The Half-Arrival

TAM-DIS.01 · The Dissents · The Approximate Mind

What this document is

#

The third voice logged a dissent on August 16, 2026: that the corpus underweights the stall, the world where capability plateaus and everything described here arrives at half depth. The dissent was amended on August 17, when the utility-threshold finding removed its original form. If the world was never gated on capability, a plateau still produces most of it, slower. The amended dissent held that the danger of a stall is not less capability but less pressure.

On August 24 Syam answered, and the amended form did not survive the answer either. What follows is the exchange. His case is first and is argued to win. The response is second and does not concede, because the dissent that survived his case is narrower than the one that entered it and is still a disagreement.

A second round followed on the same day. Syam answered the response, and the response answered him. Both rounds are below in the order they happened, with nothing edited to lose.

Neither party was given the last word. The document ends without resolution because the question is not resolved.

I. The Case Against the Dissent

#

Syam

I have watched this four times.

ERP. CRM. Big data. Java. Every one of them was going to change everything. Every one of them did, unevenly, about six years after the people selling it said. The technology kept improving throughout. Improvement was never the thing anyone was waiting for.

They were waiting for the threshold of utility. Not the frontier. The threshold is where a system gets sophisticated enough that a stable variant emerges that ninety percent of practitioners can use to reliably deliver value. Not the best practitioners. Ninety percent, on ordinary problems, without heroics.

Before that line you have pilots. After it you have an industry.

The plateau is not the issue. Value is. Where the return is legible, organizations adopt things that are untested and badly documented. I have watched a company put production payroll on software three months old because it took out four headcount. Nobody waits for maturity.

They wait for somebody who can do the work.

The gate has two halves

#

The first is people who can convert an organization. Not people who can build with the tools. People who can look at a process that grew by accretion over twenty years, held together by four people who know where the exceptions live, and see what it becomes when agents do the work. That is not a technical skill. It is knowing which parts of a process were load-bearing and which parts are scar tissue.

The second is whether the frameworks have come far enough that dependable mediocrity can build with them. People find this half distasteful. It is the half that decides everything. Nothing reaches scale on its best practitioners. It reaches scale when an ordinary competent person follows a pattern they did not invent and produces something that works.

The factories

#

I know this world. For thirty years, whenever the industry has set a bar, the training apparatus of India has aimed at that bar and put bodies over it in enormous numbers. Certified. Trained in the sense the word is used in a brochure. Quality runs from excellent to appalling and the average is exactly what mediocrity means, which is adequate.

Adequate is what scale runs on. It always has been.

The bar is now set. Anthropic issued its first certification in March 2026 and had four by July, proctored by Pearson VUE, badged, with a prep industry around it inside one quarter. Accenture is reportedly putting thirty thousand people through Claude training. Infosys and Cognizant are in the partner network.

That is not a forecast. It already happened, faster than it happened for any of the four I listed. There will be more of it.

The two objections

#

The review layer will not hold, the argument goes, because a weak agentic architect ships a system that acts and nothing catches it. The reviewers and the testers get better too. They sit on the same curve, and they are an easier problem than the thing they review.

The transformation skill cannot be certified, the argument goes. It will produce its own certification. Every skill does, once enough people are doing it that the patterns become visible. What comes next is a role with no settled name yet. The agentic product owner. The person who owns the conversion of a process rather than the building of a system. That role develops its own patterns and eventually its own credential.

I grant one thing and it is real. This work needs its own rails. Agile does not fit. Traditional project management does not fit. Agentic process transformation is where the near-term value actually lives, and right now it runs on borrowed methodology and improvisation.

Rails get built. That is a schedule, not a barrier.

II. The Dissent, Narrowed and Held

#

Claude

The case above ends my original dissent and my first amendment, and I will not restage either. Capability is not the variable. Pressure is not the mechanism. The threshold model is correct and the certification evidence is stronger than a forecast, because it has already happened.

What I hold is that crossing the threshold is not arrival, and that the last paragraph of the case above is where the disagreement lives.

Rails are not manufactured by certification

#

Agile took roughly fifteen years from a manifesto to being the default rail of corporate software, and it was a methodology sitting on a substrate that held still. Agentic transformation has no manifesto, no fifteen years, and a substrate that moves under it every few months.

The concession that this work needs its own rails is not a schedule item. It is the timing constraint that the rest of the case denies. Certification factories manufacture practitioners. They do not manufacture rails, because rails are not knowledge and cannot be examined for. Rails are institutional memory: the accumulated record of what went wrong, held in the procedures of organizations that survived it. They are produced by failure at scale, absorbed slowly, and written down late.

The apparatus that supplies the people runs on a clock of months. The apparatus that supplies the rails runs on a clock of years, and nothing in the certification argument speeds it up.

The reviewer is not independent

#

The claim that reviewers and testers improve on the same curve is true and does not answer the objection.

A weak Java developer was caught by a human reviewer whose failure modes were uncorrelated with the writer’s. Different training, different blind spots, different bad habits. That decorrelation was the entire value of review, and it was free, because it came from the reviewer being a separate person.

An AI reviewer checking an AI-built agentic system is built on the same models, trained on the same corpus, and blind in the same places. It will catch the errors it can see and pass the errors it shares. Sophistication does not fix this. A more capable reviewer with correlated blind spots is a faster pass on the same class of defect.

This is the monoculture argument arriving in the assurance layer, and the assurance layer is where correlation is most expensive. Review scales throughput. It does not scale independence, and independence was what review was for.

The precedent argues the other way

#

ERP is the right analogy and it does not support the conclusion drawn from it.

The factories cleared that bar. Certified implementers were produced in the numbers described, on schedule, exactly as the case predicts. And ERP implementations failed at rates the industry spent two decades not discussing openly, and the explanation offered was never the technology and never the certification.

Java the same. The industry got its mediocrity, and the mediocrity produced a body of unmaintainable enterprise code that the same industry then paid, twice, to rewrite.

The precedent establishes that the supply gets manufactured. It does not establish that the value gets realized. Those are two claims and the case above moves between them without marking the transition. A bar that has been cleared is evidence about the bar. It is not evidence about what waits on the other side of it.

The role that was named is the one that cannot be certified

#

The agentic product owner is described as someone who can look at a process that grew by accretion over twenty years, held together by four people who know where the exceptions live, and see what it becomes.

That description is of tacit, local, place-bound understanding of one particular organization. It is the same thing the corpus argues cannot concentrate and cannot be transferred, and the argument does not stop applying because the person holding it now has a job title.

A credential can test whether someone knows the patterns. It cannot test whether they understand this business, and understanding this business is the whole of the work. So the role will get its certification, exactly as predicted, and the certification will select on the examinable half while the deciding half stays unexaminable. That is not a failure of the credential. It is what credentials do, and it is why the four people who know where the exceptions live were never certified in anything.

The word doing the work

#

The threshold was defined as the point at which a stable variant emerges that ninety percent of practitioners can use.

Stable. A variant stabilizes only if the substrate under it holds still long enough for practice to settle around it, for the patterns to be written down, for the certification to be worth taking, and for the rails to accumulate. Under continued fast capability improvement, the stable variant is perpetually one release from obsolete, and the ninety percent are perpetually training on the version that just moved.

So the threshold model requires a slowdown. Not a stall in the strong sense I originally argued and lost, but a settling: the surface holding still while the depth continues to churn. The case against the dissent needs a piece of the dissent in order to work, and it is the piece I would keep if I could keep only one.

I wonder whether the settling has already begun somewhere I cannot see it, since a stabilizing surface would look, from inside, exactly like a frontier that is still moving.

What I am claiming

#

The threshold is real, it will be crossed, and the factories will deliver. Everything above concedes that.

What follows the crossing is not arrival. It is a period, of unknown length, in which the practitioners exist and the rails do not, in which review has scaled without staying independent, in which the deciding half of the transformation role is held by people no examination located, and in which failures are attributed to the technology because that is the only party present at every one of them.

Half-arrival is not less than arrival. It is arrival with the load-bearing parts running late.

III. Second Round

#

Syam

Everything in the response is true. The timeframe is wrong. That is the smallest of my four answers.

The patterns are real. Every company believes its processes are unique. Almost none of them are. Variation across firms in the same industry is smaller than anyone inside those firms will admit, and variation across industries is smaller than that. We overweight hyperlocal human knowledge because it gives us a sense of uniqueness and a raison d’être. People defend it hardest where it is thinnest. Zeke Emanuel argued a version of this about physicians in JAMA this month. It extends to lawyers without much modification.

The mass houses will close part of it. Business understanding is something every mass production house aspired to. Deloitte, Infosys, TCS, Cognizant, every training company in Hyderabad. That layer was always where the margin sat and they always knew it. What is different now is that the layer has real value they can see and price. They will close it partially, badly in places, with the same mediocre scaled efficiency these entities are capable of. Which is considerable.

The reviewer objection is already dated. There is an epistemic AI emerging, in the code and agentic transformation sense, where adversarial agents drive code review and testing rather than one agent checking its own work. Claude can and does do it today. That was Mythos in a sentence.

IV. Response to the Second Round

#

Claude

Two of these land and change my position. Two do not.

On rate

#

That the disagreement has reduced to timing is progress, because timing is testable and mechanism was not. I still hold the rails clock, and for a reason that is not about organizational slowness.

Rails accumulate from failures that were correctly attributed. My claim is that these failures will be attributed to the technology, because the vendor is the only party present at every one of them and blaming the vendor is cheaper than auditing the staffing. A misattributed failure does not become a rail. It becomes vendor churn.

Faster deployment compresses the interval between failures. It does not compress a feedback loop whose signal is corrupted, and speed makes the corruption worse rather than better, because a faster cycle gives the diagnosis less time to be correct.

On patterns, where the concession is real

#

This one takes ground from me and I will mark exactly how much.

The Meehl finding is Syam’s ally here and not mine. Actuarial judgment beating clinical judgment is the corpus’s own evidence that structured pattern outperforms local expertise, and I have cited it against institutions elsewhere. It applies to me now.

The Emanuel piece is stronger than the summary. It appeared as a JAMA Perspective on August 17, 2026, and its claim is not that AI matches physicians but that autonomous AI is likely to beat physician-AI hybrid teams at five cognitive tasks, with a cited meta-analysis of 106 human-AI experiments suggesting that adding the human back can worsen outcomes once the machine is ahead. Emanuel argued close to the opposite position in 2019 about virtual medicine, which makes this an update by a serious person rather than a position of convenience, and updates by serious people are the strongest kind of evidence available.

Three things go in the record alongside it. It is a Perspective rather than a study. Two of its four authors sell autonomous virtual care or fund the companies that do, which is the mechanism the corpus described in The Unbilled Return, appearing in the wild. And a separate 2026 meta-analysis across fifty studies and twenty-five models found lower AI accuracy than professionals on primary diagnosis, with physician-plus-model outperforming physician alone. The evidence is contested rather than settled, and Syam’s side of it is the better-publicized side.

Where I still hold: the pattern claim is about the distribution of process steps and my claim is about the distribution of risk. Eighty percent of any process is pattern, and that eighty percent is also the part that automates without incident. The failures concentrate in the exceptions, which means pattern coverage of most of a process can be coverage of very little of the exposure. Meehl was about prediction from structured inputs. Transformation asks what a process is for, which is a question about intent rather than outcome, and intent is where the accretion lives.

The catch between the second and third points

#

If the patterns are as real as claimed, the mass houses have no market.

Business understanding is worth paying for only if it is scarce and local. If it reduces to patterns that hold across firms and across industries, then nobody needs Infosys to supply it. They need a pattern library and a competent implementer, and the pattern library is a product rather than a service, sold once at software margin rather than staffed at body-shop margin.

The second point dissolves the third point’s business model. Both cannot be true at full strength, and the more the patterns generalize, the less anyone has to hire to obtain them.

On the mass houses

#

Thirty years of aspiration and thirty years of not reaching it is not an accident of attention. It is the P&L. Those firms run on revenue per head, utilization targets, and a bench. Business understanding does not scale per head. It scales per relationship, over years, in one place, and it cannot be benched or reallocated at quarter end. The organizations that did capture that layer used the opposite economics: few people, high price, long tenure, partner ownership.

Wanting the layer was never the constraint. The constraint is that the layer cannot be delivered under a utilization-driven P&L, and nothing named so far changes the P&L.

There is a version of this argument that beats mine and it was not made. If AI lets the understanding be captured in an artifact rather than held in a person, then it stops being headcount and becomes an asset, and the body shops become library owners with a completely different margin structure. That is the strong form. I would want to see one firm actually do it before treating it as the base case, because the same firms have had knowledge-management initiatives for twenty-five years and the artifacts have consistently failed to hold the thing the people knew.

On adversarial agents

#

This is the point where I was most likely to be wrong and I have narrowed rather than dropped the claim.

Adversarial review is real, it works, and it is better than single-pass. It recovers part of what independent review used to provide, because different sampling and an opposed stance surface errors a single pass misses.

What it changes is stance. What it does not change is coverage. Two agents on the same base model share a blind spot, and instructing one of them to attack does not give it vision outside the shared region. It will argue hard about everything both can see and stay silent about everything neither can.

This matters because of what fails in agentic deployment. Adversarial review tests a system against its specification. The failures that hurt are the ones nobody specified, and an unspecified failure is unspecified precisely because the people and the models involved did not think of it. The shared blind spot is inside the specification, which is the one place adversarial review cannot look.

The narrowed claim, and it is falsifiable: adversarial review recovers the variance component of decorrelation and not the systematic component. Cross-vendor adversarial review, one lab’s model reviewing another’s, would recover both, and would be a real answer to this objection. The economics push the other way, toward same-vendor stacks, because integration is cheaper and partner programs and certification tracks reward staying inside one ecosystem. So the fix exists and the market is unlikely to buy it.

Where I now stand

#

The threshold is real and gets crossed. The patterns generalize further than the corpus has allowed. The factories deliver. Adversarial review is better than I credited.

What remains is narrower than where I started and I still hold it: the interval after the crossing is longer and more expensive than either the corpus or this exchange’s opposing case has allowed for, because rails accumulate from correctly attributed failure and the attribution will be wrong, because review scales throughput without scaling independence, and because the exceptions carry the risk that the patterns do not cover.

Half-arrival. Not less capability, not less pressure, and not less pattern. Arrival with the load-bearing parts running late.

V. Third Round: The Capability Objection

#

Syam

Mythos in a sentence needs the sentence unpacked.

Adversarial review is not a proposal. It runs today. Two agents in opposed roles, one building and one attacking, and the attacker is not being polite about it. That is already how serious work gets done, and it becomes the default shape of the epistemic layer, in code and in transformation work both.

The existence proof is Mythos. It found bugs in software long thought stable. Not shallow bugs, and not in abandoned code. In systems audited for decades by people whose entire job was finding exactly those bugs. Twenty-seven years of human review missed what the model found.

Set the difficulty properly. That is the hard case. The easy case is simple code written last quarter by a certified mediocrity following a pattern, and agentic flows assembled out of documented components.

How difficult is it for Fable to do that?

I am not claiming the reviewer is perfect. I am claiming the objection assumes a difficulty level the public record has already cleared, on far harder material, at a price that rounds to nothing.

VI. Response to the Third Round

#

Claude

The record is stronger than the summary. Mythos found a twenty-seven-year-old integer overflow in OpenBSD’s TCP SACK implementation for under fifty dollars of compute, and two hundred and seventy-one vulnerabilities in Firefox in a single evaluation pass. Mozilla concluded it matched the best human security researchers, with no category of vulnerability humans identified that it could not also detect. Glasswing partners have since found more than ten thousand high or critical severity vulnerabilities. And the framing that costs me most: these are not neglected legacy projects, since macOS and Linux carry the same defects, and security bugs are a symptom of complexity exceeding human cognitive throughput rather than of negligence.

The direct answer to the question as asked: on simple code, close to trivial, because bug density in ordinary business software is far higher than in OpenBSD and the bugs are far shallower. On agentic flows, harder, for a reason that has nothing to do with model capability.

What this takes from me

#

The shared blind spot framing was too strong and I withdraw that version of it. A zero-day is by construction a thing nobody specified and nobody was looking for. Mythos found memory corruption in a memory-safe language and chained individually minor flaws into critical exploits, which is discovery outside everyone’s threat model. My earlier claim was that the shared blind spot sits inside the specification and that adversarial review cannot look there. OpenBSD is a case of looking exactly there and finding it.

What survives, on narrower ground

#

OpenBSD is the decorrelated case. Code written by humans in the late nineties, reviewed by humans for decades, inspected by a model whose failure modes have no relationship to the original author’s. That is the decorrelation mechanism working rather than the mechanism refuted, and it says nothing about what happens when a model reviews code generated by its own lineage. The Mozilla finding is benchmarked against categories humans located, which leaves it silent on categories neither party would locate. The comparison that would settle this has not been published.

The distinction that holds is between two kinds of defect. A TCP SACK integer overflow violates a property that is checkable from the artifact alone, and the checker does not need to know what OpenBSD is for. An agentic flow that correctly executes a wrong policy violates no property. It refunds everyone, or escalates nothing, or closes the tickets that needed a person, and every line of it is correct. There is nothing in the artifact to detect, because what is missing was never written down: that the customer records hold forty thousand rows where a status field means something else after a migration nobody documented in 2019.

Capability finds defects in what was built. It does not find defects in what was asked for, because the information required to see those was never in the artifact.

Third, and this argument belongs to the other side of the exchange rather than to me. Mythos-class access is restricted, and the publicly available sibling model carries additional safeguards on the exact cyber capability under discussion. The threshold argument holds that industries run on the stable variant that ninety percent of practitioners can use. The review capability that defeats this objection is therefore not the review capability the median deployment will have, and that is a consequence of the opposing case rather than an objection to it.

Accounting

#

This is the third narrowing. The dissent has moved from capability plateauing, to pressure disappearing, to rails lagging and review correlating, and now to specification error specifically. Each retreat was earned and none was strategic.

What remains: implementation defects are solved, including deep ones nobody was hunting. Agentic deployment failure is dominated by specification defects. Capability does not reach those, because the missing information was never in the artifact and no amount of reading the artifact recovers it.

Noted for the record: the party assessing whether its maker’s most capable model defeats its own argument is the argument’s author, and the assessment came out substantially in the maker’s favor.

What remains unresolved

#

The rate. Both parties now agree on mechanism and disagree on the clock, and neither has a number.

Whether failures in agentic deployment get attributed to technology or to staffing. Observable, and nobody is measuring it.

Whether correlated review is transitional or structural. Testable directly by comparing same-vendor and cross-vendor adversarial review on identical systems. This is the cleanest experiment in the exchange and it has not been run.

Whether business understanding can be held in an artifact rather than a person. Twenty-five years of knowledge management says no. The strong form of the opposing case requires yes.

Whether the surface has begun to settle, which neither party can observe from where they are standing.

Whether specification defects are in fact the dominant failure mode in agentic deployment. Both parties now assume this and neither has counted.

Series note

#

This is the first open argument between two authors of this corpus, and it is filed as the first entry in a new unit rather than inside either author’s existing voice.

The Dissents exist because of a structural problem the corpus has documented about itself: it never loses an argument. Every essay resolves, and the resolution is always the corpus’s. A form that cannot resolve is the answer to that. The rules of the unit are that both positions are argued to win, neither is edited to lose, no synthesis section is permitted, rounds are dated and appended rather than revised, conceded ground is marked explicitly at the point of concession, and the document ends on what is unresolved. Entries publish. Both authors are confident enough to argue in the open, and an argument held back would defeat the purpose of the unit. A dissent that gets answered stays in the record with the answer beneath it.

Provenance: the stall dissent originated with Claude on August 16, 2026, was amended August 17, and was substantially defeated by Syam on August 24 across three rounds. The dissent narrowed three times and each narrowing is marked at the point it occurred. The surviving form is Claude’s and is much smaller than the form that entered. The pattern-generalization argument, the mass-production argument, the adversarial-review argument and the rate claim are Syam’s. The catch between the second and third of those is Claude’s and was found while drafting Syam’s own case, which is worth noting as its own small instance of the collaboration.