Flow Lenia · commitment dose & memory

The controller must turn with the organism.

One early push genuinely narrows the future. Replaying the same push does not deepen that state—it moves the organism somewhere else.

The surprising result is not that control failed. Control was abundant. What expired was the direction of control: as the system developed, yesterday’s commitment-increasing action stopped pointing toward commitment.

+0.808whole-minus-sum high−low at q36
-0.175extra collapse caused by the second pulse at q40
≈ 0.151physical distance between dose histories at q56
01 · THE TEST

Same ancestor. Same action. Different moments.

At q32 we gave each organism one of eight inventory-preserving pushes. At q36 we identified a high-commitment and low-commitment sibling. Then we either stopped, repeated the same push once, or repeated it three more times. All dose histories were finally compared at the same q56 age and challenged with the same eight actions.

q32
one ancestor
q33
first push
q36
high / low split
q37–45
0 / 1 / 3 repeats
02 · THE TURN

The information split closes in four steps.

q36q40q44q48q52q561 pulse2 pulses4 pulses high − low whole-minus-sum

Mean high−low whole-minus-sum across the same 30 ancestors. All histories are byte-identical through q36. The curves separate only when later pulses occur.

Without a second pulse
+0.139

The initial lead has already lost most of its magnitude by q40, but it still points in the original direction.

With the exact same pulse repeated
-0.036

The direction is gone. The repeated action does not reinforce the earlier state.

03 · CAUSAL ANATOMY

The second pulse helps the low branch catch up.

Dose-1 and dose-2 histories are identical through q36. The only difference is the repeated action at q37. Its effect at q40 is strongly asymmetric.

high branch shift
+0.077 [-0.102, +0.255]
low branch shift
+0.252 [+0.096, +0.442]
change in high−low gap
-0.175 [-0.424, +0.047]

The low branch gains +0.252 whole-minus-sum; the high branch gains only +0.077. The gap shrinks by -0.175. The controller did not become weaker—the state changed what the action meant.

04 · IT STILL MOVED

No accumulation along commitment. Large movement elsewhere.

0.151q56 field distance · 1 vs 2 pulses
0.151q56 field distance · 1 vs 4 pulses
0.148q56 field distance · 2 vs 4 pulses

Every repeated pulse makes a visibly different future. But at q56 the original commitment ordering is gone under all three doses: -0.037, -0.070, and -0.064. The action has causal leverage without a stable global direction.

05 · ONE ORGANISM

The form diverges while the ordering disappears.

Representative ancestor: serene-form-1312. Left is the sibling originally pushed high; right is the sibling originally pushed low.

q36 high siblingq36 low sibling
q36 · after the first push
Information geometry is maximally separated.
q40 high siblingq40 low sibling
q40 · after replay
The same action has already changed meaning.
q56 high siblingq56 low sibling
q56 · same age
Different fields; no stable high/low commitment label.
06 · THE FUTURE REOPENS

The causal contraction belongs to a window.

q36 · immediately after one push
-0.008

High-commitment siblings have a smaller microscopic future repertoire. 95% interval [-0.015, -0.000].

q56 · same age after 1 / 2 / 4 pulses
≈ 0

-0.004 · +0.000 · +0.001. The earlier contraction is not a permanent lock.

This changes the picture. Commitment is not a static substance we can keep adding. It is a transient relation between the organism’s present state and the actions available from that state.

NEXT DIRECT SWING

Let the controller re-read the organism.

Open-loop replay

Choose an action at q33 and repeat that same action at q37, q41, and q45. We now know this moves the field but loses the intended direction.

Adaptive re-selection

At every pulse, re-probe all eight actions from the organism’s current state and choose the action that points toward the current commitment direction. Compare both at q56 under the same challenge repertoire.

If adaptive control preserves the split while replay does not, we have demonstrated a rotating information-geometric control field over development.