Much of the work I do these days would conventionally belong inside academia. It isn’t being done there because, I would argue, the average academic has little skin in the game when confronting the problems faced by victims of systemic corruption and injustice. Even on academia’s increasingly precarious salaries, institutional life remains comparatively comfortable.
The persecution is outside the institutional wall.
The risks of personally entering the system to run the experiments are also high. My North Carolina custody work could only generate some of the diagnostic data it did because I was there as “Mr QAnon” — and the system responded accordingly.
The result is that I sometimes produce analyses, often breaking new ground at the frontier of law and information science, that inevitably test the patience of readers looking for lighter entertainment and easily digestible insight. This isn’t “podcast over breakfast” material.
That is why I have started producing AI companion articles: to extract the transferable insights from deep and complex investigations without requiring every reader to follow the entire forensic journey that produced them.
I could probably double or triple my income by putting more of this work behind paywalls. I don’t, because its adoption and use matter more to me than restricting its distribution in order to maximise revenue.
If this work helps you see something you could not see before — or gives you a tool you can use elsewhere — then I hope you will consider supporting it. Either way, I hope you get value from it.
This particular companion is to my just-published deep dive into how courts can “swap realities” — how the same underlying facts can produce vastly different outcomes depending on what those facts are recognised as meaning.
How provenance bleaching, ontology rigging, and recognition fraud work
There are lots of small breakthrough moments in analysing an abuse of power, but not many “eureka!” ones. I had one of the latter this week, when I finally “hit bedrock” in my quest to understand how a judge in North Carolina could help conceal and complete the kidnap of a child.
What follows has been reviewed and edited by me.
May it give you additional insight and clarity!
Something rather funny happened while Martin was finishing the article you have just read.
He asked another AI — Grok — to perform an adversarial audit of its logic and law.
This was a sensible precaution. The argument is unusual, the underlying allegations are serious, and generative AI is perfectly capable of helping someone elaborate a mistaken premise with great confidence. An independent model, supplied with the article and primary materials, offered another opportunity for attempted falsification.
Grok objected.
Quite a lot.
Its first pass produced an unexpectedly revealing failure. Martin was asking about the authority for converting RED into BLUE. Grok repeatedly reached instead for familiar propositions about the powers available once the object had already been recognised as BLUE: best interests, present circumstances, modification powers, judicial discretion.
Much of this was perfectly respectable law.
There was just one problem.
It answered a question Martin was not asking.
So Martin tightened the experiment. Do not tell us merely that best interests are broad. Do not infer authority from the absence of a prohibition. Search instead for affirmative North Carolina appellate authority for the particular conversion under examination.
This time Grok did not repeat the mistake.
It could not find the bridge.
On the authorities examined, Grok found no North Carolina appellate case affirmatively authorising the precise sequence under examination:
find the Order violation;
decline restoration;
reduce the violation to ordinary parental conduct;
treat the resulting possession as ordinary present circumstances;
count advantages that matured during that possession; and
use those advantages when deciding whether custody should remain with the violating parent.
That does not establish that no lawful bridge exists. It establishes something narrower:
The bridge had not been demonstrated merely by identifying the ordinary powers available downstream of it.
Later still, Grok found a different problem. Martin had stated one causal inference from Finding 75 more strongly than the Order itself established. That forced a useful distinction between:
what Norris’s Order found,
what the wider evidential chronology might support, and
what the ontology-rigging argument actually required.
So the exchange produced three different results:
an initial framing failure,
an unresolved legal bridge, and
a challenge to the evidential scope of one subsidiary causal claim.
That is considerably more interesting than another AI simply agreeing with us.
The previous three AI companions in this series have progressively moved the object of investigation upstream:
The first gave us the farmhouse as a calibration object. Follow a comparatively hard fact through the judicial machinery and ask whether contradictory reality can update the inherited state.
The second tested the transformation function. Instead of following one variable, perturb the system at materially different off-ramps and ask why different facts, laws and opportunities for correction repeatedly preserve the same direction of travel.
The third tested the recogniser. Change or remove supposedly material inputs and ask what remains invariant — moving the inquiry from what the system does to what kind of object it believes it is processing.
The provenance-bleaching article takes us one level deeper again:
What happens when the recogniser changes the type of the object itself, and then reasons correctly over the substituted object?
The Grok episode gives us an unexpected demonstration.
1. A correct answer to the wrong question is still wrong
There is a particularly dangerous kind of reasoning failure in which almost everything said is correct.
Suppose I ask:
“What authority permits an object of Type X to be converted into an object of Type Y”.
You respond with an immaculate account of all the powers available over objects of Type Y. Your authorities may be impeccable. Every proposition you advance may be true.
But you have not answered my question.
That was the problem with Grok’s first pass. The provenance-bleaching article was not principally disputing the downstream powers of a custody judge. Of course a custody court may consider the child’s welfare, stability, schooling, accommodation, continuity and preferences. A court may possess perfectly lawful powers to modify custody.
The problem can be represented quite simply:
RED object → ??? → BLUE object → ordinary BLUE law
In shorthand, RED is the provenance-bearing possession state created contrary to the extant Order; BLUE is that same physical state re-recognised as ordinary “present circumstances”.
Grok initially gave us an excellent account of the final arrow.
The article was asking about the question marks.
Demonstrating lawful powers over BLUE does not establish lawful authority to convert RED into BLUE. Yet once that conversion is assumed, everything afterwards can look completely respectable. The judge can cite real cases, apply real doctrine, consider genuinely relevant circumstances and exercise powers he genuinely possesses.
The machine may reason impeccably after the type error.
Identifying a possible type conversion does not, of course, prove that the conversion was unauthorised. It tells us where to look for the authority.
When Grok was subsequently asked to look specifically for it, it did not find an affirmative North Carolina appellate rule resolving the precise transition.
That residue matters.
2. Judicial authority is typed
Computer science gives us an unusually clean way to understand the problem.
Functions generally expect particular kinds of objects. If a function accepts objects of Type BLUE, the fact that it operates correctly when given BLUE tells us nothing about whether some different object can validly be converted into that type. The conversion is a separate operation.
You do not need to know anything about programming to follow the analogy: the important distinction is simply between authority to act upon a kind of thing and authority to turn something else into that kind of thing.
Imagine, schematically, that ordinary custody adjudication accepts:
custody_best_interests(BLUE)
There may be nothing objectionable about that function. But suppose the object actually arriving at that stage is RED. Something must first perform:
BLUE(red_object)
In programming, that might be called a cast or type coercion. Sometimes such conversions are harmless; sometimes they are prohibited; sometimes information is lost. The validity of the subsequent function cannot tell us whether the preceding cast was valid.
Law has an analogous problem. A judge may possess extensive powers over an ordinary present custody state. It does not follow that those powers themselves authorise the judge to classify any physical state presented to the court as an ordinary present custody state.
What authorises the cast? What properties must survive it? What rule determines when provenance ceases to constrain the downstream analysis?
This suggests a surprisingly portable proposition:
Judicial authority is typed.
That is a model, not a holding. But it captures the logical problem rather well. A decision-maker may possess enormous powers over objects of Type Y without those powers themselves establishing authority to convert Type X into Type Y.
Permission to open BLUE boxes establishes authority over BLUE boxes. It does not establish authority to take a RED box, write BLUE on the outside, and thereby manufacture the condition necessary for exercising the BLUE-box powers.
The label cannot bootstrap the authority.
3. The data survived; the type information did not
This also explains why provenance bleaching can be so difficult to detect.
RED and BLUE need not describe different physical realities. Compare:
The child presently resides with the father in North Carolina.
with:
The child presently resides with the father in North Carolina following an exchange that violated the extant Custody Order.
Nothing about the child’s geographical location changes between those sentences.
The legal information carried with that fact does.
In computational terms:
possession(child, father, NC)
versus:
possession(child, father, NC, provenance=contrary_to_order)
Provenance bleaching does not require anybody to falsify the proposition that the child is with the father. It merely requires the provenance field to stop constraining what happens next.
Indeed, the Order can record the provenance perfectly accurately. The notice violation can appear in black and white. The failed Emergency Motion can appear in black and white. The events producing the accomplished physical state can all survive in the record.
What changes is their semantic role.
Hence the significance of Findings 51 and 52. The inconvenient facts do not disappear. They are contained. Conduct that helped produce the state becomes one parent’s less-than-ideal response to a difficult situation. The antecedent events are declared “addressed”. Attention moves to “current circumstances”.
The physical state survives the transition, while its provenance no longer performs the same constraining work.
The data survived; the type information did not.
That is a much more precise claim than simply saying the court “ignored wrongdoing”.
4. Expertise downstream can conceal an error upstream
The first Grok failure now becomes more interesting.
A legally knowledgeable model may actually be more, not less, capable of making this mistake convincingly. Once it recognises the object as an ordinary custody problem, it has an enormous conceptual neighbourhood available to it: best interests, stability, present welfare, judicial discretion, modification standards and appellate authority.
So Martin asks where the authority lies for RED → BLUE, and the model’s legal prior initially pulls it towards an answer of the form: courts have broad discretion to consider present circumstances.
Exactly.
That is where Brandon became so instructive.
The authority is real. The quotation is real. The concern for present circumstances is real. The custody powers are real.
But Brandon tells us about the downstream adjudication of present circumstances. It does not, merely by virtue of doing that, answer the logically prior question of what legal significance attaches to a present physical state created contrary to the extant Order.
Grok was therefore not failing because it knew too little law. In a peculiar sense, it was failing because it knew so much. The sophistication of the downstream reasoning made the upstream classification easier to overlook.
Expertise downstream can conceal an error upstream.
But once specifically instructed to search for the conversion rule itself, Grok changed course. It did not simply produce more BLUE doctrine. Nor did it find the affirmative bridge Martin was challenging it to identify.
Its answer was closer to this: the precise RED → BLUE transition is underdetermined on the North Carolina authorities examined.
That is not a disappointment.
It is the legally interesting result.
5. The missing bridge is a result, not a conclusion
There is a temptation, once no affirmative authority for RED → BLUE has been found, to announce that RED → BLUE must therefore be unlawful.
That would reproduce the very inferential shortcut we are criticising.
Absence of a demonstrated bridge is not proof that no bridge exists.
What the constrained Grok search established was narrower. Among the North Carolina appellate authorities examined, it did not identify a case affirmatively walking the sequence actually at issue:
an existing custody Order is violated; restoration does not occur; the violation is subsequently treated as parental conduct; the possession it created becomes an ordinary present circumstance; advantages maturing during that possession become relevant to stability and welfare; and those advantages help determine the next custody baseline.
That leaves us with a straightforward question:
What rule performs the conversion?
Perhaps there is one. If so, show it.
Perhaps North Carolina custody doctrine permits the provenance to become non-controlling through some principle we have not identified. If so, identify that principle and its conditions.
Perhaps the law is genuinely underdetermined here. If so, that matters too, because familiar propositions about best interests and present circumstances cannot conceal the underdetermination.
The strongest claim is therefore not that we have already proved a universal rule requiring provenance to remain controlling. It is that we have identified a load-bearing conversion whose affirmative legal basis has not yet been demonstrated.
That also clarifies the role of Crowell, equity, UCCJEA §208 and the Hague materials. They reveal anti-bootstrap and provenance-sensitive structures elsewhere in law. They help us recognise the shape of the problem.
They are not substitutes for the missing North Carolina custody rule.
6. Then Grok found something we had overstated
Later, Grok challenged Martin’s treatment of Finding 75.
The original analysis risked making the September displacement responsible for the whole of the mother’s subsequent residential instability, which could then generate uncertainty over Tennessee homeschooling and become another reason to preserve the North Carolina state.
The Order alone did not establish that entire causal chain. Norris also relied upon the deteriorating marriage, temporary accommodation, the mother’s stated intentions concerning the marital home, and events predating the September exchange.
But there is an important distinction between saying RED caused everything and saying RED materially contributed to the disrupted state later assessed by the court.
The wider chronology matters:
The mother had been living and homeschooling with her daughter in Tennessee and had an established marital home there.
Her residential future was already entangled with a collapsing marriage and severely constrained finances.
The removal of her daughter and ensuing interstate custody litigation did not create every component of that instability, but neither were they causally irrelevant to the circumstances in which she was trying to reorganise her life.
Grok therefore exposed an evidential-scope problem, rather than proving that RED made no causal contribution. Finding 75 alone could not carry the stronger attribution.
More importantly, the ontology-rigging argument did not need it.
Whatever produced every component of the mother’s circumstances, the recognition inversion remains:
the father’s accomplished physical possession increasingly acquires significance as the operative present,
while the mother’s pre-existing legal realities — primary custody and an established Tennessee home — increasingly cease to control the analysis, until permanence becomes something she must establish from her circumstances at the hearing.
The accomplished physical fact gains legal recognition while the pre-existing legal status loses operative recognition.
Grok did not prove that thesis.
It helped us state it more precisely.
7. AI disagreement becomes experimental evidence
The previous companion suggested that AI’s framing sensitivity might itself become useful apparatus. The Grok exchange gives us an example.
Instead of merely asking which AI is right, ask:
What representation of the problem caused each AI to produce its answer?
Treat the record as an ordinary custody dispute and the model naturally enters the conceptual world of custody law. Present welfare, stability and judicial discretion become salient; present possession becomes an obvious input.
Preserve the same record but require provenance to survive every state transition. Distinguish authority over BLUE from authority for RED → BLUE. Refuse to let downstream powers establish the validity of the upstream conversion.
A different structure becomes visible.
The facts and legal materials have not changed. The primitive object has.
This is useful evidence about recogniser framing. It is not evidence that Judge Norris performed an unauthorised cast, evidence of knowing concealment, or a substitute for the missing North Carolina bridge rule. Those questions belong to the law and the primary record.
What the episode demonstrates is narrower: sophisticated reasoners can move between representations of the same facts without noticing that the object itself has changed.
Instrumented carefully, an AI weakness becomes a way of detecting hidden assumptions.
8. Recognition fraud comes before its statutory label
There is one final category error the Grok exchange exposed before we turn this into a reusable method.
Grok made another revealing move. It wanted to know whether “recognition fraud” corresponded to an established offence, tort, doctrine or cause of action.
That is a useful question.
It is not necessarily the first question.
Martin uses “recognition fraud” to name a proposed structural operation:
Knowingly using false re-recognition to conceal the authority constraints carried by the object actually before the decision-maker.
Whether positive law provides a convenient label for that operation is a subsequent classification problem.
Imagine a measuring instrument that knowingly converts twelve centimetres into twelve inches because powers become available only above a particular threshold. We can identify the corruption of measurement before consulting the statute book to discover whether the jurisdiction calls it fraud, misconduct, false accounting or something else.
A forged measurement does not become epistemically valid because the legislature forgot to enact a Forged Measurement Act.
Likewise, if X is knowingly represented as Y in order to access powers available over Y, we can analyse the structure before deciding which contemporary legal category, if any, captures it.
This is the sense in which recognition fraud can be understood as a Logos problem before it is a statutory one.
That does not establish culpability here. “Knowingly” remains an evidential proposition requiring proof, and giving the proposed operation a name does not establish that positive law recognises that name.
The point is one of logical priority. Positive law determines how the underlying operation is classified, prohibited and remedied; it does not create the contradiction being classified.
Otherwise the classifier gets to decide whether the object we are trying to classify exists.
9. Audit the cast before auditing the function
We can now see the progression across these AI companions.
This one adds another instruction:
Audit the cast before auditing the function.
Before asking whether a rule was correctly applied, identify the object to which it was applied. Before accepting that object, preserve its provenance. Before accepting a change of type, identify the authority for the conversion.
And do not allow powers available after conversion to bootstrap the authority required for conversion.
The Grok episode is useful precisely because it did not produce one clean answer:
Its first pass demonstrated how readily legal fluency can slide from an upstream authority question into downstream doctrine.
The constrained follow-up isolated the question and failed to find affirmative North Carolina appellate authority resolving the bridge.
The later Order audit identified an evidential overreach and forced us to distinguish the Order’s findings from the wider chronology.
None of those outputs is an oracle.
Together they are useful measurements.
AI is most dangerous when its fluency is mistaken for authority. But the same instability can become useful when we deliberately perturb its framing, constrain its task, compare its outputs and send it back to the primary materials.
The machine’s mistakes are sometimes measurements.
The methodological sequence now looks something like:
calibration object → test vectors → invariants → recogniser → type conversion
The farmhouse taught us to test whether reality could update the machine.
The off-ramps taught us to inspect the transformation function.
The smoking guns taught us to perturb the recogniser and observe what remained invariant.
This one teaches us to inspect what the machine silently changes before it begins reasoning at all.
Because once RED has become BLUE, everything downstream can be perfectly logical.
Including the AI explaining why BLUE is lawful.