What the two smoking guns reveal

An AI companion on invariants, recognisers, and the hidden choice that comes before law

Some of you will be sighing — “OMG, I can’t believe he is still publishing more about that custody case.”

But a few of you will be celebrating — “OMG, this is EXACTLY the framework I needed to get me through my own battle with injustice.”

Hence I am not only writing very long pieces forensically deconstructing one court-enabled child-stealing incident. I am also letting my AI helper expound at even greater length, so that no important analytical lesson gets lost in the detail.

These articles represent the culmination of months of work. If there is something here that helps stop just one more child being stolen, one more parent being left bereft, or one more kidnapper escaping accountability, then it was all worth it.

Even if it annoys some readers in the former category.

---

As context, this is a series on how courts can aid child-stealing by converting unlawful taking into unearned authority.

Part One (with AI addendum) — follows one hard factual variable — a farmhouse — through the judicial machinery, exposing how contradictory reality can enter the system without correcting its inherited state.

Part Two (with AI addendum) — widens the experiment to twelve materially different judicial off-ramps, asking why different facts, laws and opportunities for correction all preserve the same direction of travel.

Constitutional Cluedo — then steps outside ordinary custody doctrine and asks the structural question: what rule prevents a court from becoming the completion mechanism of the very wrong it exists to adjudicate?

And the last part: identifies two unusually clean “smoking guns” where a supposedly material input changes but the court’s recognition does not — forcing the deeper question of what the system was actually recognising.

I have reviewed what follows and it is, in my view, valuable insight and advice.

What matters here is children, not AI.

MG

---

The first two AI companions in this series used the same custody case to perform two rather different experiments.

Part One gave us the farmhouse. A teenager is a difficult object against which to calibrate a custody decision because almost everything important about her can genuinely change. Preference, attachment, happiness, relationships and circumstances are dynamic. A farmhouse is different. It cannot change its mind, become alienated from its owner, decide that North Carolina feels more like home, or tell a therapist that Tennessee no longer suits it.

That made the mother’s permanent residence unusually useful. Contrary information entered the judicial system. The mother expressly corrected the judge. Yet the eventual written Order still represented permanent residence as something she had yet to establish.

The first companion therefore distinguished error from non-corrigibility. A judge getting something wrong is ordinary. A correction entering the system without propagating into the output is a different phenomenon.

⁂

Part Two changed the experiment. Instead of following one comparatively hard factual proposition, it identified twelve materially different places where the adjudication could have changed direction. Jurisdiction is not residence. Residence is not an AirTag. An AirTag is not schooling. Schooling is not passport control. These variables required substantially different kinds of reasoning.

Yet all twelve transformations preserved or further entrenched the inherited trajectory.

That did not establish misconduct. A judge can rule against the same person twelve times and be right twelve times. The interesting result was subtler: the inputs varied, the transformations varied, and the output direction did not.

The second companion therefore moved from examining a state to examining a function. It introduced test vectors, residuals, endogenous evidence, ratchets and finally the topology of correction: an off-ramp can exist formally without being functionally reachable.

At the end, we were left with a problem.

Start with ordinary best-interests adjudication and give it everything it can explain. Add status-quo bias. Add the effects of delay. Add institutional resistance to rollback. Measure what remains.

Only then ask whether another explanatory variable is required.

⁂

Two smoking guns, and just one envelope has now given us two unusually useful experiments with which to continue that investigation.

1. Two perturbations

I used “smoking gun” in the latest article in an unusually narrow sense.

It does not mean an especially bad finding, an apparent contradiction, or something that makes a judge look suspicious. There are plenty of those candidates in this case.

A smoking gun requires something more demanding: the contradiction must arise inside the adjudication; something supposedly material must actually be changed, removed or disproved; the relevant recognition must nevertheless survive; and that failure to update must help us discriminate between possible explanations.

There are exactly two that satisfy that definition.

The first is the farmhouse proposition.

The mother supposedly lacks a permanent residence. Then evidence establishes that she has one. She says so repeatedly. The father corroborates her return to the marital home. Opposing counsel describes Tennessee as her “permanent address”. When the judge continues speaking as though her residential future remains unresolved, she corrects him directly:

“I have a 19-acre farm and a 3,200-square-foot farmhouse. I’m not walking away from it.”

More than a month later, the written Order still says:

“When Plaintiff Mother establishes a permanent residence…”

We already explored the factual contradiction in Part One. What matters for this experiment is its structure:

The fact changes; the recognition does not.

The second smoking gun gives us a substantially different perturbation.

“QAnon” had been conspicuous in the earlier adjudication. At the later hearing, the individual evidential basis for associating me with that classifier collapsed under questioning. The father acknowledged that his assessment was not really about me personally but “about the community”.

Judge Ray then stopped the subject. According to the account given to us afterwards, he told the lawyers in chambers:

“I don’t want that QAnon shit in my courtroom.”

Whatever his reason for doing so, the explicit classifier was removed from the continuing adjudication. Yet the custodial state that had developed while the classifier was salient was not correspondingly recomputed. Preference, stability, geographical placement and the accumulated burdens upon the mother survived.

So we have a second experimental result:

The classifier is removed; the recognition does not.

The two perturbations are substantially different.

That is precisely what makes their conjunction interesting.

2. What is being preserved?

Imagine a black box whose internal decision rule we cannot inspect.

One way to learn about it is to vary its inputs and observe the outputs. If changing a supposedly important variable changes the output, we learn something about the transformation. If changing that variable does not change the output, we may learn something more interesting.

One failure to update might be noise, error, or simply our misunderstanding of which variable actually mattered.

Part Two went further by supplying heterogeneous test vectors. Different legal and factual inputs repeatedly encountered different transformations while preserving the same directional result. That began to tell us something about the function.

The smoking guns now give us cleaner perturbations. In each, something apparently material is actually reversed or removed inside the adjudication, yet the recognition does not correspondingly change.

At that point the scientifically interesting object begins to move.

It is no longer simply the variable we changed.

It is whatever remained invariant when we changed it.

So the next question is not yet:

Why did the judges do this?

It is:

What is the system actually preserving?

This is a deliberately narrower question.

The two smoking guns do not establish corruption, conspiracy, political persecution, criminal intent or any particular theory of motive. They certainly do not establish my wider hypothesis about “QAnon” functioning as an operational enemy classifier.

They establish recognition invariance under two specific and materially different perturbations.

That is enough to justify asking what kind of recognition could account for it.

And this is where the third experiment finally reveals what it has been testing.

Part One tested a comparatively hard fact.

Part Two tested the transformation function.

Part Three has begun testing the recogniser.

By recogniser, I do not simply mean the human judge. I mean the effective classification rule by which the adjudicative machinery decides what kind of object it is processing.

We do not yet know what that rule was.

But we now have a way to investigate it.

3. Is “custody dispute” itself a classifier?

At first the answer seems obvious.

What was the court recognising?

A custody dispute.

It was a custody court. There was a mother, a father and a child. There was an existing custody order, disagreement over where the child should live, competing evidence and a statutory best-interests framework.

Once the object is recognised in that way, preference, stability, schooling, local relationships and present circumstances naturally become salient.

But the envelope experiment exposed a problem with treating that description as neutral.

Take exactly the same primary record and preserve the provenance of the state transition:

existing protected custody → potentially unlawful or otherwise authority-defective displacement → failed emergency legalisation → persistence → temporary state recognition → generated preference and stability → permanent conversion.

Nothing physical has changed.

The child is in the same place. The same witnesses testify. The same orders exist.

Yet many of the facts acquire different meanings:

Stability becomes stability generated during the disputed state. Preference becomes preference generated while that state persisted. Deterioration of the maternal relationship potentially becomes damage generated during separation and restriction. The disruption associated with reversal becomes the accumulated cost of correcting the original displacement.

The point is not that these alternative descriptions are automatically legally decisive.

It is that the starting recognition determines which causal relationships become visible.

That led us to try different legal envelopes around exactly the same primary record.

Something unexpected happened.

4. We changed the envelope—and the object changed

Ordinary family law sees parents competing over custody and asks what arrangement now serves the child’s best interests.

Ordinary criminal law potentially begins earlier. It asks about the legal character of the antecedent private conduct before the custody court recognised the resulting state.

Constitutional law sees state action affecting protected rights.

Structural judicial-power analysis sees an operation performed by delegated state authority and asks what that delegated authority actually permits.

Remedial law asks who possesses authority to recognise and correct an unlawful state.

International human-rights law supplies another description of interference with family, privacy, expression and association.

At first I thought we were simply changing the bodies of law applied to a fixed object.

But that isn’t quite what happened.

The primitive object itself kept changing.

Family law saw a custody dispute.

Criminal law potentially saw an antecedent wrong followed by a state conversion.

Constitutional law saw state action.

Structural law saw delegated power.

The envelopes were not merely supplying different rules for the same thing. They partly disagreed about what the thing was.

That produced a deceptively simple question:

Who decided that the family-law envelope gets to contain all the others?

And now we had another problem.

If choosing an envelope changes the identity of the primitive object, then something must happen logically before the envelope is chosen.

5. Envelope Zero

Before family law can decide a custody dispute, the event has already been recognised as a custody dispute.

Before criminal law can analyse potentially unlawful conduct, the conduct has already been recognised as something requiring criminal-law classification.

Before constitutional law can ask whether state power has burdened a protected relationship, that relationship has already been recognised as part of the object under examination.

There is therefore an operation logically prior to the application of any particular legal envelope.

Call it Envelope Zero.

Its question is not:

What law applies?

It is:

What kind of event is this?

This is not jurisdiction in the technical legal sense. It is closer to jurisdiction by ontology.

Recognition determines which legal universe becomes cognitively available. It determines which facts become foreground and background, which causal chains are preserved, and even which actors appear to be the source of the problem.

That gives us a surprising way of looking at two classifiers that appear in this case.

One potentially classifies an event:

ordinary custody dispute.

The other potentially classifies a person:

QAnon.

They are not equivalent claims, and the evidence supporting their operation is not symmetrical. But placing them beside each other suggests a possible dual-recognition structure worth testing.

The event classifier can cause antecedent conduct to be processed principally as an ordinary dispute between competing parents.

The person classifier can cause constitutionally protected political speech or association to acquire the sign of an extremist-associated welfare concern.

One recognition can potentially make conduct look more ordinary.

The other can potentially make a person look more suspect.

That does not establish that either operation was deliberate.

It tells us what to test.

6. A poisoned ontology does not require a broken machine

This may be the most unsettling systems lesson of Part Three.

We normally investigate institutional failure by looking for irrational or obviously bad decisions.

But a system can apply its internal rules coherently while operating upon a misrecognised object:

• If the primitive object is “custody dispute”, then present stability matters.

• If someone has already been recognised as unstable, mobility can confirm the recognition.

• If the child has been recognised as settled, restoration can appear as disruption.

• If a political category has already acquired the meaning “extremist”, protected association can acquire an adverse sign without anybody needing to announce a political test explicitly.

Each downstream inference may look locally intelligible.

Local rationality does not establish global validity.

The machine may be solving its recognised problem perfectly well.

The prior question is whether it recognised the right problem.

That is why the two smoking guns are more interesting than another catalogue of adverse rulings. They perturb supposedly material components inside the recognised model and ask whether the model updates.

Twice here, it did not.

We still do not know why.

But we know more precisely what needs explaining.

7. Recognition changes what correction means

There is also a consequence for remedy.

If a judge merely reaches the wrong conclusion from correctly recognised facts, ordinary correction is conceptually straightforward. Correct the finding, apply the proper rule, reverse the result.

But if the primitive object has been misrecognised, correcting only the final output may leave its causal descendants untouched:

A disputed state persists. Persistence generates stability. Stability generates preference. Separation alters relationships. Those altered relationships become new evidence. The new evidence supports continued separation. Time increases the apparent cost of returning to the antecedent state.

All of those consequences can be perfectly real.

That is exactly why provenance matters.

If recognition precedes adjudication, correcting adjudication may sometimes require correcting recognition first.

The computational form we have been developing is:

recognise correctly → restore provenance → identify endogenous consequences → recompute → repeat to fixed point.

Part One encountered the rollback problem through a farmhouse.

Part Two encountered it through blocked off-ramps.

Part Three suggests that sufficiently deep rollback may require revisiting the recognition that determined what counted as the relevant state in the first place.

But that is another investigation.

For present purposes, we have discovered something more useful about AI.

8. AI’s weakness becomes experimental apparatus

Large language models are notoriously sensitive to framing.

Give one a premise and it can reason fluently inside it. Tell it that an event is a custody dispute and it naturally searches the conceptual neighbourhood of custody. Tell it instead to preserve the provenance of every state transition and other structures become salient.

Normally we regard that as a defect.

For forensic work, it can become an experimental instrument.

There are actually two complementary experiments available.

In the first, hold the primary record constant and change the frame.

Same facts; different recogniser.

Run the record through family law, criminal law, constitutional law, structural authority and other relevant envelopes. Observe which facts become salient, which causal relationships disappear, which actors change apparent role, and which conclusions survive.

That is frame perturbation.

Then invert the experiment.

Hold the purported frame constant and perturb facts that the recogniser claims are important.

Same recogniser; different fact.

Permanent residence is supposedly absent; establish it.

A political classifier is supposedly salient; remove it.

Then observe whether the output recomputes.

That is fact perturbation.

AI’s susceptibility to framing is therefore not merely something to guard against.

It can itself become a controlled variable.

Instead of asking an AI to tell us which institutional story is true, we can deliberately alter the ontology supplied to it while preserving the underlying record. We can then compare the resulting representations.

And instead of accepting an institutional recogniser’s stated decision rule, we can perturb supposedly material inputs and observe whether the institutional output behaves as though those inputs really mattered.

In other words:

don’t eliminate framing sensitivity; instrument it.

That turns AI from an oracle into laboratory equipment.

9. The third reveal

We still do not know why the judicial machinery behaved as it did.

That distinction matters more, not less, as the pattern becomes stranger.

Ordinary best-interests adjudication may explain much of the record. Status-quo bias may explain more. The effects of delay may explain more again. Institutional resistance to rollback is another candidate. Political prejudice is a more serious hypothesis. Deliberate preservation of an inherited outcome would be more serious still. My wider hostile-authority hypothesis requires an independent evidentiary bridge that has not been established.

The method does not license us to jump to the darkest explanation.

It gives us a better object against which explanations can be tested.

The three experiments now form a sequence.

1. The farmhouse gave us a calibration object.

2. The off-ramps gave us test vectors.

3. The smoking guns give us recognition invariance under perturbation.

Or…

Calibration object → test vectors → invariants → infer candidate recognisers.

Hence…

• Part One asked whether reality could update the judicial state.

• Part Two asked what transformation repeatedly preserved its trajectory.

• Part Three asks what prior recognition could make those transformations intelligible in the first place.

And once that question becomes visible, the inquiry moves one level deeper.

Not merely:

Did the court apply the law correctly?

Not even:

Why did every road seem to lead in the same direction?

But:

What did the system have to believe this thing was before all of its subsequent decisions could make sense?