytes like 'b', 'T', '*' cause 40+ flips, while common bytes like space and 'a' cause ~32.

Dwell Time Distribution

A "dwell" is the number of consecutive steps a neuron maintains the same sign before flipping. If the model had persistent features, we'd expect long dwells. Instead, the distribution is sharply peaked at 1 step.

Dwell Time Histogram (all neurons, all positions)

25,672 dwells last exactly 1 step (the neuron flips at the very next position). 95% of all dwells are ≤ 10 steps. The longest dwell in the entire trace is 46 steps. This is a rapidly mixing Boolean automaton.

25,672
1-step dwells (flip every step)
11,655
2-step dwells
46
longest dwell (steps)

The Most Volatile Neurons

Some neurons flip more than others. The most volatile neuron (h20) flips 667 times in 1023 positions — roughly every 1.5 steps. Even the least volatile neurons flip hundreds of times.

Top 30 Most Volatile Neurons

NeuronFlipsMean |h|% SaturatedDwell Mode
h206670.7837.3%1
h886530.4110.0%1
h366330.7706.5%1
h86270.7213.5%1
h36210.83322.5%1
h976130.84424.4%1
h1176030.87031.9%1
h195990.7798.0%1
h955970.80010.1%1
h65810.6030.0%1
h8 — the most important neuron — is also the 4th most volatile. It flips 627 times in 1023 steps. This is not a contradiction: h8 is important precisely because it actively processes information at every step, not because it holds a persistent feature.

Co-Flip Clusters

Some neurons flip together — when one changes sign, the other does too. The Jaccard similarity measures how often two neurons flip at the same position. High Jaccard values indicate neurons that are functionally coupled.

Top Co-Flip Pairs (Jaccard Similarity)

PairCo-flipsJaccardIndividual Flips
h3, h364940.650621, 633
h97, h1174550.598613, 603
h44, h694060.592536, 556
h95, h974260.543597, 613
h20, h884270.478667, 653
h19, h204180.493599, 667
h6, h204130.495581, 667
h8, h884130.476627, 653
"A context change causes a coordinated sign flip across a group of neurons."
— q234-results.tex
Co-flip clusters emerge: {h3, h36} flip together 65% of the time. {h97, h117} co-flip at 60%. These clusters correspond to groups of neurons that encode the same contextual feature (e.g., "inside a tag" or "at a word boundary") — they all respond to the same input events.

Saturation Over Time

We sampled the saturation state at positions throughout the data. At each position, we count how many neurons are "saturated" (|h| > 0.999) and how many flip from the previous step.

Saturated Neurons and Flips Over Time

The number of saturated neurons is typically 2-10 per step (out of 128). The number of sign flips varies between 20 and 80 per step. There is no trend: the model doesn't "settle down" over time. Position 1000 is just as dynamic as position 10.

Why This Matters

Static interpretability is misleading. Looking at a single time step would suggest most neurons are "settled" features. The full dynamics reveal that every neuron participates actively. Interpretability must account for temporal dynamics, not just snapshots.
The model is a rapidly mixing automaton. With mean dwell time 3.3 steps and all 128 neurons volatile, the state space is explored rapidly. This is consistent with the finding that 1023/1023 positions produce unique sign vectors.
Co-flip clusters reveal functional groupings. The Jaccard analysis identifies neuron clusters that respond to the same contextual features. This is a different (and complementary) view of the factor map found in the 20260209 experiments.

Source & Related

Papers: q234-results.pdfboolean-automaton.pdf

Programs: q4_saturation.c

Related experiments: Boolean AutomatonNeuron KnockoutOffset AnalysisPer-Prediction Justifications