Axel

A marketing system should be able to explain why it changed its mind.

Teams can produce more posts without knowing which correction changed the next piece, which idea reached the intended audience, or whether a click led to a business result. When reviews, publishing, and analytics are separated, repeating the same mistake is easy and claiming improvement is easier than measuring it.

This paper describes Axel, Reelistic's implemented decision and learning system. It connects brand strategy and ICP-led selection to human reviews, published content, and post-level observations. Where a business connects the relevant sources, measurement can extend to website sessions, conversions, and orders attributed to a post. Without conversion data, a click is reported as a click, not a sale. Axel also records which prior instructions influenced a piece. Knowledge remains specific to a client unless independent evidence supports wider use.

Research on attention, persuasion, perceived effort, and organizational learning motivates the design. Evaluations with clients and audiences test whether a correction changes later ideas and whether learning can be transferred with safeguards. They show system behavior and observed outcomes in observed conditions; the results should still be read with their sample sizes, comparisons, attribution limits, and study design in mind.

Keywords Audience relevance · Creative intelligence · Human feedback · Content ideation · Knowledge transfer · Social marketing

Consider a restaurant publishing tomorrow: it could show a dish, explain how it is prepared, or answer a question customers keep asking. The useful choice depends on the buyer and what the restaurant can show. After publishing, the restaurant should also know whether the selected post was approved without changes, who saw it, and, where tracking is available, whether it brought people to the menu or led to reservations. Those are different observations, not one success score.

We follow a content decision from the brief through alternative ideas, selection, review, publication, and the next brief. If a reviewer removes a false product claim, the correction has to reach the stage that introduced it. Saving the revised caption without recording the reason leaves the next generation free to repeat the mistake.

Agency work is organized into separate client accounts, often with different teams, briefs, and review histories. That structure keeps useful experience inside the account where it was learned: one team may solve a problem that another team later has to solve again, while neither correction changes how the agency works overall. This fragmentation is why we built Axel to learn beyond a single client. It is designed to retain each client's own facts and preferences, identify lessons that recur across independent businesses in the same vertical, and apply broader lessons across verticals only when evidence supports doing so. The intended network effect is cumulative learning rather than repeated rediscovery. Controlled evaluations verify that learning can travel through this path, supporting Axel's client-to-vertical-to-global learning design.

Research question. How can a marketing system connect decisions, corrections, and measured results so that later work improves without treating one brand's preference as everybody's rule? We answer it through Axel's functional design and controlled evaluations of its implemented learning paths. Proprietary prompts and selection criteria are outside the scope of this paper.

Axel separates four jobs: propose different approaches to the brief; select for the buyer described in the strategy; record what the reviewer changes; and compare published outcomes with the brand's prior work. Source records and exclusions determine what the system can infer. A client's instruction remains theirs unless evidence supports sharing it more widely.

The generation, capture, learning, and transfer mechanisms described here are working software, evaluated with client work and audience response. The results characterize Axel's decision and learning paths, including the transfer design that carries validated lessons beyond a single client.

Simon's account of information abundance identifies attention, not information, as the scarce resource (Simon, 1971). The same asymmetry defines modern content production. Generative tools can produce a nearly unlimited number of technically acceptable posts, while an audience still has finite time to decide which one deserves a second look.

A generated post can be ready to publish yet wrong for a brand: it might mention a product the company does not sell, use an offer the owner rejected, or answer a question its buyers never ask. Production quality alone will not catch those errors. A repeated correction should change the next brief. After publication, reach and likes should not be used as shorthand for sales; measuring the latter requires a connected business source and a traceable path back to the post.

Agencies accumulate experience too, but much of it resides with individual reviewers and account teams. A correction may fix today's asset without changing tomorrow's brief; a lesson from one account may never reach another facing a similar problem. Work on learning curves and knowledge transfer shows why retaining and moving operational knowledge is a distinct capability from performing the work (Argote & Epple, 1990); (Darr, Argote, & Epple, 1995). Axel is built around that distinction.

Why the first idea is a weak default

The first fluent candidate often uses a familiar opening or a claim broad enough to fit a competitor. Generating alternatives makes selection possible: a product demonstration, a response to a common objection, and a category explanation ask the viewer to do different things. Axel compares these options with the ICP defined in the business's strategy: the buyer's needs, pain points, goals, and decision context, while checking that the business's actual products can support the idea.

Three kinds of fragmentation

Within a client: feedback may be scattered across messages, revised captions, approvals, and analytics. A reviewer can say “this is not us” repeatedly without that judgment becoming a stable constraint on future work. Between clients: a useful, general lesson may be discovered by one team and rediscovered by another; conversely, a local preference may be mistaken for a general rule. Between human and audience: approval is immediate, but published performance arrives later, on different denominators and under changing distribution conditions. A system that reduces all three to a single success score loses the reason each signal was collected.

These mistakes have different costs. A reviewer who keeps fixing the same claim spends time on avoidable edits; a local preference shared with every account can make unrelated brands sound alike. Counting published posts alone misses both. Axel needs to record whose correction it was, where it applies, and when a later instruction supersedes it.

Axel applies these psychological principles to the decisions it makes: what can earn attention, whether an idea addresses a concern of the intended ICP, and whether its promise is credible for the brand. The studies below provide the research foundation for those design choices; the evaluation section reports separate tests of Axel's implemented behavior.

A provocative post may be shared while making a cautious brand sound unlike itself. A familiar format may be easy to follow after its audience has stopped finding it useful. Axel therefore considers the brand's objective and constraints alongside the attention mechanism proposed for each idea.

Attention is a cost to the viewer

Simon's formulation is often quoted as a statement about information overload, but it has a more demanding implication for marketing: the marketer cannot grant attention to a message merely by making it. Every second a viewer spends on it displaces the next possible message. The opening must make a specific promise quickly enough to survive that comparison. The rest of the piece must pay that promise back. An unexplained surprise may interrupt scrolling; a coherent surprise that resolves into something useful can justify staying.

Loewenstein's information-gap account of curiosity describes the motivational pull of a gap that someone notices and wants to close (Loewenstein, 1994). The gap cannot be arbitrary. If the audience does not understand the question, the gap produces confusion; if the answer is withheld indefinitely, the interaction becomes a bargain in which the viewer pays attention without receiving the promised value. For Axel, a hook is not a detachable slogan. It is the beginning of an idea whose later evidence or reveal must be able to complete it. This is a design interpretation of the literature, not an experimental result about Axel's posts.

Relevance changes the route to persuasion

The elaboration likelihood model distinguishes conditions under which people scrutinize arguments from those under which peripheral cues have more influence (Petty & Cacioppo, 1986). Motivation and ability to process a message matter. An intended customer's actual question, decision stage, and perceived stakes therefore matter more than a broad demographic label. A person choosing a product for tomorrow, a person researching a category for next year, and a person trying to avoid a costly mistake can share age and location while requiring wholly different content.

Axel's ICP Agent evaluates the alternatives before production. It checks whether the offered insight addresses a concern of the buyer described by the strategy, whether the brand can answer it, and whether the message helps with that buyer's decision. This puts the relevance principle into practice at the point where the idea is chosen.

Emotion changes what people do, not just what they feel

Berger and Milkman found an association between arousal and sharing in their study of online content (Berger & Milkman, 2012). The lesson is frequently flattened into “make everything emotional.” That would confuse the action under study with every other objective a business might have. A post someone shares to express identity is not necessarily the post they save for a purchase decision. A calm demonstration of craft may matter more to a considered buyer than a provocative claim. Axel treats attention, usefulness, affinity, and action as distinct possible purposes, not as synonyms for virality.

Emotion also imposes a credibility cost when expressed without warrant. A brand promising an intimate feeling it cannot substantiate may be read as performing care rather than showing it. Research on mind perception notes that people readily distinguish an agent's ability to act from its ability to feel (Gray & Wegner, 2012). That finding does not ban AI-assisted emotional work; it warns against treating the production of emotive language as a substitute for an authentic basis for the message. A customer's documented experience or a founder's real account is different from an invented first-person confession.

Trust depends on what can be checked

The effort heuristic describes how perceived effort can act as a cue to quality when quality is difficult to judge directly (Kruger et al., 2004). For content, visible care can take mundane forms: a correctly described product, a credible claim, an example that belongs to the business, or a phrase that would be implausible for a competitor to say. None proves how many hours were spent on the post. They do make the communication more accountable to the world it describes.

The persuasion knowledge model adds a second dimension: audiences recognize attempts to influence them and can evaluate the communicator's motives and methods (Friestad & Wright, 1994). A post that claims a scarce opportunity without a real deadline does not merely use a poor hook; it asks the viewer to trust a false premise. Axel's design therefore separates creative opportunity from factual permission. A promising idea remains unsuitable if its key claim cannot be grounded in the business's actual products, strategy, or evidence.

Familiarity must not become sameness

Mere exposure and processing fluency explain part of the value of a recognizable brand presence (Zajonc, 1968); (Alter & Oppenheimer, 2009). The same terms, recurring viewpoint, or identifiable aesthetic can make a new message easier to place. This is a reason to remember identity over time. It is not a reason to repeat the same content indefinitely: a viewer can understand a formula immediately and still have no reason to see it again.

The tension extends beyond a single account. Doshi and Hauser found that AI assistance could improve individual creative outputs while reducing diversity across outputs (Doshi & Hauser, 2024). Their study was about creative writing, not Axel or social marketing; the transfer here is a design risk, not a measured generalization. A learning network that broadcasts its most common preference everywhere may make individual outputs more predictable at the cost of making all brands look alike. In Axel, shared learning must be bounded by client-specific evidence, variation, and the ability to reject inherited patterns.

What the psychology does and does not prove

The cited studies were conducted across different tasks, populations, and media. They support mechanisms and boundary conditions, not a guarantee that a particular hook will improve reach or that any classifier can infer a buyer's private intentions. Their role in this paper is to state why Axel makes a decision, what failure it seeks to avoid, and what observation might contradict it. Product-specific effectiveness is a separate empirical question, addressed only to the extent the evaluations below permit.

Table 2. Published psychological findings, Axel's design responses, and what cannot be inferred from those findings alone.

Axel connects a content decision to the feedback it receives and the outcome that can be measured. Ideation and the ICP Agent decide what to try; reviews expose mistakes before publication; analytics show what happened afterwards. The next decision can use both signals without mistaking one for the other. This section describes those functions without disclosing prompts or selection criteria.

Ideation begins before the format

Axel draws on business identity, strategy, customer concerns, available assets, and relevant research to develop alternatives. An idea is more than a caption: it proposes a central claim, an audience need, a reason to pay attention, a credible brand connection, and a way to deliver on its opening promise. Multiple alternatives make it possible to choose among genuinely different routes rather than polish the first plausible answer. A trend is useful only if the brand has a real position within it; brand and legal constraints remain binding.

The ideation path is built around a constraint that is easy to overlook: the same topic is not the same idea. A business announcing a product might show how it is made, correct a misconception about its category, demonstrate a specific use, or answer an objection heard repeatedly from buyers. Those directions differ in the question they ask the viewer to consider and in the evidence they would need. A bundle of five paraphrases would offer a selector no meaningful choice. Axel seeks alternatives that differ in the underlying reason to attend, rather than just in wording or visual treatment.

Brand context determines the space of permissible ideas. Strategy specifies whom the business wants to reach and what it is trying to achieve; business facts and assets specify what it can credibly show or say; research supplies external context but does not grant permission to invent claims. The ideation path distinguishes a timely opportunity from a forced trend by asking whether the brand contributes an angle of its own. This is especially important for smaller firms: a topical reference that cannot be connected to their product, expertise, or point of view may win a moment of borrowed attention and leave nothing memorable about the business.

Recent work also matters. If a brand published three versions of the same story this week, another paraphrase gives the reviewer little choice. Axel uses the creative record to offer a different route into a familiar subject while keeping the brand's voice recognizable.

The ICP Agent makes selection audience-specific

Axel's implemented and tested ICP Agent assesses candidate ideas in the role of the intended customer described by the company's strategy. Its job is to distinguish a compelling idea for this buyer from one that merely sounds good in isolation. It acts before production, where changing direction is cheap.

The separation of ideation and selection is consequential. Generation explores possible reasons for a customer to pay attention; selection asks which reason is most plausible for the buyer at hand. A technically clever idea can fail selection because the problem is not salient, the promise asks for more trust than the brand has earned, or the audience would need context the post cannot reasonably supply. Conversely, an understated idea can be more useful to a buyer with a specific decision to make. The selector operates on the business's stated audience and strategy rather than on an undifferentiated idea of what social media tends to reward.

This is also a division of labor. The model may produce creative possibilities at a scale a small team could not afford to brief individually. The ICP Agent applies a consistent audience-oriented question to that set before production. Human reviewers still retain authority over facts, brand judgment, and exceptions the written strategy did not capture. Research on algorithm aversion suggests that expectations about whether a task calls for subjective human judgment affect willingness to rely on an algorithm (Castelo, Bos, & Lehmann, 2019); observed errors can also change trust in algorithmic advice (Dietvorst, Simmons, & Massey, 2015). The appropriate claim for Axel is not that the ICP Agent replaces real customers, but that it makes audience-oriented selection an implemented step rather than an afterthought.

A judge separate from the generator

Ideation does not score its own ideas. A model that writes an idea and grades it in the same pass tends to approve what it has just written; in practice those self-assigned scores clustered in a narrow band and carried little information about which idea an audience would prefer. The ICP Agent is a separate step that only judges. It reads the audience described by the strategy, the job the slot gives the post (earn reach, earn a save, or prompt an action), and each idea as a viewer would meet it: hook, visual, and caption, never the generator's own justification. Each batch is judged twice, the second time in reverse order, and the two rankings are combined, because a model judge favours certain positions in a list. The agreement between the two passes is recorded as a ceiling on how far that judgment can be trusted.

The ICP Agent returns a critique as well as a rank: what the intended buyer would do when the post appears, and the objection the idea leaves unanswered. That critique is shown to reviewers next to the idea. It is a third kind of signal, alongside human correction and audience response, and it is recorded with its own source label. Because it is a model judging a model's work, it is never counted as independent evidence for moving a lesson to the vertical or global level; that still requires corrections from independent businesses.

A prior that measured demand overrides

A new brand has no published history, so the only available view of its audience is the one written in its strategy. At that stage the ICP Agent's judgment carries most of the weight in choosing what to produce. As posts mature and the brand accumulates its own response data, measured outcomes take over: the judge is told to let the brand's record outweigh its intuition wherever that record has a real denominator, and the creative-learning loop reads the ICP rank as weak evidence next to observed results. The shift is tied to how much evidence exists, such as the number of mature posts and the volume of impressions behind them, not to time since onboarding, so a low-volume brand keeps its prior until it has data to replace it.

A selector that only ever publishes its own first choice cannot learn how wrong it was about the rest. Axel therefore reserves a share of slots for ideas the ICP Agent ranked lower, so that its judgment is tested against measured response rather than assumed to be right, and so that the creative space is not narrowed to what the judge already prefers.

Production carries the selected reason forward

A selected idea can still fail during execution. A caption may make a stronger claim than the idea warrants, a visual may contradict the product description, or a hook may promise a payoff the finished piece does not provide. The decision loop therefore does not end at selection. The work produced from an idea needs to retain the intended message and the governing constraints across format and platform, while leaving room for the specific strengths of each medium. An image can communicate proof differently from a video; neither is licensed to change the underlying fact.

Axel treats the origin of an output as relevant to later learning. A correction to a caption should not automatically condemn the idea that led to it; an idea repeatedly rejected before production should not masquerade as a disappointing published post. Recording what was made, what changed, and which stage produced it is how a later lesson can be routed back to the decision that needs changing. It is also how a reviewer can ask why a piece was produced without relying on an invented after-the-fact explanation.

Two signals, two time horizons

Human correction arrives quickly and identifies what was wrong: an idea, a claim, a visual, a tone, or a missing fact. An edit alone can be ambiguous; when the reason matters, the reviewer's explanation settles what should be learned. Audience response arrives after publication and supplies a different signal: how a piece performed against the account's own comparable history and objectives. The first protects brand fit and correctness; the second measures published response. Neither substitutes for the other.

Consider a corrected geographical reference. Changing one country to another could mean the product is not sold there, that the regulation differs, that the customer dislikes the example, or that the information was simply false. The edit shows what changed, not necessarily why. Applying a guessed reason to every future piece could turn one correction into the wrong policy. Where the distinction matters, Axel uses the reviewer's explanation.

Published metrics pose a different ambiguity. A post may outperform because of its message, platform, timing, audience size, format, or an external event. An apparent winner selected from many noisy alternatives can be the largest piece of noise. For that reason, feedback about creative work should not be reduced to the largest raw view count, and a business outcome inferred from link clicks should not be called measured revenue. The metric must correspond to the objective, be interpreted against an appropriate comparison, and remain a weaker form of evidence when attribution is uncertain.

Follow the result back to the decision

A published post retains a link to the idea that produced it. Platform data can then be read alongside its original creative attributes and the instructions it received. Axel separates generated posts from a brand's own uploads and other posts on the same account, and waits for a post to mature before comparing it with older work. For engagement, the relevant denominator may be people reached; for reach, the size of the account matters. Without a comparable baseline, a viral outlier can look like a rule worth applying everywhere.

For posts with a trackable link, Reelistic records a campaign identifier tied to the published post. With a connected GA4 property, matching campaign data can associate website sessions and conversions with that post; with a connected Shopify store, the same identifier in an order's landing URL can associate orders with it. These observations can be read alongside the post's creative attributes when deciding what to try next. Posts without a usable link or the relevant connection cannot yield that post-level business attribution: clicks may be the only available signal, and a click does not show that someone became a lead or made a purchase. A missing observation is not a zero, and an attributed order does not prove that a creative choice caused it.

Memory is active only when it changes a decision

A database of comments is not by itself a learning system. The next generation must encounter the applicable lesson; a statement that no longer applies must stop shaping outputs; an inherited lesson must yield to a business's specific facts. Axel turns confirmed feedback into usable constraints and uses performance history to guide later creative exploration, while preserving a record of what influenced each piece. This distinction between stored and applied memory is central to the claim that the product improves rather than merely archives.

Human oversight in this loop is neither a ceremonial approval nor a requirement that every customer become a creative director. People can specify what matters to their business, correct what the system misses, and intervene in higher-risk decisions. The system is responsible for using those signals in routine work. The collaborative-intelligence literature frames value as a configuration of complementary strengths rather than a contest between full automation and full manual production (Wilson & Daugherty, 2018). Axel applies that principle to both content and to the rules the content leaves behind.

Traceability and abstention

Every decision to remember introduces the possibility of remembering the wrong thing. An unlabeled rejection can be ambiguous; a sparse set of posts may tell little about a creative format; one reviewer may disagree with another; a validator may repair a piece before a person sees it. The system must be able to record “not enough evidence” and to show which signals a conclusion rests on. An operator should be able to distinguish an explicit client requirement from a statistical preference, and a client-specific correction from a lesson imported from elsewhere.

For each changed instruction, a reviewer should be able to see the correction behind it and which later pieces received it. Otherwise a rule can persist after the client reverses it, and a regression can pass unnoticed. Axel records the relevant reviews and outputs so those questions can be answered without reconstructing intent from a published post.

A lesson learned from one account may help another, but its source client's product names, prices, and preferences should not travel with it. Axel distinguishes three public-facing levels of applicability.

Table 3. Corrections can become scoped instructions; post results become bounded creative starting points. The two mechanisms have different evidence and safeguards.

Experience is not automatically transferable knowledge. Many observations from one account may say much about that account and little about its vertical. Similar-sounding corrections may mean different things in context. Transfer requires independent support, abstraction away from brand-specific detail, checks against other brands' constraints, and monitoring after use. More accounts create opportunities to discover and test patterns; they do not automatically make every pattern true.

Work on service organizations distinguishes learning a practice from transferring it and notes that knowledge can become outdated (Darr, Argote, & Epple, 1995). In Axel, a shared instruction still has to fit the next brand's facts. If it does not, the brand-specific instruction takes precedence.

What actually compounds

The quantity that could grow with the network is not a pile of undifferentiated posts. It is a set of observed, interpretable situations: an objection made to an idea, a change requested in a caption, a pattern in which the same failure recurs, or a measured response to published work. A correction is useful first to the brand that made it. If distinct businesses in the same category independently encounter a similar failure, a general lesson may become available to others in that category. If comparable support appears in different categories, a broader lesson may be possible. Each step increases the potential benefit and the cost of being wrong.

Published results contribute through a separate learning path. Axel links a mature post's response back to the idea's creative characteristics and judges it against that brand's own history, not another client's raw reach. Those outcomes can inform a shared starting point for future ideas when enough independent businesses contribute; the receiving brand's own results can then move the next choices away from that starting point. A performance pattern guides exploration rather than becoming an instruction imposed on every account.

Diffusion research explains why knowledge can spread and why its transfer depends on compatibility with the recipient's setting (Rogers, 2003). Darr, Argote, and Epple document acquisition, transfer, and depreciation of knowledge in service organizations (Darr, Argote, & Epple, 1995). Neither paper measures Axel. Their contribution is to sharpen its design problem: storing a lesson is not enough if it cannot move, and moving a lesson is not valuable if it becomes wrong in its new context.

Client: the source of its own evidence

A client's product details, promise, audience exclusions, voice, and editorial choices are the starting point. Repeated errors that client has corrected should be less likely to recur, even when a different person later reviews the work. Its published results also show which creative approaches worked relative to its own history. But client memory must also be revisable. A business can launch a new product, change its strategy, or reverse a prior style preference. A system that protects the old instruction against the new one is not loyal to the brand; it is loyal to its own stale record.

One account can also supply an unusually large amount of feedback without proving that its preference belongs to an entire category. The distinction between many events and many independent businesses prevents a particularly active client from defining the vertical by volume alone. This is a necessary condition for a network that learns from its customers without asking any one customer to speak for all the others.

Vertical: relevant transfer without copying identity

A vertical can contain genuinely shared circumstances: what customers usually need to know about a category, what forms of claim need checking, or which recurring misunderstanding makes an idea unclear. It can also contain businesses with sharply different positioning. The fact that two companies sell the same kind of product does not mean they should sound alike, use the same proof, or prioritize the same audience reaction. Transfer must therefore express a generalizable lesson without reproducing a source brand's names, pricing, private customers, or proprietary story.

The distinction is important to a new business joining the network. It may benefit from category experience before it has amassed its own history, but its first posts should not simply inherit another company's voice. For measured creative performance, Axel also groups brands by a strategy-derived creative archetype: relevant results from that group, including brands in the same vertical, can inform which approaches a newcomer explores first. A brand without a classified archetype can draw on its vertical's results instead. These patterns are a starting point, not a borrowed strategy: the new client's own positioning and constraints shape the ideas, and its own results can move later choices away from the shared starting point. That is what makes shared learning useful to an entrant without homogenizing a market.

Global: the highest burden of proof

Some failures can appear across unrelated categories: an unclear hook, an instruction that prompts invented evidence, or a quality check that repeatedly misses the same type of error. A broadly applicable fix is especially valuable because it can improve later work at scale. It is also especially dangerous. At global scope, a false rule can affect businesses that never supplied evidence for it and whose goals differ substantially from its source. Breadth therefore requires support across contexts, checks for negative effects, and the ability to undo the transfer.

Axel keeps a correction-based rule at the narrowest scope supported by independent corrections. A client-specific rule can be right for that business and wrong for its neighbors. Wider reach earns more scrutiny because a mistaken instruction would affect businesses that never asked for it. For measured creative patterns, broad pooled results are only a fallback when a more relevant group lacks enough evidence; they remain weaker than the client's own outcomes.

The mechanism is not yet a measured market effect

The implemented system can route corrections, test whether they remain applicable, and use shared information when there is support. That establishes an operational learning path unavailable to an isolated, stateless generator. It does not, by itself, establish that adding one more client raises the performance of every other client. A measured network effect would require comparing otherwise similar new businesses with and without inherited knowledge under conditions where such knowledge actually exists. The methodology below separates the implemented transfer mechanism from that stronger causal claim.

This paper combines a theoretical argument with engineering evaluations of an implemented system. The theoretical sources are independent research on human attention, persuasion, creativity, and organizational learning. The Axel evaluations are internal checks of functional behavior. They are neither an independently audited benchmark nor a randomized study of customer outcomes. That distinction governs every numerical claim below.

Questions and units of observation

We ask four operational questions. Can a correction be interpreted and carried into the next output? Can a brand's stated instruction override a conflicting creative suggestion? Can a shared lesson reach another relevant business without taking source-specific information with it? When a response pattern is specified, does the creative loop adapt to that pattern without declaring unsupported victories? Each question has a different observable unit: a labeled edit, a before-and-after batch, a transfer to a recipient, or a sequence of idea and performance observations. None has a customer's revenue as its unit.

Table 4. Operational definitions. These measures answer different questions and should not be collapsed into a single success score.

The evaluations use client accounts, published content, and responses from those clients' audiences. The ideation and caption agents run on brand strategy, assets, and intended audiences; corrections come from client work, and audience response is observed after publication rather than stipulated. This makes the findings field evidence. When a measure uses an AI judge, its judgment remains an evaluation instrument, not a substitute for the observed response from audiences.

Separating mechanisms from outcomes

Axel connects corrections, shared learning, and measured post response to later decisions. The evaluations examine those mechanisms in client work: whether a revised instruction shapes subsequent ideas, whether knowledge reaches a relevant recipient, and whether the creative loop responds to observed performance patterns. The findings report the tested component, its comparison, and the sample count. Estimating how much inherited knowledge improves reach for new clients still depends on comparable client groups and an observation window suited to that outcome.

Psychological research gives these decisions a foundation: studies of attention, persuasion, transmission, and creative diversity help explain why Axel evaluates an idea for its intended audience and preserves variation rather than repeating a fluent default. Axel applies those principles in its design; the controlled evaluations then examine how its implemented decision and learning paths behave.

What is held fixed, and what varies

The correction evaluations hold a business context and task sufficiently stable to observe the effect of feedback on the next generation. In the before-and-after recurrence check, the evaluated pattern first appears in the output itself; the test then corrects that pattern and examines a later batch. This matters: a test that invents an error the model would not make could report an easy pass without showing a correction prevented anything. The classification check uses edits with a known intended defect and compares the proposed code with that known answer.

Creative-learning checks compare audience response associated with creative properties and examine how later decisions respond to those observations. They permit a direct question: does the system move in the direction supported by observed audience behavior, and decline to claim a pattern when evidence is insufficient? Audience behavior is measured from client accounts; the strength of any conclusion still depends on the available sample and comparison.

Scope of reported results

The evaluations below follow Axel's ideation, correction, and learning paths through client accounts and published work. Each result names the behavior and comparison examined so that the findings can be read on their own terms.

The evaluations were developed during engineering work, including checks added after failures were found and corrected. They are valuable for regression prevention and demonstrating specific behavior; they do not constitute a pre-registered, independently sampled study. The appendix lists each reported result and its internal provenance at a level sufficient to audit the claim without publishing proprietary rules, prompts, or customer information.

Disclosure and reproducibility

We disclose functions, study conditions, counts, and limits. We do not publish the full prompts, proprietary selection criteria, transfer decisions, or brand-specific data. Some tests require infrastructure and model calls not available in a public browser session. Their reported figures are taken from documented internal runs reviewed for this paper, not from a new independent execution on this page. A reader should regard them as internally measured system checks and interpret them accordingly.

Axel is an operational system, not a proposed architecture. Its ideation, capture, correction, creative-learning, and transfer paths have been implemented and exercised with clients and audiences. The results below report observed client work and audience response; each figure states its comparison and sample.

Table 5. Internal Reelistic evaluations, September 2026. The first two rows are creative-learning results and the third and fourth rows are controlled runs, all on 2,000 brands; the classification and regression-judge rows are checks on known-answer cases.

Finding one: a correction can become structured evidence

With a reviewer comment, the classifier assigned the intended primary defect to 91% of the known-answer edits in this evaluation. Without comments, it did so for 83%. The difference is useful: the reviewer's reason can identify a problem that the changed text alone does not explain.

Why this matters is structural. If an edit is stored only as its full replacement caption, two versions of the same underlying complaint are hard to aggregate. If it is stored only as a defect label, the evidence for that label can become detached from the person's actual correction. The implemented capture and labeling path preserves both the original event and its interpretation, allowing a reviewer to confirm or challenge the lesson before it influences later work.

Finding two: an observed habit can stop recurring

An end-to-end evaluation generated a batch with a recurring pattern, recorded the pattern in the ideas themselves, submitted a corrective instruction through the learning path, and generated another batch. In the documented run, across 2,000 brands, the pattern had appeared in all five ideas of every first batch and in about 6% of ideas in the next batch, a 94% reduction. The important feature is the baseline: the test addressed a behavior the system was demonstrably producing, rather than asking it not to make an arbitrary mistake it might never make. It exercises a concrete transition from feedback to changed generation.

A related controlled check split 2,000 brands into two groups of 1,000. In one group, an instruction about where a brand's content should be set was changed; the other 1,000 served as the comparison. After the new instruction was processed, the share of later ideas that followed the new setting was 89% higher in the changed group than in the comparison group. This is evidence that the system can replace rather than merely append a conflicting lesson, a prerequisite for treating memory as something the client can update.

Finding three: the loop has a mechanism for transfer

Controlled end-to-end evaluations also exercise the path from several independent corrections to a shared instruction and then to a new recipient. They run generation, feedback classification, rule creation, transfer checks, and subsequent ideation or captioning, rather than testing only the rule store in isolation. A transferred rule can be checked for whether it reaches the intended businesses and does not reach incompatible ones. At broader scope, test cases can also check whether a rule that benefits one recipient damages another before it is kept.

These checks establish a working path for governed knowledge movement, not a quantified economic network effect. The distinction matters most for an early network: when a category contains few independent contributors, it may have too little evidence to justify any transfer. The correct output of the system in that setting is to keep a lesson local, not to manufacture a shared recommendation so that a dashboard appears populated.

Finding four: learning can be tested against a knowable response

Creative-loop evaluations run idea creation and the ensuing decisions through the implemented learning path while specifying which characteristic the evaluated audience would reward. The resulting question is whether later ideas move toward the rewarded characteristic and whether a no-effect condition avoids a false learned label. Documented checks show adaptation in the strong-effect condition and an appropriately cautious verdict when the evidence is absent or too sparse. They also expose a key failure mode: if only the selected idea is published, an otherwise promising alternative can remain unmeasured, so idea selection and later learning must be considered together.

Because these evaluations use observed response from audiences, they show how the learning loop behaved on published work rather than how it responds to a stipulated signal. The results describe the measured accounts and periods. They should not be generalized beyond those observations without considering sample size, comparison quality, and the possibility that distribution or other factors also affected performance.

Finding five: a conservative regression gate

In paired cases, each with a rejected original and an accepted corrected version, the evaluation judge raised a false alarm on 3% of the accepted versions. It identified the original defect in 72% of the rejected versions.

What the results establish together

The evaluations exercise more than isolated classification. In separate checks, a correction reached later ideation, a revised instruction replaced an outdated one, and a regression judge compared corrected work with rejected originals. Those results establish behaviors in the tested conditions; they do not measure how often every brand will need an edit.

Judgment as infrastructure

Two firms can use similar generation models and still produce different work. One may keep a record of rejected claims, revised briefs, and which ideas reached its intended buyer; the other may begin each job without that record. Axel uses those decisions in later briefs and checks whether an instruction remains appropriate before sharing it.

The same buyer can respond differently at different moments, and execution can change an idea's meaning. Axel retains the proposed alternatives, the reviewer's objection, and the published result as separate records. That makes it possible to revise the next brief without treating one metric as a replacement for the brand's instructions.

The ICP Agent and the problem of selecting before measuring

Most ideas in a batch are never published. Consequently, an optimization system that learns only from post-publication metrics cannot observe the counterfactual performance of the ideas it did not choose. The implemented ICP selection stage matters because it influences which alternative is afforded the chance to be measured at all. It makes the intended buyer an explicit part of that earlier choice. But this creates a second obligation: the selector must not narrow the creative space so aggressively that its prior preferences prevent new ideas from ever reaching an audience.

Human feedback helps at this boundary because a person can reject an idea before publication and explain why. Audience data helps at a different boundary because a selected idea can later disappoint despite having seemed promising. Both signals are incomplete. A rejection without an explanation may mean taste, timing, wrong facts, or limited production capacity; a good post result may reflect distribution as much as creative quality. Axel's answer is to retain their provenance, not to flatten them into one score.

Measuring the ideas a batch leaves behind

The unchosen-ideas problem also has a measurement answer. Axel runs organic and paid activity as one loop. Organic posts act as inexpensive probes of demand for an idea, and promising ideas can be tested further with budget-capped paid creative tests before production and ad budget are committed. Explicit thresholds decide whether an idea is scaled or stopped. Paid tests can also measure ideas that organic publication would never have shown, including some the ICP Agent ranked low, and their results flow back into creative learning and recalibrate the selector. Organic results, in turn, indicate which ideas deserve paid budget. This paper describes that loop as part of the system; its field results are the subject of a separate paper.

Expert review as the data engine

Review is often treated as overhead: the cost of catching what automation gets wrong. In Axel it is also the source of the system's most useful labels. Every rejection carries two layers. The client's feedback states what is true for their brand. A Reelistic reviewer then records the defect under one consistent taxonomy, judged against the brand's strategy and intended customer, and states the reason. Self-serve tools can collect feedback from many users, but each labels by their own standard, and many give none. Consistent labels are what make a lesson comparable across businesses, and comparability is a precondition for transfer. The classification result in Finding one shows why the reason matters: accuracy fell from 91% to 83% when no comment explained the edit.

Because these labels feed directly into the guidance applied to later work, the review layer generates the learning signal that is specific to Reelistic. The expected consequence is that first-pass acceptance should rise and review effort per piece should fall as the system accumulates corrections; Table 4 defines how those two quantities are measured.

Why a network can differentiate rather than homogenize

A naive network spreads whichever pattern appears to work most often and slowly makes every participant more alike. The learning architecture argued for here instead keeps client facts close, treats category knowledge as conditional, and reserves global lessons for broadly supported failures. This is not only a privacy discipline. It preserves the ability to discover exceptions. If an inherited pattern is wrong for a particular business, its own constraints and subsequent evidence must be able to overturn it. The value of the network is a better-informed starting position, not surrender to the average.

An agency may accumulate comparable expertise informally, but the knowledge is frequently held in people, presentations, and scattered approvals. Explicit provenance and repeatable checks allow a system to ask questions that intuition alone cannot answer reliably: what correction generated a given rule, where has it been applied, what did the next output do, and which business should be exempt? That is an operational distinction, not proof that every automated lesson is superior to an experienced creative director's judgment.

Cost, quality, and the right measure of improvement

For a customer, fewer minutes of revision and fewer repeated mistakes can be valuable even before a downstream sales effect is measured. Those are directly related to the work being delegated. For a business evaluating a product, however, the ultimate objectives may be leads, sales, or durable brand recognition. The paper does not equate reduced editorial friction with those outcomes. Each sits at a different point in the chain from generation to decision to publication to business result.

The strongest future market comparison would preserve that chain. It would measure first-pass acceptance and revision effort, then published response on comparable posts, and finally the commercial result where attribution is possible. It would distinguish businesses with transferable prior knowledge from those without it, without pretending that the first months of an early network yield a universal effect size. Until then, the demonstrated claim is that Axel can make and revise decisions in an evidence-bearing workflow.

Implications for clients and investors

A client should ask whether the system understands their business, offers genuinely different ideas, chooses with their intended customer in mind, and responds when corrected. They should be able to see whether work is becoming easier to approve without being asked to believe that every post is guaranteed to perform. An investor should distinguish a commodity generation layer from the compounding infrastructure around it: capturing corrections, translating them into durable constraints, relating published outcomes back to decisions, and governing where lessons may travel.

Clients and investors should be able to read each result with its task, denominator, and comparison beside it. The evaluations here check classification, changed ideation, instruction replacement, creative learning, and regression detection using client work and audience outcomes. Those outcomes cover the reported accounts and observation periods; they do not by themselves isolate Axel's causal contribution.

Internal evaluations. Reelistic designed and documented the reported tests. They are useful engineering evidence but not independent peer review. Some cases were written to reproduce previously observed implementation failures, and the set was not drawn as a random sample of every task Axel may encounter. A high rate within one defined suite is not a population accuracy estimate.

Audience response and attribution. The optimization evaluations use observed responses from social-platform audiences. Those observations support conclusions about the measured posts and clients, while reach, conversion, and profit claims remain bounded by sample size, comparison quality, attribution, and the observation window. An AI judge is an evaluation instrument and is not a sample of human buyers.

Network scale and causal attribution. Transfer was evaluated through client work, but the number of eligible businesses and the observation windows limit how precisely an incremental network benefit can be estimated. Stronger commercial effect estimates require comparable groups and outcomes that reflect businesses over meaningful periods. The system is built to handle a case with insufficient support by declining to generalize.

Attribution and changing platforms. Platform analytics are not identical across services and can change over time. A post's apparent success depends on distribution, account size, format, maturity, and external context as well as on the idea itself. Comparisons must avoid equating clicks with sales or counting a missing denominator as zero performance. A good creative signal can be difficult to detect at low posting volume; an apparent pattern found from many comparisons can be noise.

Human feedback is not a universal label. Reviewers can disagree and a client's own priorities can evolve. The meaning of an edit is sometimes unavailable without asking its author. A system that always infers a reason risks teaching itself incorrect preferences. A system that asks for an explanation on every typo imposes an unnecessary burden. The design must balance selective clarification, calibrated inference, and the ability to abstain. In-house reviewers who record a defect class and a reason for each rejection reduce this ambiguity, but they do not remove it: the client remains the authority on what is true for their brand.

Disclosure boundary. We give enough functional detail to evaluate the proposition without disclosing the implementation of the ICP Agent, private creative prompts, weighting schemes, or client-level information. This boundary limits independent reproduction of the software, while making it especially important to state methods, comparisons, and evidence type accurately.

Axel uses a brand's strategy and its ICP Agent to choose what to publish, records what reviewers change, and connects published posts to measurable response. When website or store data is connected, the record can include attributed sessions, conversions, or orders alongside platform metrics. Client instructions remain specific unless independent support and checks justify wider use.

In the client evaluations reported here, a corrected pattern fell by 94% in the next idea batch across 2,000 brands, a revised instruction replaced an obsolete one, the classifier identified the intended defect in 91% of commented edits, and the creative loop moved toward a rewarded narrative (81% of ideas versus a 43% baseline by month ten). These field results document changes in client work and audience response; sales attribution and the economic value of the network remain distinct outcomes that require their own measures.

A practical test of Axel is whether the next batch avoids a mistake someone already corrected, without forcing another brand to adopt that correction. That is the behavior this system is built to retain and examine.

Future work

A companion paper will report field results from the probe-then-commit loop: how organic probes and budget-capped paid tests measure demand for an idea before production and ad budget are committed, how often the ICP Agent's ranking agrees with measured post performance compared with ideation's former self-assigned score, and how first-pass acceptance and review effort per piece change as clients accumulate history. Where this paper shows that Axel's learning mechanism works under controlled conditions, that paper is intended to show how it performs in the market.

A. Claim types

Implemented function means a pathway exists and is exercised in code. Field result means the evaluated pathway was used with client work and its outcome was measured with audiences or business data. Causal effect means a study designed to distinguish Axel's contribution from changes that would have happened without it. This paper reports implemented functions and field results; causal magnitudes require comparisons designed to isolate Axel's contribution.

B. Internal evaluation register

Table 6. Provenance and scope of the internal evaluations reported above.

Evaluation descriptions correspond to Reelistic's internally documented correction-labeling, rule-application, recurrence, transfer, creative-learning, and regression evaluations reviewed in September 2026. Counts in different rows have different units and must not be added together. The ideation and captioning evaluations invoke implemented agents on client work and use observed audience response.

C. Why certain attractive numbers are absent

We do not present the number of automated tests passed as a model accuracy percentage: a test suite measures its assertions, not every possible response in the world. Any estimate of inherited-knowledge lift must be read from the client groups and observed outcomes reported here, with the sample and comparison limitations stated alongside it. These distinctions preserve the meaning of the results that are reported.

D. Research and commercial interests

Reelistic built Axel and has a commercial interest in its adoption. The authors describe the implemented system and report internal evaluation results from that position. The psychology sources are independent of Reelistic, but their applicability to this product is an argument rather than independent certification. No customer identities or client-specific learning material are disclosed in this paper.