The Taste Harness

Taste can't be briefed, only formed. How consequential attempts and diagnosed rejections build the context nobody could write down, and the three places that judgment has to live before it can travel.

Scott Farrell · LeverageAI · September 2026

A few days ago, near the end of a long week of building with AI, on the day I'd spent making short films, I asked the agent that had been supervising the work to write a page. Nothing elaborate: explain the films. Why they exist, what each one is for, how they connect to the people who'll watch them.

Holy dinkums, did it explain it well. It explained the branding and what each film was there to do far better than I could have, and I'm not bad at marketing and branding. The page opened on an idea none of us had written down in those words: "People don't arrive with dental words. They arrive with one tooth." It organised every film around the viewer's own concern and the viewer's own language. It gave each film one job in someone's decision. And it named the strongest device in the series, one that three separate builder agents had found on their own: the person's own words travel through the service.

That page was better than the brief I'd supplied before the work started. I wanted to know why.

The short answer is that the agent writing it wasn't a cleverer writer than the one I'd started with. It was the same kind of model. The difference was what it had lived through. It had watched builds fail. It had seen judges disagree about the same cut. It had heard me reject things. When I asked "why are we making these films?", it had a pile of specific contrasts to account for, and any answer that couldn't account for them was visibly wrong.

Do not expect the best explanation of a task from a system that has only been told about it. Give it access to consequential attempts, then ask it to account for what those attempts revealed.

That line came out of the conversation I had afterwards with my AI thinking partner, and I think it's the correct shape of an objection I've been making for years. I've always argued that you shouldn't use AI as an oracle: tell me the best thing to do, what's the best way to win this customer? Most people answer that objection with "give it more context", and that sort of makes sense. But the page didn't get better because it had more context. It got better because the work had kicked it, repeatedly, in the specific places that mattered.

This article is about that mechanism, and about turning it into something you can run on purpose. I call the result a taste harness: a way of running production so it forms the judgment nobody could brief, and a way of carrying that judgment so it survives into the next, unfamiliar job.

Context names what's available. Perturbation names how it discovers what matters.

Start with what a brief can actually do. A brand brief for a local health clinic might say: warm, credible, reassuring, recognisably local. Every word is right. And every word leaves a huge space of interpretations open. A voice can be warm and patronising. A film can be polished and emotionally remote. A reassuring delivery can be so flat that nothing registers. None of those failures contradicts the brief.

You only find where the words' edges are by making something and reacting to it. "I don't like that voice" starts as a preference. Explored, it can turn into a distinction: the problem isn't the voice's quality, it's that the voice puts the speaker in the wrong relationship with the audience. That distinction didn't exist in the brief. It couldn't have. Nobody knew it was needed until an attempt tripped over it.

The brief supplies concepts. The encounters give those concepts boundaries.

So I'd keep two words apart that most people collapse into one. Context names what the system has available. Perturbation names how it discovers what matters. Technically, everything the supervising agent learned ended up in its context window. But calling it all "context" hides the mechanism: some information was supplied at the start, and some could only be obtained by attempting something and meeting resistance.

I've written about perturbation before, as reading. In Cycle Compression I argued that a long conversation between me and a model that shares my vocabulary drifts toward self-agreement, and that web research is really an "external perturbation source" that kicks the loop sideways.¹ In The Inference Field I argued that two contexts of the same size can be nothing alike, because one was built as a pile of relevant chunks and the other as "a travelled argument", where each thing read changed what was worth reading next.² The films extend both from reading to making. The supervising agent's context was a travelled argument whose steps included building, listening, rejecting and repairing.

This isn't just my observation. Researchers at Berkeley watching people build evaluators for language-model output named the same thing "criteria drift": "users need criteria to grade outputs, but grading outputs helps users define criteria," which implies "it is impossible to completely determine evaluation criteria prior to human judging of LLM outputs."³ Donald Schön said it of architects decades earlier: the designer's intention "was not fully established at the beginning of her design process, but evolved through her appreciation of an intermediate design product. … An evolving intention is one of the outputs of her designing."⁴ Even Anthropic's guidance for writing agent skills tells you to iterate so you "discover what context Claude actually needs, instead of trying to anticipate it upfront."⁵

One qualification keeps this honest. Not every activity is perturbation. Ten near-identical versions, each greeted with "yeah, looks good", teach almost nothing. One version that looks excellent and still fails its purpose can expose an entire missing distinction. The productive ingredient isn't disturbance. It's contrast that changes the working understanding.

Taste is judgment about fit

Here's the case that made this concrete for me.

One of the clinic films was a short, beautiful piece meant to stop people scrolling: close-ups of the craft. Tooth-coloured material, shaping instruments, a polished surface. On its first cut, the judge that listened to the sound (a Gemini model, as configured that week) scored it 8.8 and called it "a masterclass in quiet, premium service advertising". A cold viewer gave it 8. The harsher judge, an Opus model reading keyframes and a transcript, gave it 6. Its reason: the craft shots never connected to anything a viewer actually notices about their own teeth, like a chipped or short edge.

Both judges were looking at the same film. They were answering different questions. The first was answering is this beautiful and on-palette? The second was answering does this connect to the person watching?

The criticism directed the repair. The first round tied each shot to something people notice, and it still wasn't enough: the tools were posed rather than working, so nothing actually changed on screen in a film whose whole premise was a close look at the work. The second round made the change happen. The material fills a chip on the edge of a tooth, and the surface turns glossy under the polisher. A third round cut a static end card. By the final cut, it cleared every judge.

That's why I now define taste, for this work, as judgment about fit: fit between expression, audience, situation and purpose. It's much richer than a catalogue of personal aesthetic preferences, and much more testable. And it tells you which meaning of "on brand" a harness is learning to check. There's the straightforward version: correct logo, palette, typography, approved wording. And there's the consequential one: does this particular experience express this company's promise to this particular audience?

Adobe's recent description of its Brand Intelligence product is a strikingly close parallel on the first half of this. It says the system learns from "'decision traces' … including comments, edits, rejected versions, and final approvals", and its validation skill takes in "intended audiences" and explains failures.⁶ Its launch post puts the problem well: brand decisions "are shaped by context, taste, and experience. Guidelines try to capture them, sometimes across hundreds of pages, but they never go far enough."⁷ That's the vendor's described architecture, not independent proof that it works. And it assumes an enterprise that already has a decision history to learn from. The business in my story didn't. Its standard had to be formed during production.

I'd put one boundary around all of this. In Hidden Gates I wrote that gates guard completeness and process, and "do not guard taste".⁸ I'm moving that line, but only partly. A harness can now operationalise some failures of fit as inspectable tests: does the craft connect to what the viewer notices? Is the voice cast from the audience's place? That's a repeatable way to detect certain failures. It isn't a taste oracle, and promotion of anything to "good" still sits with a person.

Formation, not transfer: "see how we did that"

What I was actually doing that week, I now realise, was using what I'd call the "see how we did that" strategy. Get the AI to help make one film first. Do the research, find the tools, let me review it by hand a few times and explain what I found. Then it builds better code, and it builds its own test harness as we go. I explain what I found in the film, and it updates the harness. By the end, we don't just have some source code that's an example of what we're doing. We have an increasingly competitive and accurate test harness. It contains accuracy and validity, obviously. But it's increasingly containing taste.

"See how we did that" is an apprenticeship instruction, not a recipe-extraction instruction. There are two very different things you can pull out of a successful session. One is a sequence: use these tools, write this script, render these scenes. The other is a body of judgment: what were we trying to achieve, which reasonable-looking choices failed, what did the human notice, which correction worked, and when should that lesson apply again?

The first reproduces a procedure. The second can support competent variation.

That's why the rejected versions matter so much. The final film contains the choices that survived. It can't, by itself, explain the alternatives they beat. The production history supplies that missing contrast. Learning science has said this for a long time. Schwartz and Bransford found that analysing contrasting cases creates "a 'time for telling'; learners are prepared to be told the significance of the distinctions they have discovered."⁹ The cognitive-apprenticeship literature says learners grasp "the boundary conditions and entailments" of what they know "only through encountering them in real problem solving."¹⁰

It's also why I'd say this is taste formation, not just taste transfer. Some of my taste existed before the session but was hard to articulate. Other distinctions only became clear through the collaboration. I'd recognise something wrong; the model would propose an explanation; a changed version gave me something more precise to judge. Neither of us held the fully articulated principle at the start. In Worldview Recursive Compression I described a generate, evaluate, extract, encode loop that improves a kernel after each output.¹¹ This is its missing upstream: you're not only extracting knowledge after the task. You're developing the knowledge needed to complete the task, while completing it.

Handles and the discriminations they point into

Formation leaves a particular residue: short phrases. Earlier in the work I kept using loose words for what I wanted. "Wow." "Oh, you see me." "Let the promise land." Once they'd been recorded in the prompting, they started steering builders, which meant they needed to mean something precise. So I asked the build agent to tighten them, without making them verbose, in a way that could feed the test harness.

What came back is the most portable thing the week produced:

The short phrase is the handle. The accumulated discrimination is the asset.

A handle without its discrimination is what I've elsewhere called a slogan that fails the family test: hand it to two builders and you get two unrelated films.¹² A handle that points into worked examples and a test is a different object. It can steer a builder and check a judge.

Three questions, and one rejection walked through all of them

If you only take one practical thing from this article, take this. When work is rejected, there are three different things you might be improving, and they answer three different questions. I'll call them the three questions of improvement, to keep them separate from the three questions I've used elsewhere for extracting lessons after a ticket.

The interesting recursion is that a single rejection can move through all three. So before anyone fixes anything, ask the triage question:

My earlier routing, from Look Mum No Hands, sent a rejected AI proposal to the layer that could fix it: wrong data, wrong playbook, wrong context, or wrong template.¹³ Every one of those is a defect in how the artefact got made. Taste work needs two more layers: the instrument that failed to see the problem, and the principle nobody had stated yet.

The worked case: an opening I gave 6 out of 10

The fifth film in the clinic series explained why a plan looks at more than the one tooth you noticed. It was the first film built with the lessons harvested from the first four, and it cleared the judges' bar on its first build, which none of the first four had done. I thought it was fantastic. But my verdict on its two ends was specific:

"Even the opening scene could be better. I'd give it about 6 out of 10. It has a web page on an iPhone as the thing to connect on. It assumes they've seen the offer. If you met it separately, on Instagram say, it would seem disconnected." And then: "The rest of the video is awesome, explaining why you can't fix just one tooth."

Here's where that one rejection went.

Leg one, the artefact. The scene itself connected. The problem was an inset: a phone showing the offer page, laid over the opening. As I put it once I'd looked again, "the image connects, just not the iPhone/web page. The inset image." The revision replaced the inset with a photo message from his sister: the man laughing, his daughter on his shoulders, the kind of photo you zoom into to look at your own smile. The middle, which I'd called awesome, was left alone.

Leg two, the instrument. This is where it got interesting. The judges had half-caught my point at best. When the build agent looked into why, part of the answer was its own doing: the judge brief it had written for this film said "Where it's seen: retargeting", so the judges forgave an opening that assumed the viewer had already seen the offer. Then it tested the cold viewer. A viewer told it had already seen the ads passed both openings, and called the old opening's moment of recognition "specific". A truly cold viewer, one that had never heard of the clinic and was just scrolling a feed, was harder on the old opening, and in one run said my verdict back in its own words: "I'd never have been on a … offer page. I'd just be scrolling Instagram." On the new opening, one truly cold run said the thing I'd hoped a viewer would say: "instead of enjoying it you zoom in on your own mouth. I have absolutely done that." The runs were noisier than that, and the build agent's summary made them sound cleaner than they were. But they moved towards my verdict only once the viewer was truly cold. The instrument had been measuring the wrong viewer.

Leg three, the principle. The tall cut of the new opening had put a card over the man's head. Tracing that led back to the skill itself: its advice for tall cuts said "the card sits over the quiet part of the scene", which is exactly how the phone landed on his head. The principle was wrong, not just the film. And the recognition question in the harness had its own principle baked in, "someone like me", which invites a viewer to look for a look-alike instead of their own life.

Quoted from the brand-film skill and judge as committed that day. The judge brief's wording is from the development record of the same session.

One rejection. A better film, a harness that now catches most of what I saw (with its one remaining miss written down), and a principle stated precisely enough to change the next film on a different company. That's what I mean when I say the rejection is the unit of learning.

The same shape, with a voice

Earlier that day the same thing had happened with a voice. On a film for one of my platform's concept companies, I said the opening voice was "crappy compared to the rest, seems like older tech". The immediate fix was artefact-side: the engine's own authors graded that voice D+, so it was banned. But the diagnosis kept moving. The builder had auditioned voices by pitch and by whether the transcript came out clean, and "neither measure catches quality". A transcript-reading judge went on to prefer the rejected voice every time, because it was judging a voice from mis-transcribed text; a listening judge heard my verdict and named the problem as an "incongruous British accent". The sibling book, Code as Proxy, tells that middle part as a story about what judges can perceive.¹⁴ What matters here is where it ended up. For the clinic, the supervising agent cast local voices on its own initiative and let a listening judge pick them; the record says I hadn't yet heard them myself. The harvest then turned the whole chain into a principle for the general skill: cast the voice from the audience's place. The rejection had moved from quality, to perception, to relationship, across four separate events rather than one clean step. The portable lesson isn't "use an Australian voice". It's to consider where the audience is, who the speaker is meant to be, and what relationship the delivery should establish.

Three carriers that must reinforce without converging

So where does a lesson go once it's been triaged? Not into a longer prompt. A lesson worth keeping needs three homes, each doing a different job.

When I asked how to retrofit the Film E lesson into the prompting, the build agent laid out the practice this way: the skill carries the lesson as craft ("a principle, the real example and your words, and a test"); the judge rubric carries the same test, so a film that misses it gets named with a time; code measures it wherever it can be measured. "The briefs stay short north stars and never grow rules. Your verdicts are recorded as test labels, so each new judge can be checked against what you actually said." That last sentence matters. My earlier finding, as the build record put it back to me, was that prescriptive rules made results worse.

The three carriers have to reinforce each other without becoming identical. A skill needs room to create a new good answer; a harness shouldn't reject it just because it differs from the first successful example. In The Exemplar Ratchet I called this the difference between extracting the shape and extracting the wallpaper: "If they only win when the wallpaper matches, you extracted wallpaper."¹⁵ And contrasting cases are the carrier most teams forget. Positive examples alone tell a builder what one good answer looked like. Pairs tell it which difference mattered. Intuit's researchers found that giving a model positive and negative examples beats plain few-shot prompting for "a certain style or tone that is hard to describe."¹⁶

There's a failure mode to watch in the skill carrier specifically. If a skill is rewritten wholesale after every rejection, the conditions and cases that made each lesson precise get summarised away. Researchers building evolving prompt "playbooks" call it context collapse, "where iterative rewriting erodes details over time."¹⁷ Update incrementally: lesson, condition, case, test. Never just the slogan.

Judgment on both sides of production, and the superstition trap

When I'd given my thoughts and rejections, they ended up in the prompting, obviously. But they also ended up in the testing. That's the point of the two-sided carry: the same judgment works in two places.

The second form matters because an instruction can be satisfied by intention alone. The builder can sincerely believe it created warmth or recognition. The reviewer has to inspect what was delivered. That's why the harvest asked every builder to declare its recognition and aha moments, with times, so a cold viewer could be checked against them; the method of that check belongs to Code as Proxy. The progression here is the thing to notice: a specific rejection, then an explanation of the mismatch, then a more general selection principle, then a way to examine future candidates. That's far richer than storing "Scott dislikes this voice".

Anthropic's write-up of its own design harness shows both sides of this. Its author wrote four grading criteria and "gave to both the generator and evaluator agents in their prompts", and found that "even on the first iteration … the criteria and associated language themselves steered the model away from generic defaults before any evaluator feedback."¹⁸ Criteria are a production input, not just a test.

When taste becomes the bottleneck

Something shifts once a system can make a plausible film faster than it can establish that the film deserves to ship. In my case the test harnesses started slowing down final approval, on taste. Films got stuck in a loop where they wouldn't pass the branding. My first reaction was that this was good, I guess. The "I guess" was the right instinct.

A stuck film isn't automatically progress. A rejection is valuable when it identifies a meaningful discrepancy. Repeated rejection without a clearer diagnosis can mean the criterion is vague, the judges conflict, or the builder is being pushed toward the evaluator's favourite style rather than the intended outcome. The same Anthropic post reports that the wording of its criteria "pushed designs toward a particular visual convergence", which Demonstrate Over Tell has already dwelt on. The general pattern is well documented: optimise an imperfect proxy hard enough and ground truth suffers, "in accordance with Goodhart's law";¹⁹ AI-assisted creative work scores better individually but comes out "more similar to each other".²⁰ For a brand, whose whole job is to be this company, that trade is fatal.

The triage question is the guard. It preserves the value of strict review without enthroning the reviewer. And it keeps taste from hardening into superstition. B. F. Skinner's pigeons repeated whatever they'd been doing when food arrived, and "a few accidental connections between a ritual and favorable consequences suffice to set up and maintain the behavior."²¹ Skills do the same. Rejecting one voice mustn't ban every similar voice. Discovering an effective pause mustn't make pauses compulsory everywhere. In The Perturbation Review I put the rule for local lessons this way: "n=1 may climb, and it arrives wearing its ceiling."²²

A principle becomes more useful as its conditions become clearer, not simply as its wording becomes stronger.

Why the supervisor explained it best, and why it can't be the judge

Back to the page that started this. My best explanation of why it was so good is that the supervising agent had acquired something a cold writing prompt wouldn't contain: a pattern of decisions that needed a coherent explanation.

At the start, "why are we making these films?" invites a competent general answer about awareness, engagement, trust and conversion. After the work, the question has sharper material behind it. Why did the polished version fail? Why was that voice wrong? Why did a technically correct scene still miss? What did several different repairs have in common? An answer has to account for all of that, so it can't rely on generic marketing language. That's why the top-level agent was the right one to ask. A worker knows the repair it's making. The supervisor sees the relationships among repairs, the original purpose, my interventions and the critics' disagreements, and can notice that several apparently different problems are one failure.

I'd call this theory formation from practice. The film is an attempted answer. The rejection is evidence that the answer missed something. The explanatory page is an attempt to state a principle that accounts for what the attempts revealed. Then the principle goes back into production. Making improves the explanation; the explanation improves the next making. There's a research cousin: GEPA, a prompt optimiser that reflects in natural language over what happened in each attempt, argues that "the interpretable nature of language often provides a much richer learning medium for LLMs, compared to policy gradients derived from sparse, scalar rewards."²³ Detailed experience supports better updates than a bare score.

But notice what the Film E opening showed. The viewer that knew the context passed an opening I scored 6. That's the flip side of the supervisor's strength. Economists call it the curse of knowledge: "better-informed agents are unable to ignore private information even when it is in their interest to do so."²⁴ So you need two roles, deliberately separated.

An eloquent explanation is still a candidate

The supervisor's page was persuasive. That's exactly why it needs checking. An explanation earns its place in the skill when it helps with a new case, not when it makes the previous case sound inevitable. And the development records were valuable precisely because they could contradict the neat story. Two examples.

First, provenance. The neat story is that the agents discovered the series' central idea: ask people what would you change, and what would you keep? The records say otherwise. That question, the distinction between shape and colour, and the requirement that a person's preferences survive into planning were all in the research brief I supplied at the start, because I knew a small business wouldn't have one. Attempts sharpened and bounded those ideas. They didn't originate them.

Second, approval. The local voices in the clinic films were chosen by a listening judge. The record says, in plain terms, that I hadn't heard them yet. A high score from a listening judge is not my approval, and I haven't claimed it.

The brief, the films, the page and my discussion of them afterwards are connected derivatives, not four independent confirmations. Their value comes from asking new questions of the evidence, not from repeating the same interpretation until it feels established.

When the client can't brief you

Everything so far assumes someone, somewhere, has a sense of what good looks like. Now the hard case. A small business generally isn't run by an experienced marketing professional. If you ask the owner what's their brand guide, their language, their brand promise, they won't know. It's absent. Asking them to fill in a sophisticated brand questionnaire amounts to asking them to do the professional work they need help with.

The first discipline is to keep three sources of knowledge apart, so nobody ends up believing the AI found the whole strategy in the website:

What transferred into this engagement wasn't a successful video template. It was a way of thinking about communication. A template carries a previous answer. A way of thinking helps the system decide which questions a different client needs answered.

The second discipline is to triage the gaps, because "filling the gaps" hides three different jobs:

The sociologist Harry Collins has a useful name for the first kind: relational tacit knowledge, which "could be made explicit … but is not made explicit", including "unrecognized knowledge" a person holds but doesn't see as important.²⁵ Owners are full of it. They can't volunteer it, but they can recognise it when they see it.

Which is why production is such a good instrument here. "We provide personalised care" can stay vague forever. Making a film that demonstrates it forces one question that does most of the work:

What would the viewer actually see happening that makes this care personal?

That question can expose a hole in the script, a hole in the brief, or a hole in the business's own account of what it does. The film becomes an instrument for making the business more explicit.

Teaching the system what to notice as absent

Here the professional's judgment does its most important work, and it isn't supplying answers. It's deciding what counts as an unanswered question. Someone without a communication framework could look at the client's material and see enough to start: offer, price, logo, colours, contact form. A system carrying that framework sees other gaps. What would make the viewer recognise themselves? Where does their question change the response? What makes the offer understandable rather than merely visible? What would distinguish this business from generic imagery for its category?

Those aren't missing fields in a database. They're missing relationships in the proposed experience. The review rounds made the point with numbers: across seventeen verdicts from the harsher judge, "distinctiveness" scored under 7 in ten of them, because the generated imagery was "the stock vocabulary of premium cosmetic dentistry". The distinctive material was the business's own: its real enquiry form, its own lines, a real portrait. The principle went into the skill as one thing only this company has, with a test to match: hide the logo, and name what only this company has.

Developing the brand versus inventing the business

The better the system gets at filling gaps, the more one boundary matters. A film can propose that a business visibly preserve a patient's preferences. It can't establish, from a generated consultation scene, that the business already works that way. In Demonstrate Over Tell I put the rule in five words: "A demonstration explains. Evidence substantiates."²⁶

That isn't a reason to retreat to a generic ad. It gives the artefact another use. Show the client the proposed experience, the way I'd show a buyer a specimen rather than a prescription,²⁷ and ask three questions of each part of it:

That works in both directions. The actual service constrains what the film may honestly show. And the film can make a desirable service commitment explicit enough for the business to adopt. That's brand work becoming connected to service design.

The clinic films had a sharp version of this constraint. Australian health-advertising rules restrict testimonials and before-and-after imagery for cosmetic procedures; the regulator's guidelines say plainly that "advertising of higher risk cosmetic procedures must not use testimonials due to their potential to create unrealistic expectations."²⁸ So the films couldn't show a result. The review page's answer is one of my favourite lines from the week: no results, no reveal, because such shots "also make weaker films, because they promise an outcome nobody can promise." When you can't show a result, show being understood.

Their words change what happens

The strongest idea in the series wasn't the look. It was that a person's wishes have consequences. In the films, a consultation note fills in with the patient's own words: Change: the shape of one front tooth. Keep: the colour. Those words then turn up in the planning. The review page called it "Their words travel".

I'd sharpen that one step. Their words don't merely reappear. They change what happens. A system could copy "keep the colour" onto three screens and still propose exactly the treatment it would have proposed anyway. That's theatrical personalisation. The consequential version is when the preference visibly constrains the response: what gets discussed, what's preserved, which options stay open, what the person is and isn't being asked to commit to.

The research on listening makes the same distinction from the other side. In one experiment, paraphrasing someone's opinion back to them made the listener seem more socially attractive, but was "not associated with participants' … perceptions of feeling understood by the listener."²⁹ Personalisation researchers call name-insertion without changed behaviour "fake personalization … actually not personalization at all."³⁰ In my own Voice AI Readiness work, the dangerous version was empathy theatre: a voice agent that can sound "calm, confident, and caring while being operationally powerless."³¹ Same test every time: does the service do what the tone implies?

This also gives brand consistency a more interesting meaning. The family resemblance across the four films was deeper than the palette. It was customer agency carried through the experience. Different films can look and sound different while keeping the same relationship with the customer. Four beautifully colour-matched films can contradict each other about it.

The same test at three scales

My thinking partner pointed out, after reading the records, that the same principle turns up at three scales.

• For the patient in the film: does their preference change the discussion and the plan, or just decorate it?
• For me as the author: did my rejection of that opening change the judge brief, the cold viewer, the rubric and the skill? It did: it's in the skill text, the rubric and the test labels.
• For the account of the work: did reading the development records change the story we'd otherwise have told? It did: it stopped us crediting the central idea to the agents, and stopped us treating a listening judge's score as my approval.

A contribution has been understood when it changes the subsequent work, not merely when it has been repeated back.

In Made From You I put the founder's version of this bluntly: "A correction that changes nothing is not participation."³² It's the same rule. Being heard is a consequence, not a courtesy.

The proof: a successor test, not an essay

So how do you know a harvest worked? Not by how impressive the resulting explanation sounds. The proof is whether a fresh successor, using the harvested skill and evidence, can avoid a seductive wrong answer and recognise a genuinely good unfamiliar one, without anyone repeating the whole apprenticeship.

Learning science suggests how to design it. Schwartz and Martin found that students taught by telling and students who'd first invented their own approaches looked the same on an isolated test, but "the invention condition did better than the tell-and-practice condition when there was a learning resource."³³ A recipe and a formed judgment can look identical on a familiar case. They separate on a new one, with resources in hand. So the test gives the successor its resources and a case it hasn't seen.

Here's where the evidence honestly stands. An earlier version of the skill was tested once on a very different company, a farm-contracting concept on my platform, and its builder found its own recognition moment and its own devices without copying the first film.¹⁴ The harvested skill, with this week's lessons in it, is now on all twelve company boxes on the platform, and their builders pick it up on their next film. The successor test above hasn't been run on it yet. Until it has, the harvest is a well-formed candidate, not a proven capability.

What to do with your next rejection

You don't need a film pipeline to run any of this. It applies to anything judged by fit: a proposal, a landing page, a code review, a policy explainer, a course module. The practice is small enough to start on Monday.

• Separate told from revealed. Keep what the system was given at the start apart from what attempts showed you. Don't let the second be quietly credited to the first.
• Triage every rejection before fixing it. Artefact, instrument or principle? Ask the question out loud and write the answer down.
• Turn handles into tests. For every loose word you use ("warm", "premium", "wow"), write what a viewer would actually see or be able to say if it worked.
• Route each lesson to three carriers. The skill (principle, real example, your words), the harness (the same test), the contrasting cases (this, not that, with your verdict attached). Update incrementally; never compress to a slogan.
• Ask the informed to explain and the truly cold to test. Don't let the agent that watched everything be the judge of whether it lands.
• Attach conditions, not force. Every principle carries when it applies. A lesson from one case arrives wearing its ceiling.
• For clients who can't brief you, keep the three sources apart, triage the gaps, and show them a proposed experience to correct and adopt rather than a questionnaire.
• Judge the harvest by a successor, with the seductive wrong answer and the pass/fail criteria written down before it runs.

The film is one product of this work. A more precise, transferable understanding of good film is the other. The recursive process makes both at once, and only the second compounds. The autonomy that results isn't valuable because it removes you from the picture. It's valuable because your judgment can have consequences in places where you haven't personally specified every decision.

Scott Farrell is the founder of LeverageAI, where he helps Australian leadership teams turn AI experiments into governed, compounding capability, and builds the systems he writes about. scott@leverageai.com.au