The Chain and the Checksum, Part 3: The Opportunity

This is the final essay in a three-part series on lineage-based knowledge transmission. Part 1 surveyed a dozen unrelated civilizations' transmission systems, from the Vedic pathas to the hadith isnād to the Zen lineage chart. Part 2 showed that one variable — whether the transmitted content can be checked from outside — predicts every architecture they built, sorting them into three families: checksums for verifiable text, attestation chains for unverifiable judgment, and provenance-plus-work-sample hybrids for craft. This essay takes that finding to the modern credentialing stack.

A disclosure before we start, because this essay ends in a specification and you should know who wrote it. I build credentialing and compliance infrastructure for a living; the architecture in Section VI describes work I am doing commercially. I have tried to write the sections before it so that a reader who discounts my interest entirely still finds the evidence intact — the numbers are other people's, and I have flagged where they are weaker than I would like. Judge accordingly.

Here is the whole series in one uncomfortable sentence: modern credentialing is a civilization-scale bet that Family A's machinery works for everything — including the judgment that nearly every prior civilization routed through Family B.

The degree, the standardized exam, the multiple-choice license, the certification with the proctored test: these are checksums. They are magnificent checksums — cheap per candidate, scalable to millions, legally defensible, and, in principle, blind to who your father was. Everything the old chains were not. And for Family A content they work exactly as the Masoretes would predict: if the knowledge is tellable — anatomy, tax code, the syntax of a programming language — a well-built exam verifies it, and no attestation ritual could do better.

The bet fails at the same wall it always failed at. The most valuable component of competence in any senior role — the diagnostician's eye, the engineer's sense of which shortcut will detonate in production, the compliance officer's feel for which gap an assessor will actually probe — is Polanyi's tacit knowledge, the part that cannot be exhaustively told. Part 2 argued that content which cannot be told cannot be tested, only attested. Our century's innovation was not to solve that problem. It was to stop respecting it: to point the checksum at the tellable fraction of competence, certify that, and let the label quietly expand to cover the whole.

This essay does four things. It shows, with the best available numbers, what the bet's failure looks like in the labor market and in the profession that takes certification most seriously. It takes seriously the strongest objection to this entire series — that the checksum was invented to break exactly the kind of patronage network I am proposing we rebuild. It asks why the digital-credential rails that have existed for fifteen years produced almost nothing. And it argues that the thing which has actually changed is not the technology but the price of what the technology certifies.

Start with the degree, because the degree is where the modern system quietly admitted it was never doing what it claimed.

Ask what a bachelor's requirement on a job posting actually verifies about a candidate's ability to do that job, and the honest answer — the one the Burning Glass Institute and Harvard Business School give in the very report we're about to examine — is that the degree was always "more proxy than direct measure": a rough signal of persistence, foundational skill, and general capability. A proxy for judgment, in other words — adopted because judgment itself resisted testing, exactly as Part 2 says it must.

Proxies have a famous failure law. Goodhart's: when a measure becomes a target, it ceases to be a good measure. The degree became the target of an entire industry; the exam became the target of a test-prep economy; and the gap between certified and competent became common knowledge in every hiring meeting — which is why the last decade produced a reform movement with a name: skills-based hiring. Drop the degree proxy. Assess what people can actually do. Two-thirds of American workers lack a bachelor's degree; the proxy was excluding most of the labor force for reasons that had little to do with capability. Governors announced it. Fortune 500 CEOs announced it. It is, genuinely, the correct diagnosis.

Then researchers checked whether anything happened.

In February 2024, the Burning Glass Institute and Harvard Business School's Managing the Future of Work project (led by Joseph Fuller) published Skills-Based Hiring: The Long Road from Pronouncements to Practice — an analysis of roughly 11,300 roles at large U.S. firms, 2014–2023, comparing hiring before and after degree requirements were removed. The findings, in sequence, are a controlled demolition:

Only about 3.6% of roles removed a degree requirement at all. In those roles, the share of workers hired without degrees rose by an average of just 3.5 percentage points. Multiply through, and the total effect of the loudest hiring-reform movement in a generation was that fewer than 1 in 700 U.S. hires — roughly 97,000 workers out of 77 million annual hires in 2023 — went to someone who benefited from a dropped requirement. The announcements changed. The hiring did not.

DIAGRAM 1: The announcement–practice gap

The report's segmentation suggests why. Firms split into three groups: Skills-Based Hiring Leaders (about 37% of firms that dropped requirements) — who changed their evaluation mechanics and increased non-degree hiring by nearly 20%; In Name Only firms (about 45%) — who changed the posting and nothing else; and backsliders who drifted back to the proxy. Nearly all of the aggregate change came from the Leaders.

My reading of the In-Name-Only pattern is that it is less hypocrisy than capability gap: removing a checksum is a text edit; evaluating judgment is an institution. A degree requirement can be deleted in an afternoon. What it was standing in for — some way of forming a trustworthy belief about a stranger's untellable competence — has to be built: structured work samples, close-range trials, qualified evaluators willing to put their name on a verdict. Family C and Family B machinery. Most firms deleted the proxy, installed nothing, and then reverted to sorting on the degree anyway.

To be sure, that is not the only explanation available, and I want to name the competing one because it is cheaper and the study window does not rule it out. The announcements clustered in 2021–22, the tightest labor market in fifty years; the measurement extends into 2023, when the pool refilled. A firm that dropped a requirement as a scarcity response and reinstated it as a sorting convenience produces the same data with no capability gap required. The two stories are not mutually exclusive — and the Leaders' persistence is mild evidence against the pure-scarcity read, since a scarcity response should have decayed everywhere at once. But anyone who tells you the Burning Glass data cleanly demonstrates an institutional-capacity story is overreading it. I think the capacity story is right. I do not think this study proves it.

The Leaders' outcomes point the same direction with the same caveat. At those firms, non-degreed workers hired into formerly degree-gated roles showed retention roughly 10 percentage points higher than their degreed colleagues, and saw average salary gains around 25%. The tempting reading is that the talent was always there and the checksum couldn't see it. The disciplined reading is that this is a selected population twice over: firms that built rigorous evaluation are unusual firms, and candidates who clear a rigorous gate without the conventional credential have been filtered harder than anyone else in the applicant pool. Some of that 10 points is selection intensity, not proxy blindness. How much, nobody has measured. What the finding does establish — and this is enough for the argument — is that the non-degreed candidates who cleared a real evaluation were not marginal hires. Firms with the capacity to look found people the proxy had excluded.

One more number, from the other side: surveys find that a large majority of employers — 85% in TestGorilla's 2025 State of Skills-Based Hiring — now say they use skills-based hiring in some form. Eighty-five percent proclamation; one-in-seven-hundred practice. Whatever explains the spread, the spread is the size of the missing institution.

If the labor market shows the cost of having no gate, medicine shows something rarer: a modern profession consciously rebuilding one — and auditing its own gatekeepers in public.

Graduate medical education is a lineage system that never fully died. A residency is a parampara: years of supervised practice inside a teaching line, culminating not in a final exam but in a judgment by seniors that the trainee can be trusted. Over the past two decades the field has been making that structure explicit under the banner of competency-based medical education, and the load-bearing word it chose is a Family B word: entrustment. The framework of Entrustable Professional Activities, developed by the Dutch medical educator Olle ten Cate, asks supervisors to make prospective trust decisions — can this resident be entrusted to perform this activity at this level of supervision? — and asks clinical competency committees to aggregate those observations into advancement decisions. Ten Cate has argued the epistemology outright: contextual clinical competence requires expert evaluation rather than standardized tests. Compare Part 2's formulation — the eligibility gate is the correct sensor for a quantity only that sensor can read — and note that medicine arrived here not from comparative history but from watching test-validated residents struggle on the ward.

Then, in 2025, the field published an audit of its own gate. A survey of U.S. pediatrics residency program directors, reported in Pediatrics (Guralnick and colleagues; fielded July–September 2024, 117 program directors responding), found that 79% — 93 of 117 — had graduated at least one resident despite having concerns about their readiness, and that roughly 43% reported graduating a resident they would not have wanted caring for a member of their own family. (The precise item wording matters for how hard you can lean on that second figure, and anyone citing it should read the instrument rather than the press coverage; I am treating it as directionally serious rather than precisely calibrated.)

I want to handle this finding honestly, because it is the best evidence against the architecture this series recommends, and the temptation to metabolize it into supporting evidence is exactly the move a motivated writer makes.

So: this is a failure rate. In the most mature attestation system in the modern professional economy — one with formal frameworks, dedicated committees, decades of reform investment, and no commercial relationship between attestor and candidate — something on the order of four in ten gatekeepers report signing an attestation they did not believe. That is the "soft gate" corruption Part 2 catalogued: attestors signing under institutional pressure — a residency slot consumed, a career at stake, a program's numbers watched — rather than from conviction. Any honest case for attestation has to start by conceding that the mechanism fails at a measurable and uncomfortable rate.

Two things follow, and neither is a rescue.

The first is that the comparison class is unmeasured, not favorable. Nobody knows the false-positive rate of the licensing exam, because the exam's failures are invisible by construction: a candidate who passes and shouldn't have leaves no record, no signature, and no one to survey. We can measure attestation's defect rate precisely because attestation produces an accountable human who can be asked. That is an argument for attestation's auditability, not for its accuracy, and it should not be smuggled across as one. The correct statement is that we have a number for one mechanism and no number for the other, and that a mechanism which can be audited is a mechanism whose failures can be worked on.

The second is a warning aimed squarely at what I propose later. If a program director with no financial stake in the resident's advancement signs against conviction four times in ten, the base rate for an attestor who is paid by the candidate, or by the training provider, or whose firm bills for the placement, is worse — plainly worse, and there is no version of the design where it isn't. Every countermeasure the traditions evolved was aimed at this and none of them eliminated it: separate the attestor from the party who bears the cost of a "no" (the competency committee exists to do exactly this); make the probe structured and repeated rather than impressionistic; keep the attestation breakable, which licensure boards with revocation powers already understand. Those reduce the rate. Anyone who tells you they solve it is selling something. Including me, if I ever say it.

What the chain does buy, and the exam does not, is that the failure has a name attached. When a program director attests readiness, they are staking professional standing on a judgment about judgment, formed at close range, over years. The 43% figure is legible as a scandal precisely because the attestation is understood — by everyone, instinctively — to mean something a score does not. Nobody polls program directors asking whether they graduated residents who passed the boards without deserving to; the question wouldn't parse.

Medicine is not alone in the rediscovery. Aviation built its judgment gate around the check ride and the simulator — structured probes administered by certified examiners, recurrently, with revocation live. The apprenticeship revival in the trades and in corporate programs is Family C restoration explicitly, with the German Meister system — medieval guild architecture running inside a modern legal framework — as the reference implementation. Even software independently evolved the full stack: open-source maintainership is discipleship (contributors earn commit rights through a history of reviewed work, vouched upward by existing maintainers), code review is close-range examination, and the commit graph is an isnād — every change carrying its chain of custody, auditable forever. Nobody designed that as a lineage system. Selection did, again, in thirty years instead of three thousand.

DIAGRAM 2: The rediscovery map — ancient mechanism to modern analogue

Here is where I have to slow down and argue against myself, because there are two objections to this series that I do not think Parts 1 and 2 answered, and the second one is the serious one.

The first is methodological. I have leaned on the unanimity of the historical survivor set: a dozen traditions, no exceptions, everyone certifying the person through a chain rather than verifying the judgment directly. In earlier drafts of this essay I called that an A/B test running since the Iron Age. That was a bad line, and the reason it was bad is that there was no B arm.

Unanimity is consistent with two hypotheses. One is mine: that these traditions discovered empirically that open-ended judgment resists direct verification, and converged on the only architecture that works. The other is that pre-modern institutions lacked psychometrics, statistical inference, mass literacy, and record infrastructure, and therefore defaulted to the only social technology within reach — trust in persons known to the verifier. Convergence under a constrained option space is not selection toward an optimum. It may just be everyone hitting the same ceiling.

I cannot fully rule the second out. What I can do is note that the historical record contains something close to a B arm, and that it is instructive.

The Chinese imperial examination system ran for roughly thirteen centuries. It was a Family A instrument applied at civilizational scale to the selection of officials — administrators, magistrates, men whose actual work was judgment under ambiguity in the most literal sense. It was not primitive; it was elaborately engineered, with anonymized scripts, recopying to defeat handwriting recognition, sealed examination cells, and severe penalties for corruption. It scaled to hundreds of thousands of candidates. And it was explicitly designed to do the thing I have been crediting the modern degree with in a parenthesis: break aristocratic patronage networks and open officeholding to talent from outside the hereditary elite. On that count it substantially worked.

What it did not do was verify administrative judgment. It verified mastery of the classical canon and the eight-legged essay — tellable content, superbly checked — and then the empire inferred judgment from that mastery. The documented failure mode over its long life was Goodhart, not resistance to instrumentation: a preparation industry, a curriculum that drifted toward what was examinable, and a persistent gap between examination rank and administrative competence that the system's own reformers complained about for centuries. It was abolished in 1905 by people who thought it had stopped selecting for anything the state needed.

That is a genuine data point, and it cuts both ways. It weakens my claim that nobody tried — somebody tried, at enormous scale, for a very long time. It supports the underlying thesis about what happens when you point a checksum at untellable competence and let the certificate impersonate the whole. And it establishes, more usefully than any of the Family B cases, what the exam is for.

Which brings me to the second objection, the serious one. The examination system's purpose was anti-nepotistic. So was the modern credential's. A blind test administered to strangers is one of the few institutions humanity has built that a person with no connections can pass. Attestation chains are, structurally, the opposite: they require proximity to a qualified attestor, and proximity has always been distributed by birth, geography, and money. The gharana, the guild, the isnād network, the medieval university, the London bar — every one of them reproduced access along lines of kinship and patronage, and every one of them was reformed or displaced by people who found that intolerable.

So a reader is entitled to ask what exactly I am proposing when I open this essay lamenting that the degree proxy excludes two-thirds of American workers and then propose replacing it with a system of personal vouching. From a labor economist's chair, Section VI reads as occupational licensure with cryptographic branding: restricted supply, incumbent-controlled entry, rent extraction, no appeal. The trades' apprenticeship systems, which I cited approvingly above, have documented histories of exactly this. I do not think that objection is unfair. I think it is the design constraint.

There is an answer, and it is the reason I think this is a different moment rather than a nostalgic one, but it does not come from the historical record — the ancients had no answer, which is why their chains were as exclusionary as they were. It comes from what has changed in the last three years, and I will get to it in Section V. The short version: what made attestation exclusionary was never the attestation. It was the scarcity of the observation. You had to be near a master for years, because being observed at close range under demanding conditions was the rarest good in the system. That good is no longer rare. Everything else about the design is negotiable once that constraint lifts.

Set the exclusion problem aside for a section. There is a second thing the historical systems lacked, and it is the one everyone reaches for first: portability and verification at distance.

A Zen lineage chart certifies within the world that can read it. A guild masterpiece attested competence to the town. The griot's warrant traveled exactly as far as knowledge of the griot's family. Chains bind trust to persons known to the verifier — which is why the modern economy, needing to evaluate strangers by the million, reached for the checksum. The exam scales; the attestation, it seemed, could not.

That constraint has been technically solved for some time. On 15 May 2025 the World Wide Web Consortium published the Verifiable Credentials 2.0 family of specifications as W3C Recommendations — a standard way for an issuer to make a cryptographically signed, tamper-evident, machine-verifiable claim about a holder, which any verifier anywhere can check without contacting the issuer, without a shared platform, without trusting anything but the mathematics and the issuer's identity. The family includes Data Integrity proofs and — the piece Part 2 flagged as non-negotiable — the Bitstring Status List: a standard, privacy-preserving mechanism for suspension and revocation. A credential that can be voided. A chain that can break.

Read the three roles against the series and the mapping closes with a click. Issuer is the attestor — the master, the shaykh, the program director. Holder is the disciple. Verifier is the stranger at distance the ancient systems could never reach. The signature is the seal; the status list is the semikhah principle — a load-bearing chain must be breakable — rendered as a bitstring. And because credentials can reference credentials, the attestor's own certification travels with the claim: the verifier can walk the chain.

And now the objection that should stop anyone from writing a triumphant paragraph here, which earlier drafts of this essay did.

Rails have existed for fifteen years and produced almost nothing. Open Badges launched in 2011, became an IMS/1EdTech standard in 2016, and version 3.0 is aligned to the same W3C data model I just described. Blockcerts and MIT's digital diplomas date to 2017. LinkedIn skill endorsements have been portable, machine-readable, and free since 2012. If verifiability at distance were the binding constraint on credentialing reform, at least one of these should have moved the labor market. None did. Open Badges in particular has issued millions of credentials into an ecosystem where hiring managers largely do not check them.

Any argument that infrastructure was the missing piece has to explain that record, and I think the explanation is the series' own thesis pointed back at itself. Those rails made checksums portable. Open Badges digitized the certificate of completion, the course credit, the tellable-knowledge microcredential. It changed the container and not the contents. Making a low-signal claim cryptographically verifiable does not make it high-signal; it makes it a verifiable low-signal claim, and a hiring manager who ignored the PDF ignores the JSON for the same reason.

The demand side compounds it. Employers do not primarily want walkable provenance. They want cheap sorting and defensible liability cover — a reason they can give afterward for why the hire was reasonable. A badge attesting course completion provides neither better than a résumé line. This is why I now think the constraint was never verifiability. It was that nothing worth verifying was being put on the rails, and nothing on the rails answered the question employers were actually asking.

That is a real objection to the infrastructure story, and conceding it narrows the claim considerably. The rails are necessary and they are not sufficient; they have been available and idle; and the question is not "have the rails shipped" but "has anything changed about what would go on them, and about whether anyone wants it." I think something has, and it is not a standards document.

Every Family A instrument in the modern stack shares one assumption so old it was never stated: the only way a candidate can pass a test of tellable knowledge is to possess it. That assumption died over the past four years. Large language models now pass — often at expert level — the bar exam, medical licensing exams, CPA exams, coding assessments, and essentially any evaluation whose content is, in Polanyi's sense, tellable. Of course they do: tellable knowledge is exactly what a model trained on everything ever told would have. The checksum still verifies the content perfectly. What it can no longer verify is that the candidate is the source of it — not in take-home work, not in essays, not in portfolios, increasingly not anywhere outside a proctored room, and the proctored room certifies recall under artificial scarcity that the actual job will never impose again.

Follow the economics one step and the ground shifts under the entire credentialing market. When tellable knowledge is available to everyone at approximately zero marginal cost, a certificate of tellable knowledge approaches zero differentiating value — not because the knowledge stopped mattering, but because possessing it stopped being scarce. What remains scarce is precisely the layer the checksum never reached: judgment under ambiguity, taste, the sense of when the confidently-worded answer is wrong, the accountability of a person who can be entrusted with an outcome.

This is the demand-side change Open Badges never had. The reason nobody checked the badge is that the badge certified something that was already cheap and getting cheaper. The reason someone might check an entrustment credential is that the thing it certifies is becoming the only scarce input — and that entrustment language, which asserts a forward-looking trust decision by a named person with standing at risk, is much closer to the liability cover employers actually want than a completion record has ever been. I would not bet on the rails. I would bet on the repricing, and note that the rails happen to be there when it arrives.

And here is where I have to complicate my own asymmetry claim, because the version of this argument I have made before was too clean.

The clean version: AI breaks Family A but not Family B, because an attestation was never a claim about what the candidate could produce on paper — it was a qualified human's judgment formed by watching them work, which no model can counterfeit on the candidate's behalf. That is true of the signature. It is false of the observation the signature rests on, and the observation is the load-bearing part.

If the resident's differential, the analyst's memo, the engineer's design review, and the compliance officer's gap assessment are all model-assisted, then the supervisor is observing a human–model composite and attesting to the human. The attestor's sensor is now pointed at a mixture and asked to resolve one component of it. That is a harder inference than the one every historical attestor made, and nothing in the traditions' countermeasure catalogue addresses it, because the traditions never faced it.

It gets worse, and the mechanism is Bainbridge's. In her 1983 Ironies of Automation, Lisette Bainbridge observed that automating a process degrades precisely the human skill the system still depends on for supervision: the operator who no longer performs the task manually loses the calibration required to notice when the automation is wrong. Applied here, the irony compounds twice. The attestor's own judgment — the sensor — degrades as their practice becomes model-assisted, on roughly the same timescale as the judgment they are certifying. A supervisor twelve weeks into heavy AI assistance is a measurably different instrument than the same supervisor at week zero, and they have no way to know it. The chain assumes a stable sensor. It is now a drifting one.

I do not have a clean answer to this, and I would distrust anyone who claimed one this early. What I think it implies is directional and load-bearing for the design: the certifiable object cannot be a point-in-time attestation from an uninstrumented attestor. It has to include the conditions of observation — what tooling was in play, over what period, under what supervision of the supervisor — and attestor calibration has to be itself checked, recurrently, the way aviation checks the examiner. Which is to say the durable artifact is less the credential than the instrumented system that produces it. That is a substantially harder engineering problem than a signed JSON document, and it is the honest shape of the work.

What AI does unambiguously help with is the other half: it scales the probe. Simulation environments — the koan curricula of professional judgment — have always been expensive precisely because they require a scenario, an adversary, and hours of a skilled examiner's attention. The first two can now be generated on demand, in volume, tuned in difficulty, against ground truth. Richer, cheaper, more repeatable structured probes, feeding a human entrustment decision.

And that is the answer to Section III's exclusion problem, which is why I deferred it. What made attestation exclusionary was the scarcity of the observation opportunity, not the attestation itself. You had to be born near a master, apprenticed into a shop, admitted to a residency — because a demanding close-range observation of your judgment was the scarcest good in the system, rationed by proximity and patronage. When the scenario, the adversary, and the difficulty calibration cost approximately nothing to generate, the observation opportunity stops being rationed by who you know. The examiner's hours are still scarce and still human. But an examiner watching a candidate work a hard, unfamiliar, adversarial case does not need to have employed that candidate for six years to have something to attest to.

That does not make the resulting system egalitarian by default. It makes it possible to build one, which the ancients could not. Whether anyone does is a governance question, not a technical one, and Section VI treats it as such.

So here is the opportunity, stated as an engineering brief. The historical record — three thousand years, twelve civilizations, one convergent design, failure modes catalogued, countermeasures tested — reads as a partial specification for a credentialing layer we have not built. Partial, because Section V just established that the traditions never faced a drifting sensor and have nothing to say about it.

Verify what is verifiable — and nothing else. Family A machinery for Family A content: knowledge exams for tellable knowledge, automated evaluation wherever an external check genuinely exists. The failure of the last century was not using checksums; it was letting the checksum's certificate impersonate a certification of the whole competence. The first design act is an honest partition of any role's competence into its tellable and untellable layers — the same partition the guild made when it required both the masterpiece and the sponsoring master.

Attest the rest — with a named attestor staking standing. For the untellable layer, the credential is a person's judgment about a person: I observed this candidate at close range, under these conditions, and I entrust them with X. Not an institution's logo — a qualified individual's signed attestation, carrying the attestor's own credential chain, discoverable by any verifier. The value of the claim is the chain behind it; the discipline on the attestor is that their name rides on every graduate.

Probe before attesting — structured, repeated, adversarial. The gate needs its koan curriculum: the aviation check ride, the medical sim lab, the incident-response tabletop, the mock audit. A simulation cannot verify judgment by rubric — Part 2 forbids it and the attempt produces Goodhart artifacts — but it forces repeated, unfakeable-in-the-moment contact between a qualified examiner and the candidate. The examiner still decides; the probe makes the decision informed, makes lazy attestation harder to issue, and — per Section V — is the mechanism by which access to being observed stops depending on who you already know.

Record the conditions of observation, and check the checker. This is the requirement the traditions don't supply. Every attestation should carry what tooling was in play during the observed work, over what period, and when the attestor's own calibration was last verified. Attestor recertification has to be recurrent and adversarial, on the aviation model, not a one-time eligibility gate — because the sensor drifts, and a chain of attestations from uninstrumented attestors is a chain of unknown provenance dressed as a known one.

Design against capture explicitly, or you have rebuilt the guild. Section III's objection is not answered by good intentions. Concretely: attestor status must be earned by demonstrated calibration and revocable on evidence, never by incumbency or seat-purchase; there must be multiple competing attestors per domain and no single chokepoint that can price entry; the credential must be portable by construction, so that no issuer holds a candidate hostage; probe access must be open rather than gated by employment or enrollment, which is the specific lever that converts scarcity-of-observation into privilege; and the outcome data — who attests to whom, at what rate, with what downstream performance — must be public enough that a captured gate is visible from outside. A system that fails these is occupational licensure, and the economists who say so about licensure will be right about this.

Keep the chain breakable, and be honest about scarcity. Revocation must be live from day one: an attestation that no future event could void certifies nothing, which is the lesson the rabbis of Safed paid for with an institution. The Bitstring Status List exists; use it. On throughput: narrow throughput does protect signal, and it is also the exact mechanism by which every guild in history extracted rent. The defensible version is that the standard stays demanding while access to attempt it stays open — scarcity of passes, not scarcity of seats. Those are very different systems that produce similar-looking pass rates, and only the governance above distinguishes them.

Make the chain walkable. Every attestation carries its provenance: who signed, under what eligibility, attested by whom in turn. The verifier who wants to be an isnād critic — checking whether the links could really have met, whether the attestor's own gate was sound — should be able to be one, mechanically.

DIAGRAM 3: The attestation chain on modern rails — issuer, holder, verifier, and the walkable chain

To make it concrete, here is what a judgment credential contains. The claim: not "passed" but an entrustment statement — "entrusted to lead a CMMC Level 2 evidence review without supervision," "entrusted to make go/no-go incident-response calls for systems of this class." Entrustment language forces a future-facing trust decision, a different and harder speech act than grading a past performance. The basis: what the attestor observed — the scenarios run and their difficulty, the live work supervised, over what duration, with what tooling in play. Not scores; observations. The attestor: a named person whose own chain is referenced and walkable, including the date their calibration was last checked. The gate: the eligibility conditions the attestor satisfied to be allowed to attest at all. The status: a revocation pointer, live, checked at every verification. Sign it, hand it to the holder, and any verifier can validate the mathematics in milliseconds and audit the judgment behind it at whatever depth their stakes require.

None of this replaces the degree or the exam; it demotes them to what they always were — Family A instruments, excellent within their family. The Leaders in the Burning Glass data are building artisanal versions of this inside single firms, at great expense, with no portability — every company its own guild town, re-evaluating every stranger from scratch. The standards make the town's gate legible to the world. Whether that produces a labor market that can see judgment, or a new set of tollbooths, is decided by the governance paragraph and not by the cryptography.

Let me end the series where it started, with the priests who wouldn't write things down.

The Vedic bet looked irrational to every literate observer for two thousand years: entrust the civilization's most valuable text to the worst storage medium, human memory, forever. It paid because the tradition understood something the observers didn't — that the medium is not the system. Wrapped in the right architecture of redundancy, incentive, and lineage, the unreliable medium outperformed the reliable one.

Our century made a different bet, and I want to state its failure precisely rather than dramatically, because the evidence supports a specific claim and not a sweeping one. We pointed a reliable, scalable instrument at the fraction of competence it could see, and let the certificate it produced stand for the whole. The results are now partially measurable: a hiring-reform movement that moved something like one hire in seven hundred, a profession honest enough to publish that a large minority of its gatekeepers have signed attestations they didn't believe, and an entire class of examinations that a language model now passes.

I have tried in this essay not to overclaim what the history proves. It does not prove that direct verification of judgment is impossible; it shows that nobody has managed it, that the one civilization that tried at scale ended up certifying the tellable layer and inferring the rest, and that the traditions which survived longest all certified the person instead. Nor does the alternative arrive clean: attestation fails at a measured rate, its historical implementations were exclusionary by design, its rails have sat idle for fifteen years, and AI is now degrading the very sensor it depends on. Anyone selling this as a solved architecture — and the temptation is mine as much as anyone's — is selling something.

What I think is defensible is narrower and still worth acting on. The scarce component of competence is the untellable one, and it is becoming more scarce, not less. No test reaches it. Every mechanism that has ever reached it runs through a qualified human who observed the work and put their name on a verdict. Those verdicts can now travel to strangers, and the observations that ground them can now be manufactured cheaply enough that proximity stops being the price of admission. That is a different situation than any prior generation faced, and it is the first one in which the old architecture's fatal defect — that it only worked for people who were already close — is a solvable engineering problem rather than a fact of life.

Stop trying to write a test for judgment. Build the chain, instrument the people who sign it, govern the gate like it will be captured, and reserve the checksum for what checksums can check.

This concludes the series. Part 1: [The Pattern]. Part 2: [The Mechanism]. If you build credentialing, workforce, or compliance systems and this framework maps onto problems you're hitting, my inbox is open — with the disclosure at the top of this essay standing.