Synthetic Work

For the things we have to learn before we can do them, we learn by doing them.

§1. The Investigation

It is round three of an audit, and Elena Reyes is staring at a wire transfer she does not believe.

She is a first-year associate at a firm in San Francisco, six months out of school, on a forensic engagement that began as a routine quarterly review and turned into something else two rounds ago. The client is a midsized logistics company. The transfer is for $84,000, dated a Friday afternoon in February, sent to a vendor whose name does not appear in the master ledger Elena pulled in round one. She has already requested the original purchase order. The system told her the document is in the controller’s email and she will need to ask him for it. She has interviewed the AP clerk, who was nervous in a way Elena could not place. She has emailed the controller twice and received one reply, polite and unhelpful. The coach in the corner of her screen has not told her what the transfer means. He has asked her, twice now, what she thinks the AP clerk was nervous about.

She is going to have to make a call. The next move is hers. Whatever she does — pull the controller into a meeting, request his calendar for that Friday afternoon, walk over to the AP clerk’s desk again with a different question, escalate to the partner — the world is going to react to it. The controller has a history Elena can ask about. The clerk has a relationship with the controller Elena does not yet understand. The partner has expectations Elena has been managing since round one. Whatever Elena does, the people in this case are going to remember it, and round four is going to look different because she did it.

This is the kind of world the field has been trying to build for twenty-five years. We have known it was the right answer since the late 1980s. We have written about it, designed for it, and shipped fragments of it under a dozen names. We have never been able to build the whole. The thing Elena is standing in is the thing we have always been trying to make and have never quite been able to afford. That is the subject of this paper.

§2. The fragment we could afford

What we have been able to ship, instead, is the fragment. The fragment goes by many names — branching simulations, immersive learning, serious games, virtual practice environments, scenario-based eLearning — and the names are different because the field keeps reaching for one that finally captures what it has been trying to do. The architecture underneath the names is the same every time. A constructed world. A cast of characters. Situations the learner navigates. Consequences that follow from decisions. Feedback when things go wrong. The architecture is the same because it is the only honest answer to the question situated-learning theorists have posed since the late 1980s. If you take the theory seriously, this is what you build, and the field has been taking the theory seriously for twenty-five years.

And the field has built it. Virtual restaurants for shift management and customer service. Immersive simulations for clinical procedure and patient communication. Branching scenarios for B2B sales, negotiation, leadership conversations. Full-day exercises for crisis response and incident command. Tabletop walkthroughs for cybersecurity, regulatory inspection, executive crisis. Where the design was rigorous, the results outperformed content-based training reliably enough that the methodology has persisted across decades and across every wave of platform churn — desktop CD-ROMs to web players to mobile to VR to AI. The question is not whether scenario-based learning works. It does. The question is why, after twenty-five years of evidence that it works, it has not displaced the course library as the default.

The answer is in the budget. Every additional branch costs money. Every additional feedback variant costs money. Every additional character reaction, every additional scene the learner can enter, every additional consequence that can play out — each of these is an authored asset, and each of them is paid for. You make the world as rich as the budget allows and then you ship. Get the design right and the program outperforms whatever course library it replaced; get the budget right and you can build one of these for the role that matters most. Build it for every role and every skill that matters in an enterprise of any size, at the depth the methodology actually demands, and the math collapses. The methodology has been right for twenty-five years. The economics have been wrong for twenty-five years. That is the whole story.

§3. Two learners at the counter

To see what the budget actually buys, watch two learners run the same scene.

Take a virtual restaurant of the kind we have been building for years. The learner is playing assistant manager. A regular comes to the counter — three times a week, knows the staff by name, visibly stressed for the past month. Her order is wrong for the second time this week. She is not yelling, but she is at the edge of it.

The scene plays out, and the system presents four options:

A. Apologize and offer a free replacement meal. B. Apologize and offer a refund. C. Apologize and offer a discount on a future visit. D. Explain that the kitchen has been short-staffed and ask for her patience.

Maya is paying attention. By the time the menu comes up, she has formed a different read of the situation. The mistake is not the order. The mistake is that nobody has acknowledged the customer as a person yet. What this regular needs is for someone to come around the counter, ask if she is okay, and treat her like a human being for ten seconds before they start writing tickets. The free meal is fine. The refund is fine. But the right move is upstream of either. The menu has no slot for that move, so Maya picks A, because A is the closest thing on the list. The scenario logs a correct answer.

James has formed a different read too, and his is worse. Second complaint in a week, same customer, he has seen this pattern before — regulars sometimes work the system. His instinct is to push back gently, explain the kitchen pressure, see if she really wants to make a thing of it. He picks D. The scenario flags it as wrong and tells him a free replacement is the right call when a regular’s order has been wrong twice.

Both learners exit the scenario with the right answer logged. Neither has done anything to their actual model of the situation. Maya’s better read never got expressed, never got tested, and over many such moments she stops generating it. James’s worse read got contradicted at the answer level — the menu told him D was wrong — but his underlying belief that this customer is probably running a hustle is exactly where it was when he started. Tomorrow, on a real shift, Maya will reach for the transactional move because that is what she has practiced. James will reach for the suspicion because nothing in the scenario ever made him sit with the consequences of being suspicious.

What this scene shows is two ceilings working at once. Maya runs into the first one: the move she actually wanted to make was not on the menu, so it never got tested, and over many such moments her better instinct atrophies into the habit of picking the closest option. James runs into the second one: the feedback he received was perfectly correct at the answer level and did nothing to his actual model, because the system had no way to know that the model behind his answer was the part that needed correcting. The system saw a wrong selection and corrected it. The belief that produced the wrong selection went home with him.

The deeper failure is the one underneath both cases, and to name it we have to take a brief detour through how learning actually happens. Roger Schank made the case for it in Dynamic Memory (1982), and the cognitive science community has been building on it ever since. The cycle goes like this. A person encounters a situation. The mind reaches for an expectation — a prediction about how this kind of situation usually goes. The expectation produces an action: the person commits to a move that will work if the expectation is right. Reality then responds. If reality matches the expectation, the expectation gets quietly reinforced. If reality contradicts the expectation, it does not. The mismatch — Schank calls it expectation failure — is what forces the model in the person’s head to revise itself. No mismatch, no revision. No commitment, no mismatch. Learning runs on prediction.

Now look at what the menu does to the cycle. The learner walks into the scene and starts forming an expectation — Maya’s is this customer needs to be acknowledged before transacted with, James’s is regulars sometimes work the system. Then the menu appears, and the cycle stops. The learner does not commit to her own expectation; she scans four pre-authored expectations and picks the closest one. Reality, in turn, does not respond to her expectation — it responds to whichever option she selected. Maya never gets to find out what would have happened if she came around the counter. James never gets to find out what happens when his suspicion makes the regular escalate. The expectation each of them brought into the scene stays inside their head, untested. The system has no way to see the expectation, because the menu sat between the expectation and the action and replaced one with the other. There is no expectation failure because there was no expectation — only recognition. Which of these four did the author intend? That is a multiple-choice question in the strict cognitive sense, and a multiple-choice question is the wrong instrument for changing what someone believes about a customer at the edge of yelling.

This is the methodology defeating itself. The whole point of putting the learner in the situation was to produce the prediction-meets-reality moment that would update the model. The menu prevents the prediction. What the budget bought us is a scenario architecture whose central mechanism — pre-authored options chosen from a list — eliminates the cognitive event the architecture was supposed to deliver. The world Maya and James walked into looked like the world of practice. It was, in fact, the world of recognition. They walked out the way they walked in.

§4. Until now

Until now, the deal was this. You designed a world, and then you spent the budget enumerating it. You wrote out every move the learner could make, every reaction the world would offer back, and every line of feedback the coach would deliver. You accepted that the enumeration would be the ceiling on what the learner could practice and the ceiling on what the system could coach against. That deal has changed.

What changed is that the assets the world is made of no longer have to be authored one at a time. A generative system can produce a character who responds in voice and stays in character across a conversation that takes any shape the learner chooses to give it. It can produce a document the learner asks for, with the right form and the right details, that did not exist before the learner asked. It can produce the kitchen’s reaction when Maya comes around the counter, the regular’s reaction when James pushes back, the controller’s calendar when Elena requests it, and a coach who watches all of it and intervenes when intervening is the right move. None of these assets had to be enumerated in advance. The system knows the world well enough to generate them when the learner’s move calls for them.

This means the menu is no longer the ceiling on what the world can produce, and the pre-authored response is no longer the ceiling on what the system can coach against. The learner’s own expectation can be the action. Reality can respond to that expectation, not to whichever option was closest. Expectation failure becomes available — not as a feature the designer chose to build, but as a property of a world that can answer back. The whole architecture changes shape on the other side of the change.

§5. Real, open, persistent

Run Maya through the new world and watch the cycle run end to end. She walks up to the counter, sees the regular at the edge of yelling, and forms her actual expectation: this woman needs to be acknowledged before transacted with. There is no menu. She comes around the counter, asks the regular if she is okay, gives her ten seconds of human attention before mentioning the order. The regular’s face changes. She exhales. She says her son has been in the hospital for a month. The line behind her keeps moving, because the staff know what to do when the assistant manager steps away. The kitchen finds the original ticket. Maya makes the meal right. The coach says nothing during the scene, because Maya is doing the right thing, and intervening on a learner who is doing the right thing is how you teach learners to second-guess themselves. After the regular leaves, the coach asks Maya what she thought was happening, why she made the move she made, and whether she would have made the same move with a customer she did not recognize. Maya now has language for an instinct she previously could only act on. The model in her head, on this kind of customer, on this kind of stress, on this kind of moment — has moved.

Run James through the same scene and the cycle runs differently. He walks up to the counter, sees the regular for the second time this week, and forms his actual expectation: she is working the system. He says, politely, that the kitchen has been short-staffed and asks for her patience. The regular’s face hardens. She says she has been coming here three times a week for four years and the staff know her by name and her son is in the hospital and she does not need a lecture about kitchen pressure. She walks out. The line goes quiet. The shift manager comes over and asks James what just happened. James explains his read. The coach steps in and asks James to walk through what he saw that made him reach for the suspicion — the way the customer was standing, the second complaint in a week, the pattern he thought he was seeing. The coach does not tell James he was wrong. The coach asks James to consider what he did not see. The expectation James brought into the scene is now visible to him, and to the system, and to the coach. The cost of acting on it is the regular walking out. The model in James’s head, on this kind of read, has been opened up for revision.

Three architectural commitments make those scenes possible, and the field has been reaching for all three for twenty-five years. The world has to be real — consequences flow with weight, the regular’s exit is a real exit, the line behind her is a real line, the next customer is a real next customer who saw what just happened. The world has to be open — Maya’s move and James’s move both have to be admissible, even though no author enumerated either of them in advance. And the world has to be persistent — round two starts where round one ended, the regular remembers James, the staff remembers Maya, the shift remembers the moment when the line went quiet. Realness, openness, persistence. The methodology has demanded all three since the late 1980s. The economics permitted at most a fraction of one.

What realness actually requires is that the world’s response to a learner’s move be generated, not retrieved. The regular’s exit is not a pre-authored bad ending; it is what happens when a customer at the edge of yelling encounters the wrong move from the assistant manager. A different move — Maya’s, or a third learner’s, or a fourth — produces a different response, also generated, also coherent. The weight in the consequence comes from the fact that the world is responding to this move, not picking from a list of canned outcomes. When the regular walks out, the cost is real because the move is real. The learner cannot dismiss the consequence as a thing the author decided to throw at them.

What openness actually requires is that the world admit moves the author did not enumerate. Maya coming around the counter was not on anybody’s list of options. The system did not have a branch labeled come around the counter. It had a model of the world rich enough to answer the question what happens if the assistant manager comes around the counter and gives the customer ten seconds of human attention. That is a different thing. The first is a tree of designed paths. The second is a world that can be acted upon. A learner whose better instinct is to do something the designer never imagined is, in an open world, finally able to test that instinct. Whatever happens next, the instinct has been on trial.

What persistence actually requires is that the world remember. James leaves the shift, comes back tomorrow, and the regular is not there — but the staff remember the moment when the line went quiet, and the shift manager has views about it, and the next time James reaches for a suspicion read he does so under the eye of people who saw what happened last time. Round two is shaped by round one. Stakes accumulate. A regular who walks out today is a regular who is not there tomorrow, and tomorrow’s shift is the world without her. Persistence is what turns a scene into an arc, and an arc is what lets judgment-under-pressure be rehearsed at all, because judgment-under-pressure is the thing that compounds across a career and cannot be exercised in a single decision.

These three together are the architecture. None of them was technically impossible before generative AI; all of them were economically impossible at the depth the methodology demands. To author a world rich enough to be real, open, and persistent across the moves a learner might make and the rounds the world might have to hold — at the depth a serious training program needs and across the breadth of roles an enterprise actually employs — required authoring budgets nobody has. The assets that constitute realness, openness, and persistence are precisely the assets that no longer have to be authored one at a time. What changed is not that we found a new pedagogy. What changed is that the pedagogy we have always wanted is now within reach.

§6. What the substrate produces

To see what a real-open-persistent substrate produces in practice, look at two worlds we have actually been building. Neither one looks anything like a branching scenario, and neither one looks like the other — that is the point.

The first is a sales pitch world for a global infrastructure firm. The first-year consultant sits down to a video call with Tanaka, the chief risk officer of a regional bank. Tanaka is twenty-five years into financial services, conservative, skeptical of AI, exhausted by the regulatory pressure that has been climbing every quarter for three years. The consultant has been told the bank is a target and Tanaka is the relationship to win. There is no script. The consultant has to open the call, find Tanaka’s actual concern, build credibility without overpromising, and earn a second meeting. Tanaka responds in voice — sometimes warmly, sometimes guarded, sometimes with a question that exposes a claim the consultant did not have evidence for. The video avatar carries the affect; the consultant can see Tanaka’s posture shift when a particular line lands. The coach is in a sidebar. The coach intervenes when intervening would not interrupt the relationship being built — which is rarely during the call and often after it.

The first call with Tanaka is one call in a pursuit. The pursuit lasts three or four meetings before it resolves — the bank signs, the bank passes, or the deal stalls long enough to be paused — and in every meeting Tanaka remembers the previous one. If the consultant overpromised in meeting one, Tanaka in meeting three is harder to win back. If the consultant under-asked in meeting one, Tanaka in meeting three has stopped expecting the consultant to lead a real discovery. The arc is the unit, not the call. When the Tanaka pursuit ends, a different pursuit begins — a different CRO, at a different bank, in a different industry, with a different pressure. Across an onboarding program, a first-year consultant runs twenty or thirty of these pursuits, each one its own arc, each one drawing on what the previous ones taught.

But Tanaka is not the only stakeholder at her bank, and the persistence inside her pursuit is more interesting than a single character with a memory. Real selling is multi-stakeholder, and so is the world. Between meeting one and meeting two, Tanaka talks to the CFO and to the head of compliance, and what she tells them is not a transcript of what the consultant said — it is Tanaka’s version of it, filtered through her own concerns and her own reading of the consultant’s posture. When the consultant walks into the CFO meeting, the CFO opens with what Tanaka relayed plus the CFO’s own concerns layered on top. If the consultant told Tanaka something that landed cleanly with her but read as evasive when she repeated it to the CFO, the CFO is already skeptical before the meeting starts, and the consultant has to figure out why. When Tanaka returns for meeting three, she remembers the consultant through her filter — what she paid attention to, what she did not, what she has come to believe in the days since. The consultant cannot count on having said something to one person and not having to live with how that person reported it to the next. That is what selling actually is, and that is the kind of persistence the world has to hold: not a perfect record, but a network of imperfect narrators each carrying their own version of what happened, the way an actual organization carries its own version of any consultant who has ever walked through it.

The shape of the program, then, is not a library of practice scenarios; it is one apprenticeship, lasting weeks, in which a first-year learns the job by doing the job in a world where the job has consequences and the consequences propagate through people who do not all see them the same way. That shape has a name. Allan Collins, John Seely Brown, and Susan Newman called it cognitive apprenticeship in 1989, by analogy to the craft apprenticeships that produced expertise in everything from blacksmithing to law before knowledge work made the master-and-apprentice arrangement economically infeasible. (Allan Collins the cognitive scientist is no relation to the co-author of this paper; the convergence of the names is awkward and we are stuck with it.) Cognitive apprenticeship was never a missing idea. It was a missing budget. It is the species of synthetic work designed for learners who have to acquire an entire role from scratch, and it is what we mean when we say virtual apprenticeship: the apprenticeship architecture, instantiated in a generated world, available to any learner who needs to learn a role end to end.

The second world is an investigation. The first-year audit associate — Elena, from the opening of this paper — is six months into an engagement that started as a routine quarterly and is now a forensic. The world contains characters: a controller, an AP clerk, a partner, a CFO who has not yet entered the case but will if Elena asks the right questions. The world contains documents: a master ledger, a wire transfer record, an email thread Elena does not yet have access to, a vendor file that does not exist in the system but does exist in the controller’s email. Elena requests documents and the world produces them, with the form and the detail an actual document would have. She interviews characters and they respond in voice, sometimes cooperative, sometimes evasive, sometimes giving her the wrong answer for reasons she will discover only if she follows the right thread. The same multi-narrator filter that runs the sales world runs this one too: the AP clerk has been talking to the controller between Elena’s interviews, and the controller, when Elena finally gets him on the phone, is the controller plus what the AP clerk told him about the questions Elena was asking. She decides whether to escalate, when to escalate, and to whom — and the partner above her has expectations Elena has been managing since round one. The simulation does not tell her what the wire transfer means. It puts her in the world where wire transfers like that one mean something specific, and lets her find out which specific thing this one means.

The audit world has a different shape from the sales world. There is no single conversation to win. There is an investigation that runs across rounds, with each round opening up new documents and new characters and new versions of the original question. The pacing is different. The stakes are different. The kind of judgment being rehearsed is different — Elena is rehearsing forensic reasoning under conditions of partial information and stakeholder pressure, where the consultant in the first world is rehearsing relationship-building under conditions of skepticism and time pressure. Same substrate. Different shape. Different role. Different kind of judgment.

This is what we mean by saying that the choice of shape is now an instructional design decision. Until the substrate existed, the shape of a synthetic-work environment was constrained by what the authoring budget could afford to enumerate, and the shapes that got built were the shapes whose enumeration cost was tractable — short scenes, branching trees, single decisions. Long-arc investigations were unaffordable. Multi-round relationship campaigns were unaffordable. World-with-memory was unaffordable. With the constraint lifted, the shape of the world becomes the design question — what kind of judgment does this role require, what conditions of partial information and consequence does the judgment compound under, and what shape of world rehearses that — rather than a budget question. Sterling Crest is one shape. The audit is another. Maya’s restaurant is a third. The substrate underneath all three is the same. The shape is chosen to fit the role.

A first-year consultant who has spent twelve weeks in a virtual apprenticeship that put her through thirty pursuits — different industries, different pressures, different cultural contexts, different kinds of skepticism, different stakeholders to win and lose — has practiced something a course about consultative selling cannot teach. A first-year audit associate who has spent sixteen weeks investigating eight engagements, each one branching from a routine quarterly into a forensic that demanded different kinds of forensic reasoning, has practiced something an audit textbook cannot teach. A new assistant manager who has spent six weeks running shifts in a restaurant whose regulars remember her, whose staff has views about her, and whose customers compound across days, has practiced something a service-recovery course cannot teach. The course was an artifact of authoring scarcity. The world is what we build now.

§7. What we can finally build

Sterling Crest and the audit are two specific worlds for two specific roles. The reach of the substrate is wider than that, and it is worth taking a moment to look at the shape of the problems it is now positioned to solve. The list is not exhaustive — every reader will think of additions, and the additions are part of the point.

Consider the question of how organizations will produce expertise in a labor market where entry-level knowledge work is being automated. Some observers describe what is emerging as a diamond-shaped organization — top-heavy with senior judgment, hollow in the middle, thin at the bottom. The shape produces a question nobody has had a good answer for: if AI does what new hires used to do, where do tomorrow’s senior people acquire the experience that makes them senior? Senior judgment was always built out of the accumulated weight of having done the junior work, with all its mistakes and corrections, under the eye of someone more experienced. Without the junior work, the apprenticeship breaks. Synthetic work does not put the entry-level jobs back, but it does something that may be more useful: it provides a place where the deliberate, supervised practice of junior judgment can happen at scale, in a world that responds to the practice with the same kind of consequence the actual job used to provide. The expertise that used to require five years of progressively harder work, learned by getting it wrong and being corrected, becomes available without the organization having to maintain those five years of work just to give the practice a place to happen.

Consider the related but distinct question of onboarding. Every organization that hires at scale has lived the same problem: a six-week or twelve-week onboarding program runs the new hire through the curriculum and lands them at the actual job, where they are not ready for it. The gap between the program and the work has been a constant complaint of every L&D function for as long as L&D functions have existed. The gap exists because the program teaches about the work, and the actual job is doing the work. Synthetic work closes the gap by making the onboarding program itself an accumulation of pursuits, engagements, and shifts at the depth and shape of the real role — so that on day one of the real job, the new hire has done some version of the work forty times and the failure modes are recognizable rather than novel. The point is not that onboarding gets shorter. The point is that the learner who exits onboarding has practiced the job rather than read about it.

Consider the question of what happens when senior practitioners leave and take their judgment with them. Every domain has knowledge that lives only in the heads of the people who have done the work for twenty years — the way an experienced underwriter senses a fraudulent claim, the way a veteran maintenance technician hears a bearing about to fail, the way a senior negotiator reads when the other side is bluffing. Knowledge management has tried for thirty years to capture this kind of expertise and has mostly produced documents nobody reads. Synthetic work captures it differently. An experienced practitioner thinks aloud through a real case. The world that produced the case gets reconstructed around the practitioner’s reasoning. A junior practitioner then runs the same case, again and again, until the senior’s recognition becomes the junior’s. The expertise does not get documented. It gets reproduced as a world the next generation can practice in.

Consider the problem of high-stakes events that happen too rarely to rehearse. Crisis response. Catastrophic failure. Hostile takeover. Major fraud discovered mid-quarter. The kind of moment that occurs in an organization once every four years and that you cannot afford for your people to be improvising on. Old approaches to this problem produced binders nobody opens until the moment, by which point the binder is the wrong artifact. Synthetic work produces something different: a crisis the leadership team has been through six times under varying conditions, with consequences that compounded across each run, so that when the real crisis arrives, it is the seventh run rather than the first. Rare events become rehearsable.

And consider the kind of compliance and regulatory training that has to produce judgment, as distinct from the kind whose real job is producing documentation. Financial decisions under capital adequacy regulations. Clinical decisions under patient safety standards. Security decisions under active threat. The real question in each of these is not whether the practitioner knows the rule but whether they will recognize the situation that calls for the rule, under conditions of pressure, ambiguity, and competing demands. Awareness training cannot teach that. What can teach it is the right scenario, generated under the conditions that produce the failure, rehearsed across enough variations that the recognition becomes reliable. That is what synthetic work delivers, and that is what the genuinely consequential parts of compliance training have always needed.

These are five problems. There are more. The question is not whether the substrate can be applied to a given organizational pain point; the question is whether the pain point is actually about producing judgment under conditions of partial information. When the answer is yes, the response is the same one in every case: build the world, and let the practice happen against it. We have spent twenty-five years watching the field reach for the methodology and lack the budget to deliver on it. The budget has shifted. The methodology is finally available at the breadth the field has always known it needed.

§8. What survives

It is worth being clear about what does not change. The methodologies the field has built over the last twenty-five years are not made obsolete by the substrate; they are finally able to be delivered at the depth they always demanded. Three in particular survive intact.

The first is the principled choice of which mistakes the world should be built around. The methodology has a name — Critical Mistake Analysis — and it has been the load-bearing input to every serious synthetic-work program our team has built. CMA starts from the failures: the patterns of error that show up in incident data, after-action reviews, call recordings, secret-shopper video, verbal protocols. It traces those failures to their actual root causes, which are almost never knowledge gaps but decision-making and judgment failures. And it ranks them by frequency, criticality, and trainability, producing the small set of mistakes that are worth building the world to produce. The substrate does not change any of that. The substrate makes it possible to build the world deeply enough that the mistakes get reproduced under the same conditions that produce them in real practice — which is what the methodology was always trying to deliver and never quite could. Without CMA upstream, the substrate just produces a vivid world that may or may not rehearse the things the role actually fails on. With CMA upstream, the substrate produces practice that targets the specific decision points where real practitioners go wrong.

The second is the design framework for what makes a synthetic-work environment ring true. NIIT has used a framework called E=MC5 — Education equals Mission, Context, Challenge, Choice, Consequence, and Coaching — and the framework names the components a synthetic world has to hold together to produce learning rather than entertainment. Mission gives the learner a reason the world matters. Context places the learner inside the situation rather than outside it. Challenge sets the pressure that forces a real expectation rather than a speculative one. Choice makes the expectation visible by requiring commitment. Consequence is the world responding with weight. Coaching is the layer that turns the consequence into revision. None of those components is supplied by the substrate; the substrate is the medium through which a designed program delivers them. A generative world without mission is a sandbox. A generative world without challenge is a chat. A generative world without coaching is an exercise the learner walks away from without revising what they came in believing. The frameworks the field has been using to design good scenario programs are the same frameworks needed to design good synthetic-work programs. They are the difference between a world that teaches and a world that merely runs.

The third is the architecture that gives the long-arc programs their name. Allan Collins and his coauthors did not invent cognitive apprenticeship in 1989 to solve a budget problem. They invented it because it was the architecture by which apprenticeship had always produced expertise, and they argued — correctly — that knowledge work needed it just as much as the trades did. The method survives. What changes is that an apprenticeship that previously required a master per learner, or a war-game program that required a year of authoring per role, is now buildable by a small team and runnable across an entire organization. The architecture was right. The deployment was rationed.

It is also worth being clear about what synthetic work does not replace. Knowledge transfer is still real work, and a procedural reference is still the right artifact when what the learner needs is a procedure. Regulatory awareness training, when its real job is producing documentation that the regulator wants to see, is a different thing from synthetic work and should be designed for what it actually has to do. So is the kind of compliance training whose real purpose is to inoculate the organization against liability. Synthetic work is the right answer for the part of training that has to produce judgment — for the moments where the learner has to commit to an expectation under conditions of partial information and live with the consequence. That is most of what people get paid to do in knowledge work, but it is not all of it, and a paper that argued otherwise would be doing the same thing the rest of the field has been doing for fifty years: claiming a single modality could solve every learning problem. The course library was wrong about that. We do not want to be wrong about it in the other direction.

A final note on what becomes free. There are properties of synthetic-work environments that used to be customization work — separate authoring runs, parallel investments, deliverables in their own right — and that are now simply properties of the substrate. Cultural localization is one. The sales world we described, instantiated in Japanese for Japanese learners with a CRO whose pressures are the Japanese regulatory environment, was not authored twice. The world is rich enough to be expressed in either language, with either set of cultural and business pressures, without the authoring team enumerating the variation a second time. Industry variation is another. Stakeholder variation is a third. What used to require a project plan now requires a prompt and a refinement loop. Things the field has been paying for as customization, the field will stop paying for, and the budget that used to go to enumeration can go to the design choices that actually matter — which mistakes to target, which conditions to rehearse them under, and which kinds of judgment the world has to produce.

§9. What comes next

Every organization that hires for judgment has been training people somewhere. The somewhere has not usually been a training program. It has been the actual job, with real customers and real clients and real patients absorbing the early errors of practitioners who had no other place to make them. We have called this on-the-job learning and it has been the dominant practice surface for every cognitive profession we have, because the methodology that should have replaced it could only ever be afforded in fragments. The cost of the fragmentation has been the practitioners’ early years, paid for by the people they served while they were still learning to serve.

That cost is no longer necessary. The substrate is here. The architecture has a name and a shape, and the organizations that are building against it now will produce a different kind of practitioner — one who arrives at the actual job already having made the mistakes that the job used to require for them to make. The economics of the field have shifted. The methodology can finally be deployed at the breadth it has always demanded.

The question, for you, is whether your people are going to acquire their judgment the way they have always acquired it — on the customers and clients and patients in front of them — or whether they are going to acquire it in a world built for the purpose. Both options are still on the table. They will not both stay there.

§10. References

Collins, A., Brown, J. S., & Newman, S. E. (1989). Cognitive apprenticeship: Teaching the crafts of reading, writing, and mathematics. In L. B. Resnick (Ed.), Knowing, learning, and instruction: Essays in honor of Robert Glaser (pp. 453–494). Lawrence Erlbaum.

Schank, R. C. (1982). Dynamic memory: A theory of reminding and learning in computers and people. Cambridge University Press.