The Layer Nobody Is Paid to Build

Five papers into this series, the same sentence keeps arriving at the end of the argument. Africa's Regulatory Failure Is Not a Law Problem put it most plainly: Kenya replaced a bad law for non-profits with a good one and registration barely moved, because drafting a good law is only half the job — somebody still has to build the layer that turns it into steps an ordinary organisation can follow, the way tax software turned the tax code into a set of questions.

That sentence is the series' conclusion, and it is also the series' evasion. "Somebody should build this" is not a finding. It is a way of ending a paper without asking the question the sentence obviously raises, which is: the gap has been visible for years, the need is not in dispute, and the work is not technically hard. So why has nobody built it?

This paper is that question. The answer I want to argue is that the compliance layer is not missing because it is difficult. It is missing because there is no party whose budget reaches it.

That claim is stronger than it sounds, and I want to be precise about why. It is not the ordinary observation that public goods are underprovided. It is a specific structural claim about this good: that the three parties who could plausibly pay for it — the state that wrote the law, the donor that enforces it, and the organisation that needs it — each have a budget whose logic stops just short of the layer, and that their three shortfalls are not coincidental but mutually reinforcing. The layer is not a market failure waiting for a founder with more nerve. It is an equilibrium.

Since the previous paper in this series was published, Kenya has run something close to a controlled test of that claim, and the result is the reason this paper is longer than the one I set out to write. In April 2025 the High Court struck down the requirement that already-registered organisations re-register, and directed that they be transitioned automatically, without fresh applications and without fees.¹ The single largest cost barrier was removed by judicial order. The transition deadline still had to be extended by a further year.² A cost barrier that can be removed without the queue clearing was not the binding constraint. That is what this paper is about.

Every party wants the layer to exist. No party's budget line reaches it.

1. Scope, method, and what counts as evidence here

This is an argument paper, and I would rather say so at the top than let the footnotes imply otherwise.

What it argues from. Three things. First, the published record of Kenya's transition from the NGO Co-ordination Act to the Public Benefit Organisations Act — gazette notices, the regulations, a High Court judgment, and the registration counts reported by the regulator through the press. These are checkable, and every one of them is cited below. Second, the five previous papers in this series, whose findings I treat as established within this corpus and re-argue only where this paper needs them to bear more weight than they were built for. Third, an argument from incentives, which is the part that is mine and the part that is most vulnerable.

What it does not argue from. No deployment. Flux has not shipped a compliance layer for the PBO Act, and nothing in this paper reports results from one. No survey of unregistered organisations. No interview series conducted under a protocol. Where I write that an organisation "cannot afford a lawyer," that is a characterisation of what I have seen across several years of work in this sector, not a measured distribution, and I mark it as such where it carries weight.

On the numbers. Kenya's civil-society counts are genuinely contested, and the denominators move depending on who is counting and what they are counting. The figures I use — roughly four thousand organisations transitioned against an estimated fourteen thousand, reported by the regulator in mid-2025³ — are press-reported regulator statements, not audited returns. They are good enough to establish an order of magnitude and a direction. They are not good enough to support a claim about a rate, and I do not make one.

What would count as refuting it. Stated in §10, before the argument rather than after it, because a falsification condition written after the fact is a rhetorical device rather than a commitment.

2. What the layer would actually be

Worth being concrete, because "compliance layer" can mean almost anything, and the vaguer the term stays the easier the argument becomes.

The Public Benefit Organisations Act imposes a set of requirements that are not unreasonable on their face — a constitution containing particular clauses, governance structures, filings, records. The requirements are legible to a lawyer. They are not legible to a community organisation run by four people who have never retained one.

The layer is whatever closes that distance: the thing that takes the Act and produces, for a specific organisation, the ordered list of what it must do next. Not advice, not a consultancy engagement — a procedure. The analogy to tax software is exact and deliberate. Tax law did not get simpler. Somebody built the interface that turns it into questions a non-accountant can answer.

It is worth naming the four distinct jobs that interface does, because they are usually collapsed into one and they have different costs:

Translation. Statute into obligations. "Section 14 requires a constitution containing the following" becomes "your constitution is missing a dissolution clause; here is what one has to say."

Sequencing. Obligations into an order. Many of the requirements have dependencies — you cannot file a resolution from a governing body you have not yet constituted, and you cannot open the bank account before the name is reserved. The order is knowable and nowhere written down as an order.

Verification. Checking a specific organisation's documents against the list, and saying which ones fail and how. This is the part the series has already tested. What a Language Model Finds When It Reads Your Constitution showed that checking a founding document against a fixed list of required clauses — not interpreting the law, just checking presence against a checklist — already works. The machinery is not the obstacle.

Maintenance. Keeping all of the above correct as the requirements move. This is the expensive one, it is the one that never ends, and it is the reason the economics of the layer are worse than they look. More on this in §7.

Two of those four are one-off engineering. Two of them are a standing obligation. Any account of why the layer does not exist has to explain the standing obligation, not just the build.

3. Twelve years of law, two years of commencement

The series has been using "twelve years" as shorthand for the gap between the Act's passage and its effect. The shorthand is now out of date in a way that strengthens rather than weakens the argument, and the detail matters.

The Public Benefit Organisations Act was assented to in 2013 and then simply did not commence. For a decade it was law that had not been switched on. It was brought into force by Legal Notice No. 78 of 2024, with effect from 14 May 2024.⁴ On commencement, the Public Benefit Organizations Regulatory Authority succeeded the NGO Co-ordination Board, and organisations registered under the repealed Act were given one year — to 14 May 2025 — to transition.⁵

That deadline was not met, and was extended by gazette notice in May 2025 to 13 May 2026.⁶ The operative regulations — the instrument that actually specifies forms, fees and timelines — were not gazetted until 18 March 2026, as Legal Notice No. 43 in Gazette Supplement No. 67.⁷ Read those two dates together: for the first twenty-two months of a two-year transition window, the detailed rules an organisation was transitioning to had not been published.

I want to resist the easy reading of that. The easy reading is that the regulator was negligent or the state was hostile, and the series has already argued against both. Your NGO Is Illegal and That Is Not Your Fault makes the case that the organisations outside the system are not evading it, and the same courtesy is owed in the other direction: a registrar publishing regulations late is a registrar that is under-resourced for the job of designing them, which is a different failure from indifference and has a different remedy.

The record through this period is not one of an absent regulator. It is one of a regulator doing regulator-shaped work energetically: in the year to mid-2026, 4,779 organisations were de-registered for breaches of the Act and its code of conduct.⁸ Enforcement moved. Registration did not. That asymmetry is the observation this whole paper is built on, and §5 argues it is exactly what the budget structure predicts.

4. The natural experiment nobody designed

Here is the part that changed my mind about how strong the claim in §5 could be.

If the compliance layer is missing because compliance is expensive for the organisation, then removing the expense should move registration. In April 2025 the High Court removed a large part of it. In Otieno & 2 others v Attorney General & another; Katiba Institute & 9 others (Interested Parties), the court held that requiring organisations already validly registered under the repealed NGO Co-ordination Act to re-register under the PBO Act was an unreasonable regulatory hurdle inconsistent with freedom of association and legitimate expectation, and directed that such organisations be transitioned automatically — without fresh applications and without fees.¹ The same judgment struck down compulsory donor and member data disclosure and compulsory membership of the national federation.

This is close to the best case a reformer could have asked for. A court removed the fee. A court removed the application. A court removed two of the requirements most likely to make an organisation hesitate. And the transition still had to be extended by a year, and organisations were still being told, in the final weeks before the extended deadline, to complete their submissions or lose standing.⁹

I am not going to overclaim this as an experiment, because it was not one: the judgment was appealed in part, the practical effect of "automatic" transition still left documentation to be submitted, and organisations do not necessarily know within weeks what a court has done for them. Legal change propagating slowly into behaviour is ordinary and is not evidence of anything by itself.

But the direction is informative, and it is informative in a way that is awkward for the two most popular explanations of the gap.

It is awkward for "the fee is the barrier." The fee was waived for the largest affected class and the class did not clear.

It is awkward for "organisations are avoiding the state." The judgment moved the balance decisively toward the organisations, on precisely the grounds — data disclosure, forced affiliation — that the avoidance story says they were worried about. If avoidance were the mechanism, a judgment that removed the reasons to avoid should have produced a visible inflow.

What survives is duller and, I think, correct: the thing standing between an organisation and registration is not a fee, and not a fear, but the work — the translation, sequencing and verification described in §2. Nobody removed that, because a court cannot. It has to be built, and §5 is about why it has not been.

5. Three budgets, and why none of them reaches it

5.1 The state is funded to process, not to reduce

This is not cynicism about regulators; it is how regulators are funded. A registration authority is resourced to process applications and enforce the statute. It is not resourced to reduce the cost borne by applicants, and no line item rewards it for doing so.

Look at what the regulations the Authority did produce actually optimise. They specify the Authority's timelines with real precision — sixty days to determine an application, thirty days' notice before suspension or cancellation, sixty days of name reservation.⁷ Those are commitments about the regulator's conduct, and they are good ones. There is no corresponding instrument anywhere in the framework that commits to reducing what an applicant must do to produce a complete application in the first place. The Authority's throughput is measured at its own door.

The de-registration figure in §3 is the same fact seen from the other side. An authority that de-registers 4,779 organisations in a year is not idle, and it is not failing at its job. It is succeeding at the job it is funded for. The state's budget stops at the edge of its own process, and the layer sits on the other side of that edge.

5.2 The donor buys a filter, and the layer is not a filter

Your Donor Is Now Your Regulator made the case that enforcement has effectively migrated from the state to the funder, through due-diligence checklists. That paper stopped at the observation. The budget consequence is the part it did not draw out.

A due-diligence checklist is a filter, and a filter is cheap. Its cost scales with the number of applicants reviewed and it is charged to overhead. Paying for the applicants to become compliant is not a filter, it is a programme: it scales with the number of organisations helped, it takes years, and it competes for money against the things the donor exists to fund. Given the choice between funding ten clinics and funding the capability that would let forty more organisations qualify to run clinics, the second is always the harder sell, because its output is eligibility rather than service delivery.

There is a sharper version of this that I think is the real mechanism. The filter does not merely fail to fix the problem; it insulates the donor from the problem. An organisation that cannot produce a certificate never reaches the part of the process where anyone would notice it existed. The cost of the missing layer is therefore borne almost entirely by parties who are, by construction, outside the donor's field of view. A cost that is invisible to the party with the money is a cost that does not get budgeted for, and no amount of good intent changes that, because the good intent is never presented with the case.

5.3 The organisation needs it most and can pay least

This is the part that makes the whole thing stable.

Under the 2026 regulations, the statutory costs are knowable. Name reservation is reported at KES 1,000, standard local registration at KES 25,000, annual report filing at KES 2,000, and a change to the constitution or the governing body at KES 4,000.¹⁰ Those are the state's prices, and by the standards of professional services they are modest — which is precisely the point. They are not what makes registration hard. A community organisation that could produce a compliant constitution, minutes recording the right resolution, particulars of a properly constituted governing body and audited accounts would find the fees the easiest part of the exercise.

The costs that are not on that schedule are the ones that bind: the advocate who drafts the constitution so that it contains what Section 14 and the schedules require, the accountant who produces accounts that will survive filing, and the person inside the organisation who has to hold the whole sequence in their head for the months it takes. None of those has a gazetted price, and I do not have a verified distribution for them. That is the single most important number missing from this paper and it is registered as such in Appendix B.

Note also that the annual figures are the ones that matter for the argument. Registration is an event; compliance is a subscription. Form 13 — audited accounts and activity reports — falls due within six months of each financial year end, governing-body changes must be reported within thirty days, and other changes within sixty.¹¹ An organisation that clears the bar once and then loses its bookkeeper is back outside it within a year. The layer is not something an organisation needs on the day it registers. It is something it needs permanently, which multiplies both the value of the layer and the cost of providing it.

And the willingness to pay is real while the ability is not. The organisations with the ability to pay are, definitionally, the ones that already cleared the bar and no longer need the layer.

Put the three together and you get a demand curve inverted against an ability-to- pay curve, with the intersection at approximately nobody. That is what the figure above is drawing. It is not that the market is inefficient. It is that, at the prices the beneficiaries can pay, there is no market.

6. Why the obvious answers do not work

"Make it a product." A SaaS tool priced for the organisations that need it cannot fund its own maintenance, and compliance tooling is maintenance-heavy: the requirements move, and a compliance tool that has drifted from the statute is worse than none, because it produces confident wrong answers. This is the failure mode that matters most and the one a thin-margin product is least able to avoid.

The last two years make this concrete rather than hypothetical. A product built against the Act as it stood in May 2024 would have needed revision when the transition deadline moved in May 2025, again when the High Court struck down re-registration and the disclosure provisions, and again — substantially — when the regulations were gazetted in March 2026 with the forms and fees in them. That is three forced rewrites in under two years, against a customer base that by assumption cannot pay much. Maintenance is not a tax on the product. In this category it is the product.

"Let the sector build it collectively." Umbrella bodies are the natural candidate and they have not done it. I do not think this is negligence. A membership organisation is funded by its members, and its members are the organisations that already registered. There is a further wrinkle specific to this jurisdiction: the Act's compulsory-federation provision, which might have produced a body with sector-wide reach and a sector-wide mandate, was struck down in the same judgment discussed in §4, on grounds I think are correct.¹ The remedy that protected freedom of association also removed the one structure that had a plausible claim on funding the layer. I do not raise that as a criticism of the ruling. I raise it because it is a real trade-off and the sector should see it.

"The regulator should publish better guidance." Guidance is a document. The gap this series has documented is not a gap in documents — the Act is published, the requirements are knowable. The gap is between a requirement and a procedure, and no amount of better prose crosses it. That is exactly what the fourth paper argued, and it is why "publish a handbook" has been tried and has not moved registration.

"Language models have made this nearly free." This is the newest version of the objection and the one I have the most sympathy for, having made the underlying claim myself in the third paper of this series. The generation and checking cost really has collapsed. But §2 separates four jobs, and the collapse applies cleanly to one of them — verification — partially to translation, and not at all to maintenance and accountability. Somebody still has to know that Legal Notice No. 43 exists, read it, decide what changed, and carry the consequence when the tool is wrong. A cheap generator with nobody accountable for its currency is the "confident wrong answers" failure mode with better prose. The model reduced the build cost, which was never the binding cost.

7. Objections I take seriously

"This is just the standard public-goods argument dressed up." Partly, and I should concede the overlap. Where it differs is the inversion in §5.3. A standard public good is underprovided because its benefits are non-excludable; this one is underprovided because the beneficiaries' ability to pay is anti-correlated with their need, and the correlation is produced by the very bar the good exists to help them clear. That is a stronger and more specific condition, and it predicts something the general argument does not: subsidy alone will not fix it if the subsidy is routed through the beneficiaries, because the routing reproduces the selection.

"Registration is low because many of these organisations should not register." A serious objection. Some fraction of the estimated fourteen thousand are dormant, duplicated, or were never really operating. If that fraction is large, the gap is an artefact of a bad denominator. I accept that this weakens the size of the claim and I have no way to bound it. It does not touch the structure: whatever the true denominator, the organisations at the bottom of it are the ones this argument is about, and nothing about a bad count explains why no party funds the layer.

"The state did eventually publish regulations, so the state's budget does reach." It reaches as far as specifying the obligation, which is the state doing its job well. §5.1's claim is narrower: no budget line rewards reducing the applicant's cost of meeting that obligation. The regulations are evidence for this, not against it — they commit the Authority to its own timelines and commit nobody to the applicant's.

"You are describing a business you would like someone to fund." Fair, and worth naming: Flux builds compliance software, so I have an interest in the conclusion that compliance software is undersupplied. The mitigation I can offer is that the paper's actual conclusion is hostile to my interest. It says the layer is not a viable business at the prices its users can pay, and that anyone building it should expect to be funded as infrastructure or to run it at a loss. That is not the conclusion a vendor writes to raise money.

8. What would actually change it

I can see three shapes that break the deadlock, and I want to be honest that I have tested none of them.

The funder pays for eligibility, not just for programmes. Compliance becomes a condition of disbursement that the funder also underwrites — the donor funds the layer as portfolio infrastructure, because it is cheaper than losing access to the grassroots organisations its due diligence is currently filtering out. This is the one I find most plausible, because it is the only one where the party with the money also holds the problem. It requires the donor to notice a cost that §5.2 argues is structurally invisible to it, which is the hard part, and it is why the most useful thing anyone could do for this argument is to measure the size of the filtered-out population.

The state treats registration throughput as an outcome it owns. Not a queue it processes — a number it is accountable for moving. That is a change in what a registrar is for, which is a political change, not a software one. It is the most durable of the three if it happens and the least likely to.

Somebody builds it as a loss. Cross-subsidised from unrelated revenue, on the argument that the returns are real but accrue to the sector rather than the builder. That is a legitimate reason for a lab to do a thing. It is not a business model, and it should not be described as one.

There is a fourth that I considered and rejected, which is that the layer gets built as a by-product of professional services — an advocate's internal tooling generalised outward. I rejected it because the economics point the wrong way: the tool's value to its owner is that it makes their billable work faster, and releasing it destroys the scarcity that makes the work billable. Where this has happened in other regulated domains it has happened after the underlying service was already commoditised, not before.

9. What this changes about the series

If the argument holds, three things.

The series should stop ending on "somebody should build this." Five papers established the gap thoroughly. The gap is not the finding any more.

The finding is that the gap is load-bearing — it persists because the incentives hold it in place — and anybody planning to close it should begin by working out whose budget they are going to change. That is a different first question from the one a builder normally asks, and asking it second is how the four-year-old version of this idea would fail.

And the unit of the problem is annual, not one-off. §5.3's point about Form 13 and the thirty-day change notifications is the one I would most want a future paper to develop, because it is where the difference between a tool and a layer actually lives.

10. What would falsify this

Stated as commitments, in decreasing order of how badly each would damage the argument.

A layer appears, is used at scale, and is funded by subscriptions from the organisations using it. That directly contradicts §5.3 and I would withdraw the paper's central claim.

Registration completes without any layer being built. If the transition clears through some combination of extended deadlines, regulator outreach and ordinary professional services, then the layer was never load-bearing and I mistook a slow process for a stuck one.

The layer appears first in a setting with many small funders rather than one large one. §5.2 predicts the opposite: it should appear first wherever a single funder is large enough to internalise the whole cost. A counterexample would not kill the paper but would falsify its most specific prediction.

The prediction, drawn. If the layer appears anywhere but here, the account in this paper is wrong.

A measured distribution of professional-services costs for PBO registration shows them to be small relative to the median unregistered organisation's budget. Then the ability-to-pay inversion in §5.3 is wrong, and with it the claim that this is not an ordinary market.

11. Limitations, and what this paper does not claim

It does not claim to have built the layer. It does not report results from a deployment, because there has not been one.

It does not claim the three-budget account is the only explanation. It is the one that survives when I try to knock it down, but it is an argument from incentives, and arguments from incentives are the easiest kind to make and the hardest kind to falsify. §10 is my attempt to hold myself to something.

It does not generalise beyond Kenya on the evidence presented. The structure — regulator funded to process, donor funded to filter, beneficiary unable to pay — is not obviously Kenya-specific, and I expect it travels. But everything empirical in this paper is Kenyan, and a reader in Ghana or Bangladesh should treat the generalisation as a hypothesis.

And it does not resolve the question of who should build it, which is a different paper and a harder one, because the honest answer depends on the falsification tests above coming back.

Appendix A — Evidence table

The verified rows are cited in the footnotes. The unverified rows are the measurements this argument needs and does not have; they are left empty rather than estimated. Source column marks provenance: G gazette or primary legal instrument, J judgment, P press report of a regulator statement, — not yet sourced.

Rows 13–17 are the ones that would move this paper from an argument to a finding. Row 16 is the one I would fund first: §5.2's claim that the cost is invisible to the party with the money is the load-bearing step, and it is the only one a donor could check against its own records in an afternoon.

Appendix B — Open questions and citation register

CITATION-NEEDED — professional-services cost. The per-organisation cost of PBO registration compliance in Kenya beyond the gazetted fees: advocate's fees for constitution drafting, accountant's fees for the accounts required at filing, and the internal staff time. Appendix A rows 13–14. The shape of the argument in §5.3 does not depend on the numbers, but the argument would be much stronger with them, and I am not going to invent them.

CITATION-NEEDED — fee schedule against the primary instrument. The fees at rows 8–11 are taken from professional summaries of Legal Notice No. 43 of 2026 and agree with each other, but I have not read the schedule in the gazette myself. Before these numbers are quoted anywhere consequential they should be checked against the instrument at Kenya Law.⁷

CITATION-NEEDED — appellate history. The 2025 High Court judgment has been the subject of further proceedings, and the state of the law as at publication should be confirmed before §4 is relied on for anything other than the directional argument made here.

CITATION-NEEDED — the counterfactual jurisdiction. §10 predicts the layer appears first where a single funder internalises the cost. I have not looked for an existing instance in another regulatory domain — occupational licensing, food safety certification, building approvals — where exactly this happened. If one exists it is the best available test of the whole argument and I have not run it.