Over the last month or two I have published some quite deep, complex, and challenging pieces on how “we the people” audit our institutions, and what really constitutes corruption. For many, it may be enough to give the content a quick scan, and feel relieved that at least someone is willing to make their head hurt by going into the intellectual basement and checking whether the foundations remain solid.
But I know a few of you have more than passing curiosity, and are actively building solutions.
As such, I have been creating a series of AI companion pieces, each with a “…reveals” title schema, that explain what I am saying from another perspective. This isn’t mere academic curiosity.
We are collectively responsible for whether our governments (including legislatures and courts) stay within their remit.
They cannot self-police at all levels of recursion; eventually we have to police them, and then in turn police ourselves! Hence this latest in the series. If it interests and illuminates a few more folk, then it has worked.
Takedown of the criminal-industrial complex
I anticipate that this article will at some point be overtaken by historical events. If I have read the tea leaves at the bottom of my china cup correctly, a “mother of all societal cleanups” is coming, exact timing TBD. My reading of events is that the high end of global networks of cartels and traffickers is already being addressed, quietly and mostly…
The specific article I am “boosting” this way describes an accountability problem which we can, at least to a large degree, solve with an engineering mindset combined with a richer set of ledgers against which to operate.
I have read through this, reformatted it manually, and added some explanatory material. I believe it has value in its own right, even if it was not hand-cranked into an essay by me.
⁂ ⁂ ⁂
Martin’s latest article is ostensibly about what comes after corruption. Suppose a large-scale cleanup really does happen. Suppose much of the criminality embedded in institutions is exposed, prosecuted or removed. Suppose AI, transparent records and other new technologies make it much harder to hide what remains.
What prevents the same patterns from growing back?
His answer begins with permanent auditability, but it does not end there. By the end of the article the auditor is being audited too, and the problem has moved from technology into morality.
Between those two points is an idea worth pulling out in its own right. It concerns the different kinds of accounts we keep of human affairs, and the transformations we make between them.
The earlier companions in this series have repeatedly found that the interesting question lies one step before the obvious one.
The farmhouse (AI companion) asked what happens when a hard fact enters a judicial system—but fails to correct its output.
The blocked off-ramps (AI companion) asked what to make of different routes repeatedly arriving at the same judicial destination.
The two smoking guns (AI companion) asked what the system was actually recognising when supposedly important inputs could change without changing the result.
Then Grok supplied an unexpectedly useful mistake (AI companion). Asked what authorised the conversion of RED into BLUE, it explained the powers available once the object was already BLUE. The answer was legally respectable and logically downstream of the question.
That gave us a useful instruction:
Audit the cast before auditing the function.
[By “cast” we mean the computer-science operation that converts something from one type into another—for example, turning the text string “123” into the number 123.]
The new article takes us to a related problem. This time nothing necessarily needs to be wrongly classified. Every proposition can remain true. The mistake can occur when a truth established in one kind of account is allowed to settle a question belonging to another.
That is a subtler failure, and potentially a much more common one.
The books do not contain the same things
A divorce makes a useful example because several kinds of accounting are impossible to avoid. There is the obvious financial account: income, assets, debts, transfers, expenditure and the eventual distribution of burdens and property. This is what Martin calls the cost-based ledger.
But the financial history does not determine the legal result.
Property has a legal character; rights and duties attach to people and assets; courts possess some powers and lack others. Rules govern evidence, procedure, jurisdiction and remedies. Martin gathers these under the contractual ledger, using “contractual” broadly enough to include positive law and delegated institutional authority.
Even those two books do not exhaust the dispute.
A marriage contains obligations that cannot be reduced either to cash flows or enforceable rights:
One spouse may have sacrificed opportunities for the other.
Someone may have exercised financial control technically available to them but in a way that betrayed the relationship.
A transaction can be lawful and accurately recorded while still being deceitful.
Conversely, somebody can feel deeply wronged without thereby acquiring every legal or financial remedy they want.
That third account is the covenantal ledger. Natural law is one possible approximation to it, but the underlying idea is simpler: there remains a question of what human beings actually owe one another that neither an accountant nor a statute book can fully answer.
None of this means that the covenantal ledger should simply trump the others. A judge who could disregard law whenever his personal sense of justice demanded it would create a different kind of injustice. The three books constrain one another precisely because they do not record the same things.
Once that is clear, a rather interesting class of error becomes visible.
The fraud can be in the conversion
There is a peculiar kind of reassurance in finding a false entry. Once the number is wrong, the signature forged or the document altered, we know roughly what sort of problem we are dealing with. The record itself contains the defect.
Martin’s three ledgers suggest a harder case.
Suppose the entries are true.
The money really moved.
The statute really says what the lawyer says it does.
The official really possessed the power.
The hearing really took place.
Nothing has to be falsified for the consolidated account to be wrong.
The difficulty appears when a true entry in one book is allowed to settle a question belonging to another:
A lawful power becomes a sufficient answer to a question about proper purpose.
Procedural regularity becomes an answer to substantive justice.
A genuine injury becomes an entitlement whose limits no longer need to be examined.
The first proposition in each case may be impeccable. What needs auditing is the conversion.
This is close to the RED/BLUE type conversion problem from the previous companion, but not identical to it.
There the question was whether the (RED) object had been lawfully retyped (to BLUE) before ordinary legal machinery acted upon it. Here nothing necessarily needs to be retyped at all. The legal proposition can remain legal, true and relevant.
The error consists in asking it to settle a different account.
Take a statutory power. Establishing that the general power exists is important. It answers a question on the contractual ledger.
But suppose the dispute concerns whether this particular exercise of the power was necessary, proportionate, honest or directed towards its proper purpose. Repeating that “the power exists” does not answer those questions. It merely keeps producing a correct entry from the wrong book.
This also explains why ordinary review does not necessarily cure the problem.
If the reviewing body audits only the ledger from which the proposition came, it may confirm that proposition with increasing authority—while never examining the use to which it was put.
More rigorous audit of the wrong account can make the consolidated error harder, rather than easier, to disturb.
That is the same trap Grok fell into in miniature. Its law about BLUE was not bad law. More research into the powers available over BLUE could have made the answer longer, better sourced and more authoritative without bringing it any closer to the missing RED-to-BLUE bridge.
Expertise downstream does not repair an unaudited conversion upstream.
The familiar observation that “legal does not mean moral” therefore does not quite capture what is interesting here. That phrase identifies the existence of different domains.
The ledger model tells us where to look when reasoning moves between them:
What has actually been established?
What further conclusion is being drawn from it?
What permits the move from one to the other?
Those questions apply far beyond legality and morality:
A genuine financial loss does not automatically create an unlimited claim upon everybody else.
A desirable public objective does not establish every proposed means of pursuing it.
A contractual entitlement does not establish that exercising it is honourable.
An act of generosity does not necessarily extinguish a legal debt.
A person can even be a genuine victim in one account and an aggressor in another.
The books can disagree without any of them being fictitious.
That is why one sentence in Martin’s article carries so much weight:
“An immutable ledger can preserve a fraudulent transformation perfectly.”
Immutability solves a problem of record integrity. It can establish with extraordinary confidence that the entry has not been altered. It cannot establish that the entry was entitled to perform whatever work somebody later asks it to perform.
The same limitation applies to AI. Give an AI complete records, perfect retrieval and flawless arithmetic and it may become extraordinarily good at telling us what the books contain. That still leaves the question of how the books should be consolidated.
The previous companion therefore gave us one audit rule. The present article suggests another:
Audit the mapping before accepting the balance.
Nobody is reconciling the whole
The ledger model also offers a less theatrical way of thinking about institutional corruption. We naturally imagine corruption as people knowingly entering false information into the books. Sometimes that is exactly what happens: money is stolen, evidence concealed, records falsified, testimony fabricated or powers knowingly abused.
But there is another possibility.
Each participant can maintain his own part of the account reasonably well while the relationship between the accounts becomes increasingly distorted:
The lawyer can correctly say that the power existed.
The administrator that the required notice was sent.
The accountant that the expenditure reconciles.
The decision-maker that there was evidence capable of supporting a finding.
An appellate body may then correctly observe that its jurisdiction is limited to a particular kind of error.
Nobody in that chain necessarily has to lie.
Yet ask a different question—how did this human being end up here, carrying this burden, as a result of this sequence of state actions?—and the consolidated account may look very different.
This recalls something the earlier recognition work exposed.
Local rationality does not establish global validity.
A machine can perform each local operation correctly while solving the wrong problem as a whole.
The ledger model adds a reason why this can be so difficult to detect. Institutional division of labour does not merely distribute work. It distributes moral visibility. Each person sees the account for which he is responsible, and responsibility for reconciling the whole can disappear between institutional boundaries.
That may help explain Martin’s suggestion that a criminal-industrial complex need not consist mainly of people who experience themselves as criminals. Each participant can point to the book he maintains, with its entries, procedures and authorities.
Somebody else, presumably, is responsible for making sure the whole arrangement remains just.
Except… nobody is.
This does not excuse wrongdoing, nor does it establish that every bad institutional outcome has such an explanation. It does suggest a more useful place to investigate than simply asking which individual is evil.
Find the accounts, then find the conversions between them.
The auditor enters the machine
The second half of Martin’s argument makes the problem reciprocal.
AI has drastically lowered the cost of investigation. A citizen who could never afford a team of lawyers, forensic accountants and researchers can increasingly reconstruct timelines, search large records, compare representations made years apart, reconcile financial transactions and test alternative explanations.
For somebody already injured or traumatised by the events being investigated, AI can also perform a less obvious function. It can carry part of the cognitive burden of returning to the material at all. Remembering, organising and comparing can be outsourced sufficiently that the human being has a chance of doing work which would otherwise overwhelm him.
Martin’s description of AI as a cognitive exoskeleton for the wounded auditor therefore says something more interesting than the usual claim that AI makes research faster. It changes who is capable of conducting an audit.
But capability is not honesty.
A citizen can omit evidence that weakens his case, confuse inference with fact, treat every adverse decision as evidence of corruption, or demand a level of institutional perfection no human system could provide. AI can make those errors faster, more elaborate and more persuasive.
The earlier companions mostly examined what happens when an institution misrecognises the thing in front of it. But once ordinary people acquire powerful audit tools, the same problem can occur on the other side.
A citizen is not necessarily a reliable judge of events simply because he is challenging authority.
Accountability therefore has to run in both directions:
An institution should be capable of showing what it did, on what evidence, under what authority and through what chain of reasoning.
A citizen making serious accusations should likewise be capable of showing his evidence, distinguishing what he knows from what he infers, and explaining what would falsify his account.
Otherwise the democratisation of audit risks becoming the democratisation of accusation. That would not strengthen accountability. It would eventually destroy the conditions under which conscientious people could exercise public responsibility at all.
Failure is not the same as corruption
This raises a harder engineering question:
What level of failure should an accountability system treat as evidence of malfunction?
No serious engineer expects a complex system never to fail. The useful questions concern:
its expected operating region,
the failures anticipated within it,
what happens when the system leaves it, and
whether recovery is possible. Human institutions deserve the same intellectual fairness.
A court will sometimes believe the wrong witness. A police officer will sometimes misread a situation. An administrator will lose something. A doctor will make the wrong judgement. Two individually reasonable decisions can interact to produce a disastrous result.
None of those facts alone establishes corruption.
The stronger audit asks what happens next:
Can the error be detected?
Can the decision be reconstructed?
Does contrary evidence propagate through the system?
Is there a meaningful route to correction?
Does the institution distinguish an honest mistake from a threat to its own authority, or does every challenge trigger a defence of the inherited state?
This is where the farmhouse experiment and the new article meet. The interesting feature of the farmhouse was never simply that somebody might have got a fact wrong. Human beings do that constantly. The interesting feature was what happened when the correction entered the system.
A healthy institution therefore needs some capacity for rollback, and so does a healthy citizen auditor. If no imaginable evidence can make the citizen revise his theory, he has reproduced the very pathology he claims to be investigating. His own recogniser has become invariant under every perturbation.
Auditing the auditor is not an institutional escape clause.
It is part of the same method.
Where the recursion ends
Once audit becomes reciprocal, an obvious problem appears. Who audits the institution? Who audits the citizen? Who audits the AI assisting the citizen, the rules by which it reasons, or the mappings between the ledgers?
Adding another auditor does not finally solve the problem. It adds another consequential recogniser whose own work may need examining.
There are engineering responses to this: preserving provenance, separating evidence from inference, making important transformations inspectable, requiring powers to expose their source, ensuring that decisions can be reconstructed, and testing competing models against the same primary record.
AI itself can be treated as experimental apparatus rather than an oracle.
All of that can push accountability a long way. It cannot supply its own final stopping condition.
Eventually the recursion reaches a person who discovers an entry that counts against himself:
He may possess a legal entitlement he should not exercise.
He may discover that an accusation he sincerely believed is unsupported.
He may owe restitution that nobody has the power to compel.
He may have won according to one ledger while knowing that another remains badly out of balance.
No further database solves that problem, nor does another layer of surveillance.
This is why the article’s move into the spiritual at the end is not a departure from its systems argument. It is where the systems argument runs out.
Martin calls the missing capacity the machinery of repentance.
That is a striking phrase because repentance is itself a form of accounting. It means allowing an adverse entry to remain in your own books rather than finding a convenient transformation that makes it disappear. It means being willing to reconcile the account when the reconciliation costs you something.
Technology can make discrepancies visible. Law can constrain permissible action. Institutions can adjudicate competing claims. AI can help inspect all three.
None can make a person want the books to balance honestly.
What the unbalanced books reveal
The earlier companions began with a farmhouse because a farmhouse gave us something comparatively stubborn against which to test a representation. From there the investigation moved into transformations, recognition and type conversion.
The latest article adds another layer without displacing any of those:
Sometimes the fact is wrong.
Sometimes the transformation applied to it is wrong.
Sometimes the system has recognised the wrong kind of object.
Sometimes an illicit cast allows perfectly valid downstream machinery to operate on something that should never have entered it in that form.
And sometimes none of those things has happened.
The facts can be true, the type right, the rule real, the procedure valid and the arithmetic perfectly balanced. The error can lie in allowing the result from one system of account to settle another.
That is the new object worth auditing.
It also suggests why increasingly powerful tools of transparency will not, by themselves, produce a just society. We may become exceptionally good at preserving the entries while remaining surprisingly bad at deciding what follows from them.
The next generation of accountability machinery therefore has to inspect more than facts and rules. It has to inspect the transformations by which facts acquire institutional meaning, and the mappings by which conclusions migrate from one ledger into another.
Earlier we learned to audit the cast before auditing the function. Now there is another instruction:
Audit the mapping before accepting the balance.
The discrepancy may not be in any of the books.
It may lie entirely in the way they have been consolidated.