26: How Does Finite Selection Route Causality?
Chapter 25 established a non-local effect in a system whose transition rule was local.
A perturbation at one frontier site changed which distant candidates received evaluation slots.
When candidate subsampling disappeared, the far-field effect disappeared with it.
The mechanism was:
LOCAL FRONTIER CHANGE
↓
FINITE EVALUATION SLOTS
↓
CANDIDATE SUBSTITUTION
↓
FAR-FIELD REDISTRIBUTION
That result answered one question.
It created another.
If finite selection changes where causal opportunity is expressed, does it also change how much causal consequence accumulates?
Or does it merely route the same perturbation through different computational pathways?
That distinction became Chapter 26.
And the eventual answer was more interesting than either alternative alone.
Redistribution Is Not Amplification
The obvious experiment would have been to vary the evaluation budget.
But that experiment would have been confounded immediately.
A larger budget evaluates more candidates.
More evaluated candidates produce more expected attachments.
More attachments can create more downstream attachments.
So an apparent increase in causal amplification could reduce to:
MORE COMPUTATION
↓
MORE CONSTRUCTION
rather than:
DIFFERENT ALLOCATION
↓
DIFFERENT CAUSAL AMPLIFICATION
Chapter 26 therefore had to control construction quantity while changing allocation structure.
The primary scientific question became:
At matched expected background construction rate, does strong candidate subsampling produce a larger finite downstream causal consequence than true exhaustive evaluation?
The stronger mechanistic question emerged only later:
If the downstream consequence does not change much, does finite selection nevertheless change the pathway through which the immediate causal effect is realized?
The Perturbation
As in the previous chapters, we started from the same checkpoint and the same frontier probe x.
Two branches were created:
FORCE
x is occupied
PREVENT
x remains empty
Both branches used:
same checkpoint
same probe
same future environment
same cell-keyed randomness
The intervention was transient.
After one continuation update, x was removed from FORCE if it remained occupied.
The finite-horizon outcome was:
G_T(H)
with:
H = 12
where G_T is the cumulative FORCE-minus-PREVENT realized construction difference within the chosen local horizon.
This is not a formal branching ratio.
It can contain overlapping pathways, substitutions, descendants of descendants, and branch-specific frontier changes.
So Chapter 26 calls it:
finite-horizon causal amplification
and nothing stronger.
No criticality claim follows from it.
The Allocation Arms
The continuation compared:
f = 0.10
f = 0.25
f = 0.50
f = 0.75
f = 1.00
unbounded
where f controls the fraction of the PREVENT frontier used to determine the finite evaluation budget.
The true unbounded arm is different.
It evaluates the complete current frontier of each branch.
That distinction turned out to matter.
flowchart LR
subgraph Finite
A[PREVENT frontier F_PREVENT] --> B[Budget B = f × F_PREVENT]
B --> C[Same B used by FORCE]
C --> D[Evaluation count matched across branches]
end
subgraph Unbounded
E[PREVENT evaluates F_PREVENT] --> F[FORCE evaluates F_FORCE]
F --> G[Coverage matched, not count]
end
But before we could interpret any amplification result, we first had to solve a harder control problem.
V1 — We Matched a State, Not a Process
The first version of the experiment calibrated each allocation arm to the same expected construction rate at lag 1.
The natural reference policy was:
f = 0.10
offset = 0
For broader evaluation arms, an additive score offset lowered individual attachment probabilities until the PREVENT branch matched the same expected attachment count.
At lag 1, the calibration worked extremely well.
It looked as if we had separated:
HOW MANY CANDIDATES ARE EVALUATED
from:
HOW MUCH CONSTRUCTION IS EXPECTED
But the offset was solved once.
Then the crystal evolved.
Different selected candidates produced different attachments.
Different attachments produced different morphologies.
Different morphologies produced different frontiers and different probability distributions.
The single calibration no longer represented the changing process.
The V1 audit showed exactly that.
LAG 1
matched expected construction
↓
LAGS 2–12
construction rates drift apart
The lesson was simple:
A controlled initial condition is not a controlled dynamical process.
V1 remained part of the scientific record.
But it was not the experiment we intended.
So we built V2.
V2 — Match the Trajectory
V2 recalibrated the background process at every continuation lag.
A dedicated PREVENT-only reference trajectory used the natural:
f = 0.10
offset = 0
policy.
At each lag τ, that trajectory defined:
C_target(τ)
the expected background construction target.
Then every allocation arm solved a fresh offset on its own current PREVENT state:
flowchart TD
A[For each lag τ] --> B[Reference PREVENT defines C_target(τ)]
B --> C[Arm PREVENT state chooses its candidate set]
C --> D[Solve offset such that expected PREVENT attachments = C_target(τ)]
D --> E[Apply same offset to FORCE]
E --> F[Execute both branches]
F --> G[Next lag]
G --> A
FORCE was deliberately not calibrated independently.
That would have normalized away part of the causal response.
Instead:
PREVENT
defines background policy
FORCE
receives the same policy
and remains free to differ causally
This gave us the control we actually wanted.
The Calibration Itself Became Part of the Experiment
The manipulation was not merely:
same rule
different allocation
That description is too loose.
Expected construction was held constant by trading:
FEW EVALUATED OPPORTUNITIES
with relatively high per-candidate probability
against
MANY EVALUATED OPPORTUNITIES
with much lower per-candidate probability
The broader the evaluation policy, the more negative the calibration offset had to become.
Across the V2 trajectory, the approximate mean offsets were:
f=.10 0.000
f=.25 -1.59
f=.50 -2.49
f=.75 -2.96
f=1.00 -3.28
unbounded -3.28
The reference f=.10 offset is zero by definition.
That is not a measurement.
It is an assertion about the reference policy.
The broader-arm offsets are measurements of how strongly the native attachment rule had to be suppressed to preserve the same PREVENT construction expectation.
The offset also moved over time.
For the broadest arms it became less negative as the trajectory evolved.
That is another reminder that:
Controlling a process requires a moving control when the state itself moves.
The Validity Gate Passed
The full V2 run used:
192 independent groups
768 probes
4 probes per group
12 continuation lags
6 allocation arms
with fresh seed:
20260913
The supported probe regime remained:
occupied_neighbors = 1
The dynamic matching validity gate required:
>= 95% of group × probe × arm × lag records
within 2% of target
and:
every population-level arm × lag mean
within ±2%
Both gates passed.
The record-level pass fraction was:
1.0
and cross-arm PREVENT expected-construction dispersion was effectively numerical zero.
So V2 really did compare different allocation regimes under dynamically matched background construction.
The Frozen Amplification Test
The primary contrast was:
G_T(f=.10)
-
G_T(unbounded)
Strong candidate subsampling versus true exhaustive evaluation.
The predeclared smallest scientifically meaningful effect was:
±0.15 attachments
The result was:
mean difference
+0.00130
95% CI
[-0.08984, +0.08854]
achieved MDE80
0.11536
The complete confidence interval lies inside:
[-0.15, +0.15]
So the result is:
BOUNDED_NEAR_ZERO
at the frozen effect scale.
This is stronger than a nonsignificant result.
It is also narrower than saying the two regimes are equivalent.
The experiment resolves the predeclared:
±0.15
scale.
It would have been:
UNRESOLVED
under a:
±0.10
smallest-effect threshold.
That precision boundary matters.
The scientifically bounded sentence is:
At dynamically matched PREVENT background construction rate, strong candidate subsampling did not produce a mean twelve-step causal consequence differing from true exhaustive evaluation by the predeclared ±0.15 attachment scale.
The Same Mean Outcome Does Not Mean the Same Computation
The primary arm means were:
f=.10 0.1497
unbounded 0.1484
Almost identical.
But the mechanism audit showed that these similar outcomes were not produced in the same way.
To understand why, we had to look inside the immediate ring-one causal effect.
First Filter: Did the Perturbation Survive?
The FORCE intervention can disappear before the continuation even begins.
The intervention-step background loss rate is:
δ = 0.08
If x is removed by that loss draw, FORCE and PREVENT collapse back to the same occupancy state.
Then:
FORCE = PREVENT
↓
E1 = 0
G_T = 0
structurally.
The realized survival fraction was:
0.9141
against the expected:
0.92
So roughly eight to nine percent of recruited probes were not weak causal responses.
The perturbation itself had been annihilated.
That is the first source of structural zero inflation.
Second Filter: Did the Selector Expose the Difference?
Even when x survives, a finite selector may fail to evaluate any ring-one opportunity affected by the intervention.
For an affected frontier set of size k, frontier size F, and finite budget B, the without-replacement protocol expectation is:
P(any affected candidate selected)
=
1 - C(F-k, B) / C(F, B)
So the probability that an immediate ring-one causal effect can be expressed is approximately:
P(E1 active)
=
P(x survives)
×
P(any affected ring-1 candidate selected)
No fitted parameter is required.
The prediction matched the observed active-E1 fraction closely:
arm predicted observed
f=.10 0.363 0.401
f=.25 0.679 0.698
f=.50 0.876 0.880
f=.75 0.914 0.913
f=1.00 0.918 0.910
unbounded 0.918 0.910
The largest discrepancy was under the strongest subsampling arm.
By f=.50 and above, prediction and observation were within about one percentage point.
This explains something that originally looked like an estimator problem.
At f=.10:
~60% of probes
have effectively zero E1
not because the causal system is simply noisy, but because the protocol often prevents the causal difference from becoming expressible.
That gives us a new hierarchy:
flowchart LR
A[PERTURBATION DELIVERED] --> B[PERTURBATION SURVIVES]
B --> C[IMMEDIATE CAUSAL DIFFERENCE EXPRESSED]
C --> D[DOWNSTREAM CONSEQUENCE REALIZED]
A -.->|≠| B
B -.->|≠| C
C -.->|≠| D
The primary f=.10 arm had:
E1 ≈ 0 for ~59.9% of probes
G_T = 0 for ~83.1% of probes
So most delivered perturbations did not produce a realized twelve-step construction difference.
Why the Immediate Mean Stayed Nearly Flat
Before the mechanism audit, the arm-level ring-one means looked almost invariant:
f=.10 0.11305
f=.25 0.10860
f=.50 0.10573
f=.75 0.10757
f=1.00 0.10810
unbounded 0.10824
It was tempting to say that dynamic construction matching mathematically pinned the immediate causal input.
That was wrong.
We tested that hypothesis.
The simple probe-level law failed.
The apparent stability came from something more interesting.
There were two causal channels whose relative weights changed in opposite directions.
Two Ways the Same Local Intervention Can Matter
For an n=1 frontier probe, x has one occupied neighbour and five empty neighbours.
Some of those empty neighbours are not frontier cells before FORCE occupies x.
When x is present, they become new frontier candidates.
Those contribute through a:
FORCE-ONLY / PROMOTION CHANNEL
Other empty neighbours were already frontier cells in PREVENT.
When x becomes occupied, their local neighbourhood changes.
Those contribute through a:
SHARED PROBABILITY-SHIFT CHANNEL
The exact ring-one accounting is:
E1_ring1
=
FORCE-only contribution
+
PREVENT-only selector contribution
+
shared probability-shift contribution
The equality itself is a code-level accounting identity.
We asserted it across:
4,608 probe × arm rows
to floating tolerance.
The scientific result is not that the sum equals itself.
The scientific result is how the channel weights change.
The Causal Pathway Rotates with Evaluation Breadth
Conditioning on the 702 probes where x survived, the mean contributions were:
FORCE-only shared shift total E1
f=.10 0.0694 0.0542 0.1237
f=.25 0.0433 0.0761 0.1188
f=.50 0.0361 0.0798 0.1157
f=.75 0.0337 0.0840 0.1177
f=1.00 0.0330 0.0853 0.1183
unbounded 0.0332 0.0853 0.1184
That is the mechanism Chapter 26 was missing.
As evaluation becomes broader:
FORCE-ONLY / PROMOTION CONTRIBUTION
falls
while
SHARED PROBABILITY-SHIFT CONTRIBUTION
rises
The two compensate.
The total immediate causal input stays near the same arm-level scale even though the computational route changes substantially.
flowchart LR
subgraph StrongSubsampling
A[Finite selector] --> B[FORCE-only promotion dominates]
A --> C[Shared probability shift weaker]
end
subgraph BroadEvaluation
D[Broad selector] --> E[Shared probability shift dominates]
D --> F[FORCE-only promotion weaker]
end
B & C --> G[Similar total E1]
E & F --> G
G --> H[Arm-level mean nearly stable]
So the correct conclusion is not:
allocation does not change E1
and not:
construction matching algebraically fixes E1
It is:
Changing evaluation breadth reroutes the immediate causal effect between selector-mediated promotion and probability modification of already-shared opportunities. Those pathways compensate strongly enough that the arm-level mean ring-one effect remains nearly stable.
That is a genuine mechanistic finding.
Strong Subsampling Makes Selection Matter More
A compressed two-term model reinforces the same picture.
Conditioned on intervention survival, the fitted promotion proxy declines systematically as evaluation broadens, while the shared-shift proxy rises:
arm promotion proxy shared proxy
f=.10 0.580 0.379
f=.25 0.224 0.565
f=.50 0.121 0.608
f=.75 0.038 0.669
f=1.00 0.015 0.686
unbounded 0.017 0.686
These coefficients are descriptive.
They are not literal attachment probabilities.
The exact selected-set decomposition is the mechanistic object.
But the progression is clear:
STRONG SUBSAMPLING
↓
which opportunities receive evaluation
carries more of the causal effect
BROAD EVALUATION
↓
probability changes on already-shared opportunities
carry more of the causal effect
The same intervention can matter because:
a new opportunity receives computation
or because:
an existing opportunity becomes more likely to construct
Finite selection changes the mixture.
Redistribution Still Exists
None of this removes the Chapter 25 result.
Under strong subsampling:
E1_far(f=.10)
≈ -0.0387
True unbounded evaluation has:
E1_far = 0
by construction.
That zero is not an empirical finding.
When every far frontier candidate is evaluated in both branches and their far probabilities are identical, the expected selector-mediated far difference must be zero.
So:
unbounded E1_far = 0
is a correctness assertion.
The experiment’s scientific evidence comes from the contrast:
finite selection
→ nonzero selector-mediated far redistribution
true exhaustive evaluation
→ selector pathway absent
The non-local allocation effect remains real.
But Redistribution Did Not Become Additional Mean Amplification
Now we can interpret the frozen primary result more precisely.
Finite selection changed:
where evaluation slots were spent
how the ring-one effect was divided between causal channels
whether the immediate effect became expressible
Yet the mean twelve-step causal consequence remained bounded near the exhaustive result at the predeclared scale.
A post-hoc analysis asked whether residual differences in immediate causal input could be masking an allocation effect.
At group level:
ΔG
=
α
+
β ΔE1
gave:
α ≈ -0.0063
95% bootstrap CI
[-0.0944, +0.0794]
This does not replace the frozen primary test.
It is interpretive evidence.
And it points in the same direction.
So the stronger distinction is now:
CAUSAL REDISTRIBUTION
≠
ADDITIONAL DOWNSTREAM AMPLIFICATION
Finite selection can reroute causal influence without producing a scientifically meaningful additional mean downstream consequence at the tested scale.
f=1.00 and Unbounded Mean Different Things
In V2:
B = |PREVENT frontier|
for the f=1.00 arm at each lag.
So PREVENT is exhaustively evaluated.
But that same evaluation count is handed to FORCE.
If FORCE has a larger frontier, some FORCE opportunities still compete for a finite number of slots.
Thus f=1.00 is not stale-budget V1 behaviour.
It is a different resource-allocation convention.
Finite arms conserve approximately:
EVALUATION COUNT ACROSS BRANCHES
True unbounded evaluation instead conserves:
COVERAGE
because each branch evaluates its own entire frontier:
B_PREVENT = F_PREVENT
B_FORCE = F_FORCE
That distinction appears directly in FORCE expected construction.
At lag 1:
FORCE expected surplus over target
f=.10 +0.041
f=.25 +0.045
f=.50 +0.048
f=.75 +0.049
f=1.00 +0.036
unbounded +0.093
The unbounded arm lets FORCE evaluate all of the extra opportunities created by the perturbation.
Finite arms hold the evaluation count to the PREVENT-defined budget.
So the f=1.00 versus unbounded difference is best understood as a methodological contrast between two resource-allocation semantics:
FINITE POLICY
matches evaluation count
UNBOUNDED POLICY
matches evaluation coverage
What Happened to Common Random Numbers?
Every arm used the same keyed random numbers.
But pairing only reduces variance if the two arms tend to put the same cells on the same side of their probability thresholds.
The calibration offsets differ greatly.
So the same random uniforms are being compared against very different probabilities.
The group-level correlations with unbounded G_T were approximately:
f=.10 0.13
f=.25 0.05
f=.50 0.12
f=.75 0.44
f=1.00 0.82
The closer the allocation/calibration policy gets to unbounded, the stronger the pairing becomes.
That explains why the primary contrast did not receive the dramatic variance reduction we might have expected from common random numbers.
Again, this does not invalidate the result.
It defines its precision.
The experiment resolved:
±0.15
not:
±0.10
What Finite Selection Actually Controls
The picture is now more specific than the original Chapter 26 question.
Finite selection controls at least three things.
First:
SPATIAL ALLOCATION
It creates outside-cone redistribution by making distant candidates compete for finite evaluation slots.
Second:
CAUSAL PATHWAY
Under strong subsampling, selector-mediated promotion carries a larger share of the immediate local effect.
Under broad evaluation, shared-neighbour probability shifts carry more.
Third:
CAUSAL EXPRESSIBILITY
A perturbation must survive loss and gain access to an affected evaluated opportunity before an immediate causal difference can appear at all.
What finite selection did not do, at the frozen resolution, was create a large additional mean twelve-step causal consequence.
That gives us a much better substrate statement:
Finite computational selection changes where causal influence is expressed, how that influence is routed through local computational pathways, and how often it becomes immediately expressible. Under dynamically matched background construction, those changes did not produce an additional mean twelve-step causal consequence of ±0.15 attachments or greater relative to true exhaustive evaluation.
Chapter 26 Status
CHAPTER 26
STATUS: COMPLETE
Primary frozen result:
BOUNDED_NEAR_ZERO
at ±0.15 attachments
Best bounded claim:
At dynamically matched PREVENT background construction rate, strong candidate subsampling did not produce a mean twelve-step causal consequence differing from true exhaustive evaluation by the predeclared ±0.15 attachment scale.
Mechanistic finding:
Evaluation breadth changes the mixture of immediate causal pathways: strong subsampling increases the FORCE-only promotion contribution, while broader evaluation shifts causal weight toward probability changes on already-shared opportunities. The two pathways compensate strongly at the arm-average level.
Protocol-level zero-inflation result:
Immediate causal expression is gated first by intervention survival and then by whether finite selection exposes an affected ring-one opportunity.
Measured:
Finite candidate subsampling continues to produce selector-mediated far-field redistribution.
Structural assertions:
ring-one channel accounting sums exactly to E1
unbounded E1_far = 0
reference f=.10 offset = 0
Not established:
formal branching process
criticality
subcriticality
supercriticality
phase transition
self-sustaining propagation
coherent structure
individuality
organism
life
No V3.
No further Chapter 26 audit.
What Survived the Hypothesis?
The original hypothesis asked whether candidate subsampling amplified causality.
At the tested scale, it did not.
But the experiment revealed more than a null.
flowchart TD
A[FINITE SELECTION] --> B[Changes which opportunities receive computation]
B --> C[Changes whether perturbation becomes expressible]
C --> D[Changes mixture of local causal pathways]
D --> E[Creates non-local redistribution]
E --> F[Does not produce additional mean<br/>12-step causal consequence<br/>above ±0.15 scale]
That gives us a clearer distinction than the chapter began with:
CAUSAL ROUTING
≠
CAUSAL AMPLIFICATION
And another:
PERTURBATION DELIVERED
≠
PERTURBATION SURVIVES
≠
CAUSAL DIFFERENCE EXPRESSED
≠
DOWNSTREAM CONSEQUENCE REALIZED
Those are not biological categories imported into the model.
They came from the computational substrate itself.
The Next Missing Ingredient Is History
Chapter 18 showed that persistent material state can alter later construction while it remains causally accessible.
Chapter 19 showed that different experiences can leave distinguishable persistent organization without establishing a meaningful common-challenge response.
Chapter 24 briefly attempted to use recent process history, but the tested substrate had no independent material-history state.
Now we can build the missing experiment deliberately.
The next chapter should hold constant:
present occupancy geometry
occupied-neighbour count
promoted-neighbour count
shared-neighbour count
allocation policy
background construction rate
perturbation
while changing:
PERSISTENT MATERIAL STATE
And the perturbation should survive its intended causal exposure by design, rather than recruiting probes and later discarding the roughly eight percent whose intervention is annihilated by loss.
Then the question becomes:
Can two states with the same visible geometry route the same perturbation differently because their material histories differ?
That would be genuinely new.
Not history inferred from geometry.
Not a variable named memory.
A separate persistent state channel whose causal consequence can be measured.
If history changes the promotion channel, the shared probability-shift channel, the total immediate effect, or the downstream consequence under matched present geometry, then the past has become causally active.
If it does not, stored state remains storage without demonstrated use.
That is Chapter 27.
Where We Are Now
The path from Chapters 24 through 26 has become unexpectedly coherent.
CHAPTER 24
A local intervention can alter distant expected construction
through a finite selector.
↓
CHAPTER 25
The effect scales with finite allocation
and vanishes when subsampling disappears.
↓
CHAPTER 26
Finite selection changes:
- spatial redistribution
- causal expressibility
- causal pathway mixture
but not additional mean downstream
causal consequence above ±0.15
under dynamically matched construction.
We started by asking whether finite computation could make a perturbation grow.
Instead we learned how finite computation routes a perturbation.
That is a better answer.
And it gives us a clean next experiment:
What changes when the material itself can remember?
