15: Can the Past Redirect the Future?

The last two chapters have narrowed the question to one local event.

The What Does One Attachment Cause? chapter forced a single attachment and found an immediate causal effect consistent with the local rule’s mechanical prediction. A positive transient cumulative consequence was not established, while the late transient accumulation rate became practically consistent with zero under the frozen criterion.

The previous chapter then showed that finite computation can redistribute evaluation opportunity outside the one-step reach of the local rule, change how the perturbation is expressed, and gate whether affected opportunities receive evaluation at all — while resolving no meaningful change in mean twelve-step consequence at the declared ±0.15 scale.

Every variable in both chapters was a fact about the present. Current occupancy. Current frontier. Current budget. Current probabilities.

But two crystals with the same visible geometry can still differ in hidden state.

A real history-dependent system would have to produce such a difference from its own past. This experiment does something narrower: it writes the hidden state deliberately and asks whether that difference is sufficient to change the response to the same perturbation.

So this is not yet a test of endogenous history encoding.

It is a test of whether hidden material state can matter causally when the visible present is held fixed.

This is not the question of whether the crystal has memory. The Crystal Gets a Past refused that word when a causal past turned out not to be a readable one, and Can Experience Change the Material? refused it again when distinguishable traces failed to produce the required differential response.

The question here is smaller and prior to all of that:

Can two states with the same visible geometry respond differently to the same perturbation because they contain different hidden material state?

First establish whether hidden state can change the response.

Only later ask whether the crystal can create, preserve or use such state on its own.

Same Shape, Different Hidden State

The experiment gives the crystal a second kind of state.

Some occupied cells carry a decaying scalar value that contributes to the ordinary attachment score:

The material is deliberately weak and transient.

Its half-life is six updates and the trace has already aged three before the test begins, so each carrier starts at about 0.707. Two cells carry it, for a total starting mass near 1.414.

Newly attached cells do not inherit the trace, and the trace neither spreads nor transfers. While a carrier remains occupied its strength decays; if the carrier is lost, its material disappears with it.

That weakness is the design. We are not constructing a memory architecture with retention policies and propagation rules — the Can Experience Change the Material? chapter built something like that and found the interesting question was elsewhere. We are giving the crystal one hidden variable and asking whether it can matter causally at all.

And note the phrase that has to be used carefully from here on. The conditions do not have the same complete state. They have the same visible occupancy geometry and different hidden material state.

That matching is exact at the intervention checkpoint. ACCESSIBLE, REMOTE and ERASED begin from the same occupied set, birth-time map, step, stream seed and probe. The only difference between the history arms is which already-occupied cells carry the hidden scalar.

The invisibility is the entire point:

SAME VISIBLE GEOMETRY
≠
SAME COMPLETE STATE

Accessible, Remote, Erased

Three conditions, and the choice of primary comparison matters more than it looks.

Accessible. Two occupied cells near the probe carry the trace, and the probe’s sole occupied neighbour is always one of them — guaranteeing that the stored state is locally causally accessible to the perturbation.

Remote. The same number of carriers with the same material mass, placed beyond the twelve-step direct local reach of the probe.

Erased. No material at all.

The primary contrast is accessible versus remote, not accessible versus erased. The intention is to hold material quantity fixed while changing whether that material lies on a direct local causal route to the probe.

That does not automatically make REMOTE a perfect null — a point the first experiment will expose.

At the intervention point, visible occupancy, probe geometry, external input, random-number construction and perturbation are matched. Evaluation is true unbounded in all three arms, deliberately removing the finite-selector routing mechanism isolated in the previous chapter.

The futures are then allowed to diverge normally.

The First Experiment Wasn’t the Experiment

It appeared to work. The immediate causal response differed sharply between accessible and remote conditions, and the twelve-step consequence looked lower under accessible history too.

Then an audit of the implementation found that the intervention was not the intervention.

The intended design was FORCE occupying x for one causal exposure while PREVENT kept it empty for the same exposure. What the code did was insert x in FORCE, and start x empty in PREVENT — leaving the PREVENT branch free to attach x naturally during the first growth update.

An empty cell at the start of a control branch is not the same thing as a cell being prevented from appearing.

Worse, the contamination was correlated with the treatment. The accessible trace deliberately included x’s only occupied neighbour, which raises x’s own attachment probability. The audit recovered exactly how much:

probability PREVENT naturally attaches x

accessible   0.428
remote       0.377
erased       0.378

Five percentage points. The condition under test was changing whether the control branch remained a control.

So the downstream result is:

INVALID

Not failed. Not inconclusive. Not underpowered.

Invalid.

More data would only have measured the wrong intervention more precisely.

This adds another failure mode to the book:

too imprecise to answer
≠
valid experiment against a failed hypothesis
≠
invalid intervention measuring the wrong causal contrast

Something Survived the Failure

One quantity escaped the PREVENT-x defect, for a structural reason.

The immediate expected causal response was computed from the branch probabilities before the realized growth step. Whether x later reappeared in PREVENT could not affect a number that had already been calculated. So the immediate effect is measurable even in the invalid run:

ΔE₁ (accessible − remote) ≈ −0.0182

with a narrow interval entirely below zero.

Accessible hidden material reduced the immediate causal response.

Which is backwards. The material gain is positive; the material raises attachment probabilities. Adding it should, on the obvious reading, make the perturbation matter more, not less.

That contradiction is where the mechanism is.

Why More Material Produces Less Response

The attachment rule is logistic, and a logistic function does not convert score into probability at a constant rate. Its slope is:

Accessible material changes the baseline operating point of the shared frontier candidates around the probe.

For the affected candidates, that operating-point shift reduces the local slope of the logistic response.

FORCE then adds the same local score contribution as before, but the resulting probability increment is smaller.

    flowchart TD
    A["Accessible hidden material"] --> B["Higher baseline probability on shared candidates"]
    B --> C["Candidates move toward the flat part of the logistic"]
    C --> D["Same perturbation produces a smaller Δp"]
    D --> E["Reduced immediate causal response"]
  

The V1 immediate-effect audit also supplied a candidate-level mechanism check. That immediate quantity remained interpretable because it was computed before the contaminated PREVENT growth step.

For accessible versus erased, the reduced probability increment on shared candidates contributed about −0.01814, against a total immediate difference of about −0.01930. Under that audit, the operating-point effect accounts for most of the measured immediate difference.

The corrected V2 experiment below independently tests whether the negative immediate contrast survives the repaired intervention. It does not rerun this candidate-level saturation decomposition as a new primary test.

It is worth being precise about what changed, because the loose version of this sentence is wrong. The response rule did not change. The logistic is the same logistic, with the same parameters, computing the same function. What changed is the operating point at which the perturbation acts:

Locally accessible hidden state changed the effective causal sensitivity of the same fixed response rule.

The V1 candidate-level audit supplies the operating-point explanation for that sign, and corrected V2 independently reproduces the sign itself. What V2 does not do is repeat the decomposition. The mechanism is therefore consistent with the V1 audit and replicated in sign, not independently re-derived under the corrected intervention.

That is sharper than history changes response. The rule is fixed; the state moves the perturbation to a different operating point on that rule.

The Remote Arm Found a Back Door

The audit turned up a second problem, and it should feel familiar.

Through the material dynamics alone, remote carriers were beyond the probe’s twelve-step local causal reach and should not have affected the probe locally.

Yet the protocol produced a tiny local difference.

Not by propagating. By calibration. The protocol dynamically matches expected background construction, exactly as the previous chapter’s corrected design required — and it does so with a global score offset. Remote material changes expected construction where it sits; the controller compensates; the compensation applies everywhere, including near the probe.

remote material
↓
global expected construction changes
↓
calibration offset changes
↓
local probabilities shift slightly

The previous chapter found that a global computational mechanism can couple spatially separated regions. Here the same structure appears one level up: our own experimental controller had become a coupling channel between regions the physics kept apart. A compensator that acts globally is, by construction, a path between everything it touches.

That is not a property of the material. It is a property of the instrument, and it means the remote arm was not automatically the clean null it was assumed to be.

Fix the Experiment, Not the Hypothesis

Three corrections, with every scientific parameter frozen — same gain, same half-life, same history age, same horizon, same effect threshold. Nothing was tuned. These are construct-validity repairs.

The intervention. PREVENT now explicitly blocks x during lag one; FORCE explicitly contains it for one full causal growth exposure. After lag-one growth and loss, x is removed from both FORCE and PREVENT and its absence is asserted. From lag two onward, both branches return to ordinary dynamics.

The control. V2 keeps the global dynamic construction-matching controller, so the controller remains part of the experimental instrument. The repair is therefore not to pretend that REMOTE is automatically isolated from it.

Instead, remote carriers are matched to accessible carriers on their baseline influence on the frontier: exactly the same number of adjacent frontier cells, with total baseline attachment-probability mass matched within the frozen 0.05 tolerance, while remaining beyond the twelve-step direct local reach.

After that repair, the remote-minus-erased immediate difference falls to about 8.2 × 10⁻⁵.

So the controller-mediated leakage is negligible at lag one on this measure.

The same V2 run also contains a twelve-step REMOTE-versus-ERASED secondary control:

REMOTE − ERASED

ΔG_RB       ≈ −0.052
95% CI      ≈ [−0.213, +0.051]

That interval spans zero, but its achieved MDE80 of about 0.172 exceeds the frozen ±0.15 scale, so the experiment could not have certified the remote arm as practically null over the full horizon even if it were.

So REMOTE should not be promoted into a demonstrated zero-effect arm.

It is the matched comparator the primary experiment declared: equal material mass, identical visible occupancy geometry, no direct local material route at the intervention, and closely matched background frontier influence.

The three contrasts are at least mutually consistent. Accessible-minus-erased came out at −0.450, interval [−0.705, −0.227], which is close to the sum of the accessible-minus-remote and remote-minus-erased differences. That is arithmetic rather than evidence, but a decomposition that failed to add up would have been a warning sign.

The primary estimand remains:

ACCESSIBLE − REMOTE

under this calibrated experimental regime.

The estimator. The realized twelve-step attachment difference is noisy — a sum of Bernoulli outcomes measuring a small effect. So the primary quantity becomes the expected local causal difference at each lag, summed over the horizon:

At every lag, the FORCE and PREVENT probabilities are evaluated on their own realized current states. The branches are therefore free to acquire different occupancy, material and frontier geometry through actual stochastic events.

Expectation replaces only the final Bernoulli measurement of construction opportunity at each realized state. It does not replace the trajectory.

The secondary realized estimator is the corresponding twelve-step cumulative difference in actual attachment counts inside the same local support, again excluding x.

This is worth stating carefully, because it is easy to misread. Both branches still evolve through actual stochastic events — cells attach, cells are lost, geometry diverges, material decays. The expectation is used only to measure the causal difference at each realized state more precisely than a single coin flip per candidate would allow. The realized outcome is kept as a secondary check.

The corrected run used 192 groups and 564 supported probes, and passed every validity gate: group coverage, dynamic matching, population matching, intervention assertions, remote-carrier matching. Only then is it worth interpreting.

The Immediate Effect Replicates

ΔE₁ (accessible − remote) = −0.01499     95% CI [−0.01725, −0.01281]

Same sign, with a slightly smaller magnitude.

The immediate hidden-state modulation survives the corrected intervention and control design.

So the sensitivity reduction is not an artifact of the broken PREVENT semantics, not an artifact of the old remote placement, and not the calibration leak. With visible geometry matched and the intervention properly implemented, hidden material state changes the immediate causal response of the same perturbation.

The Later Future Changes Too

Over twelve updates:

ΔG_RB       = −0.397     95% CI [−0.679, −0.119]
ΔG_realized ≈ −0.357     95% CI [−0.673, −0.040]

The expected estimator and the noisier realized estimator agree in direction and rough magnitude.

Under the corrected, dynamically calibrated protocol, locally accessible hidden material produces a more negative twelve-step FORCE-minus-PREVENT consequence than equal hidden material placed on the matched remote route.

Both estimators support that direction.

The experiment does not establish that REMOTE itself is a zero-effect arm, nor does it yet establish the predeclared minimum effect magnitude.

And here the frozen decision rule does something that a looser protocol would have let slide.

Direction Is Not Magnitude

The predeclared smallest effect of interest was ±0.15.

The result separates two questions cleanly.

For the expected estimator:

mean       −0.397
95% CI     [−0.679, −0.119]

The interval lies entirely below zero, so the negative direction is supported.

But its upper endpoint, −0.119, still includes effects whose magnitude is smaller than 0.15. The interval therefore does not establish the predeclared minimum magnitude.

The frozen protocol also required sufficient prospective precision:

achieved MDE80     ≈ 0.357
required           ≤ 0.150

That gate fails as well.

So:

DIRECTION
SUPPORTED

PREDECLARED MINIMUM MAGNITUDE
UNRESOLVED

Those are compatible statements.

Calling the whole hypothesis SUPPORTED would claim a minimum magnitude the experiment did not resolve.

Calling the entire result INCONCLUSIVE would throw away a directional result supported independently by both the expected and realized estimators.

The frozen primary verdict remains UNRESOLVED, while the negative direction is an established secondary statement about that result.

The Trace Fades While the Difference Grows

Now the question the main test does not answer.

The material decays. The causal difference accumulates. Is the later effect simply proportional to how much material is still present — material persists, material keeps pushing, effect persists?

The observed trajectory does not support that simple proportional-dose account.

The trace starts at a total mass of about 1.414. It falls below half that around lag 4. At the preceding lag, the cumulative expected causal difference was only:

−0.0995

against a final twelve-step value of:

−0.3972

So roughly 75% of the final causal difference accumulated after the material had already fallen below half its starting mass.

The trace drops below a quarter around lag 8. After that point, a further −0.141 accrues — about 36% of the final effect, after the trace has lost three quarters of its strength.

Split into descriptive epochs:

EARLY    lags 1–4     −0.120
MIDDLE   lags 5–8     −0.158
LATE     lags 9–12    −0.119

Descriptively, the accumulated effect is not confined to the period when the trace is strongest.

The middle epoch contributes at least as much as the early epoch, even though the original material trace has already substantially weakened.

The late epoch still contributes on average, at a point where mean accessible material has fallen to around 0.183. Its interval includes zero, so this remains a description of the trajectory rather than a separate confirmatory result.

A regression of the per-lag increment on contemporaneous material mass points the same way, but its rows are group-by-lag and therefore not independent, so it is reported in the ledger as descriptive only and carries no weight here.

The closeout diagnostics do not establish that material amount is irrelevant. The material state caused the initial sensitivity difference, and the experiment does not isolate a residual-material component from a trajectory-mediated component.

The narrower descriptive conclusion is:

The observed time course does not look like a simple model in which the later causal increment merely tracks the instantaneous amount of material remaining.

A Trajectory-Redirection Interpretation

The material-decay analysis leaves a suggestive pattern: substantial cumulative difference accrues after the original trace has weakened, while simple contemporaneous material-mass diagnostics do not track the increment well.

One plausible causal account is:

hidden material state
↓
changes immediate causal sensitivity
↓
changes which construction events occur
↓
changed events alter later geometry and state
↓
later geometry changes what the perturbation's consequences can do
↓
causal difference continues accruing as the original trace decays

One plausible interpretation is that early hidden-state modulation changes construction events, those events alter later geometry and state, and those altered states then contribute to subsequent divergence.

But this is an interpretation of the time course, not a mediation result.

The experiment does not partition the later difference into:

continued action of the remaining trace

versus

consequences already embedded in changed downstream states

Both may contribute.

What is directly measured is narrower:

The cumulative causal difference continues to change after the original material trace has substantially weakened.

That pattern is consistent with trajectory redirection, but it is not a measured transfer of causal influence from the trace into some new carrier. Nothing was tracked moving from one representation to another.

An earlier draft named this Material-State Trajectory Redirection. The label is less important than the boundary on the claim: trajectory redirection is a plausible interpretation of the observed temporal pattern, not an independently isolated causal channel.

Being Conditioned by Hidden State Is Not Reading the Past

This is the distinction the experiment actually earns, and it reframes several earlier failures.

The The Crystal Gets a Past chapter established that a past can be causally consequential without being recoverable as a stable history signature.

The Can Experience Change the Material? chapter went further: even persistent, accessible and spatially distinguishable traces failed to produce the required differential response to a common challenge.

Those chapters separated several questions we had initially treated as one:

did the past leave a consequence?
can the past be distinguished?
can the present differentially use that consequence?

This chapter asks a prerequisite question and gets a positive answer: if the system contains a hidden state difference of the tested kind, that difference can alter its causal response even though the visible occupancy geometry is unchanged.

The experiment does not establish that the crystal generated that state from its own past.

RECORD-LIKE ROUTE               STATE-CONDITIONED ROUTE

past                             past
↓                                ↓
stored representation            changes sensitivity
↓                                ↓
later retrieval                  changes events
↓                                ↓
behaviour depends on past        changes later state
                                 ↓
                                 behaviour depends on past

A genuine history-dependent system could in principle use either route.

Only the first requires a representation that can later be read as a record of earlier events.

This experiment establishes the causal-sufficiency prerequisite for the second route:

A hidden state need not be readable as a record in order to alter future response.

Whether the crystal can generate such a state from its own history remains a separate question.

Do Not Call It Memory

The temptation is obvious. There is hidden state. It changes later response. Its consequences outlive most of the trace. Why not memory?

Because we wrote the state. It was placed by the experiment, not acquired by the crystal. Nothing here encoded anything, selected what to retain, retrieved anything, reconstructed a past event, distinguished one history from another, or improved at anything.

What has been demonstrated is a more primitive causal capability:

EXPERIMENTALLY WRITTEN HIDDEN STATE
↓
CAUSAL RESPONSE MODULATION
↓
DOWNSTREAM TRAJECTORY DIFFERENCE

For this to become endogenous history dependence, the crystal would also have to generate the relevant hidden-state difference from its own prior dynamics.

A stronger memory claim would require additional machinery or evidence: endogenous encoding, retention, discrimination, retrieval or some other demonstrated use of stored history.

The Can Experience Change the Material? chapter’s failure is the reminder of how much further there is to go: two histories that leave distinguishable traces still did not produce a differential response to a common challenge. Here the response does differ, but the experiment deliberately placed hidden material where it entered the local causal mechanism and compared it with matched material positioned outside that route.

The crystal did not discover or encode that placement itself.

For the same reason, be careful with the word experience. This experiment models the causal effect of a candidate retained material trace. It does not model the crystal having an experience and encoding that experience into the trace.

Experimental Note

State matching. ACCESSIBLE, REMOTE and ERASED are built from the same checkpoint object: identical occupied set, birth-time map, step, stream seed and probe. The arms differ only in which already-occupied cells carry the hidden scalar. No occupied cell is moved, added or removed. Equal initial material mass is asserted at construction.

Material dynamics.

gain g_m                0.30, added to the attachment score
half-life               6 updates
history age at t0       3 updates
carriers                2
initial mass            ≈ 1.414 (each carrier ≈ 0.707)

no inheritance, no spreading, no transfer
material dies with its carrier

Arms. ACCESSIBLE places carriers near the probe, always including x’s sole occupied neighbour. REMOTE places equal mass beyond HORIZON + margin, matched to the accessible carriers on exact adjacent-frontier count and on baseline adjacent-frontier probability mass within a frozen 0.05 tolerance. ERASED carries none. Remote matching is performed outcome-blind, before any result is examined. Evaluation is true unbounded in all arms, removing the finite-selector routing of the previous chapter.

Calibration remains active. V2 does not remove global construction matching. A dedicated ERASED PREVENT trajectory defines the target expected attachments at each lag. ACCESSIBLE and REMOTE each solve one additive offset on their own current PREVENT state so that expected total attachments match that target; the same solved offset is applied to FORCE, which is never separately normalized. The offset acts on the attachment score, before the logistic.

Intervention.

lag 1     FORCE contains x for one full causal growth exposure
          PREVENT removes x from the frontier before evaluation

after lag-1 growth and loss
          x removed from BOTH branches, absence asserted

lag 2+    ordinary dynamics; natural reoccupation permitted
          x excluded from all measured outcomes

Estimators. Local support L is 1 ≤ d(x,y) ≤ 12, excluding x.

Δ_t     = Σ_{y∈L} [ p_FORCE(y,t) − p_PREVENT(y,t) ]
G_RB    = Σ_{t=1..12} Δ_t

realized = cumulative FORCE-minus-PREVENT actual attachment counts
           over the same support and horizon

Probabilities are evaluated on each branch’s own realized state at each lag. The trajectories remain fully stochastic; expectation replaces only the final Bernoulli measurement.

Estimands. Primary: G_RB(accessible) − G_RB(remote). Secondary: the realized version, and REMOTE − ERASED as a control.

Scale and precision.

groups                  192
supported probes        564
group coverage          1.00  (required ≥ 0.90)
record-level match      1.00  (required ≥ 0.95, within 2%)
every arm × lag mean    within 2%
intervention assertions PASS
remote matching         PASS

SEI                     ±0.15 attachments
MDE80                   one-sided minimum detectable effect at 80% power;
                        the frozen rule requires MDE80 ≤ 0.15 for SUPPORTED
statistical unit        group; bootstrap percentile intervals

Analysis-only diagnostics. The V1 construct-validity audit and the V2 trace-decay closeout are analyses of existing samples. Neither is a confirmatory experiment, and neither changes the frozen V2 primary status.

Evidence Ledger

Where Does This State-Conditioned Process End?

We have now established something that earlier chapters had not.

With visible geometry matched, experimentally written hidden material state can change the response to the same perturbation.

And in the corrected V2 sample, the cumulative causal difference continues changing after that material trace has substantially weakened.

That is not yet memory.

It is a demonstrated hidden-state causal response plus a descriptive trajectory result. To call it endogenous history dependence, the crystal would still have to generate the relevant hidden state from its own past.

No retrieval or history-specific decoding has been demonstrated.

Which sharpens a question the book has left open twice.

The Is There Actually One Thing Here? chapter looked for a privileged boundary around the connected crystal and failed to find one, twice: no scale showed excess predictive coherence beyond a family null, and the candidate outer boundary localized causal effects no better than a circle drawn arbitrarily through the interior. What survived was spatial causal locality — consequences stay near their causes — which is true of any local field and establishes nothing about individuals.

But Is There Actually One Thing Here? tested regions defined primarily by geometry.

We now have a better experimental target.

Instead of asking whether a region is special because its outline looks object-like, ask whether perturbations originating inside it are preferentially contained there — and whether perturbations originating outside penetrate it less — beyond what geometry alone predicts.

That reopens the boundary question without assuming in advance that there is an object, a process or an individual waiting to be found.

Is there a region whose causal containment exceeds what its geometry alone would predict?