Self-Improving AI Marketing Campaigns - A Grand Unified Theory

A Grand Unified Theory of Structure, Content, and Evolutionary Feedback

Deterministic scaffolding first. Stochastic enhancement second. Doctrine that amends itself last - and a substrate that makes all of it compound.

Status: Public draft, open framework License: Free to use, adapt, and redistribute

What this is: a unified, layered theory of how marketing campaigns improve themselves over time - from the simplest weight-tuning loop a two-person team can run this week, to a fully reflexive system in which market response revises the organization's own codified doctrine. The theory is company-agnostic and vendor-agnostic. It treats direct response marketing as the original feedback science, treats AI as an instrument that compresses that science's cycle time, and treats a version-controlled knowledge substrate as the only place learning can durably live.

Why it matters:

• Most "AI marketing" fails on marketing fundamentals, not model quality. Every failure of an AI-augmented campaign system can be traced to a violated law of direct response - an unmeasured funnel, a confounded test, a proxy metric optimized into corruption - not to insufficient intelligence in the model. The theory therefore begins with the deterministic laws and adds intelligence only where it survives them.
• Learning that lands in a vendor's database evaporates. Learning that lands in a sovereign substrate compounds. The single most consequential architectural decision in a self-improving system is not which model to use; it is where the learning is recorded.
• Self-improvement is not one capability but seven, stacked. Weight adaptation, variant selection, semantic representation, generative variation, structural evolution, and doctrinal self-amendment are different machines with different sample appetites, different risks, and different governance requirements. Conflating them is how teams end up running a reinforcement learner on forty data points, or A/B testing subject lines while the offer is broken.

Contents:

1. Lineage: a century of closed loops
2. First principles: the two dichotomies and the deterministic spine
3. The anatomy of self-improvement: nine questions
4. The economics of learning: samples, decay, and the Red Queen
5. The layer model
6. Credit assignment and time
7. Governance and guardrails
8. A bestiary of failure modes
9. Past the close: one graph, one loop, the whole journey
10. The asymptote: what "fully self-improving" means
11. Appendix A: schemas and pseudo-code
12. Appendix B: a diagnostic

1. Lineage: a century of closed loops

Direct response marketing was the first empirical science of persuasion, and it is worth being precise about why, because the entire theory in this paper is an extension of its epistemology rather than a replacement for it.

In 1923, Claude Hopkins published Scientific Advertising and described a practice his agency had already been running for two decades: keyed coupons. Every advertisement carried a code identifying which publication, which headline, which offer had produced the response in hand. Ads were not judged by acclaim, awards, or the copywriter's confidence; they were judged by counted replies against counted cost. "The time has come," Hopkins wrote, "when advertising has in some hands reached the status of a science." What he had built, in modern vocabulary, was a closed feedback loop with clean attribution: a hypothesis (this headline, this offer, this audience), an instrumented deployment (the key), an observed outcome (coupons redeemed), and a disciplined iteration (change one thing, run it again).

Everything that followed in the direct response canon is a refinement of that loop. John Caples and later David Ogilvy industrialized headline testing and established that the variance between a strong and weak headline dwarfs the variance from everything else on the page - an early, hard-won lesson in needle-mover identification: of the many things one could vary, only a few move the metric, and the discipline is finding them. Gary Halbert taught the primacy of the list - "a starving crowd" beats any copy improvement - establishing the hierarchy that still governs today: market first, offer second, copy third. Dan Kennedy formalized the triangle of message–market–media match and the follow-up doctrine (the fortune is in the follow-up; most sellers surrender after two touches when seven to twelve are required), converting persistence from a personality trait into a measurable structural property of a campaign. Eugene Schwartz, in Breakthrough Advertising (1966), contributed the deepest theoretical object in the field: market sophistication - the recognition that a market is not a static audience but a state machine, advanced through five stages by the cumulative effect of all advertising it has ever absorbed, such that the copy strategy that works is a function of where the market currently sits, and every winning campaign pushes the market one step closer to exhausting the strategy that won. Later formulations - Russell Brunson's hook–story–offer and value ladder, Alex Hormozi's value equation, systems-level treatments of client acquisition as a chain of state transitions - added compositional structure: the campaign as a pipeline of psychological state changes, each with its own conversion rate, its own bottleneck, its own testable variables.

Three properties of this century-old tradition matter for what follows.

First, direct response was always a selection system. Variants were generated, exposed to the market, measured, and the winners inherited the next budget. The vocabulary of evolution - variation, selection, inheritance, fitness - maps onto keyed-coupon practice without metaphorical strain. What the tradition lacked was not the loop but cycle time and carrying capacity: a human copy chief could run a handful of disciplined tests a month across a handful of variables.

Second, the tradition's laws are laws of measurement, not of persuasion. Hold everything constant except one variable, or the test yields nothing. Reach the sample size, or the reading is noise. Wait out the latency window, or you will kill a working system inside its lag. Judge against the metric that pays the bills, never a vanity proxy. These are not stylistic preferences; they are the conditions under which a feedback loop produces information rather than confident error. They bind any learner - human, statistical, or neural - that plugs into the loop.

Third, the tradition already knew its winners decay. Schwartz's sophistication stages are a formal statement that the fitness landscape is non-stationary and reflexive: the market's response function changes because of your campaigns. Every winning stimulus habituates its audience; every saturated mechanism forces the field up a sophistication stage. A theory of self-improving campaigns that models the market as a fixed distribution to be learned has misunderstood the field at stage zero.

The thesis of this paper is that modern AI - inference, embeddings, generation, and the surrounding machinery of instrumentation - does not replace this epistemology. It compresses its cycle time by orders of magnitude, extends its carrying capacity from a handful of variables to thousands, and, at the highest layer, allows the loop to close over something no coupon key could reach: the organization's own written doctrine. But it inherits every law. The layers in Section 5 are, in a precise sense, the historical practice of direct response re-instrumented at increasing altitudes - and every failure mode in Section 8 is a century-old sin wearing new tooling.

2. First principles

Three foundational distinctions govern everything downstream. Two are dichotomies; the third is an ordering.

2.1 Structure versus content

Every marketing campaign decomposes into structure - the shape of the journey - and content - the words, images, and offers that fill the shape.

Structure is the directed graph of the campaign: which touches exist, in what order, on which channels, with what timing, branching on which observed outcomes. Structure encodes the persuasion grammar the field has validated over a century: attention before interest, interest before commitment; problem–agitate–solve; hook–story–offer; multi-channel follow-up with escalating cadence; a human-escalation path when automation exhausts its inventory. Structure changes slowly because human psychology changes slowly. It is the Lindy-tested layer.

Content is what a specific prospect encounters at a specific node: the subject line, the first sentence, the mechanism claim, the proof element, the call to action, the offer's terms. Content decays quickly - measured in months in competitive channels - because content is what the market habituates to. Schwartz's stages are a theory of content decay under a comparatively stable structure of human decision-making.

The distinction is load-bearing for self-improvement because structure and content have different variation costs, different sample requirements, and different blast radii. Swapping one subject line for another is a cheap, reversible, locally-measurable change. Rewiring the journey graph - adding a branch, changing follow-up cadence, inserting a new channel - changes the denominator of every downstream metric and requires an order of magnitude more evidence to evaluate. A well-designed system therefore lets content vary early and often, within a structure that is versioned, validated, and mutated only deliberately. The layer model makes this precise: content variation is a Layer 2–4 activity; structural variation is Layer 5; and the doctrine that generates structures is Layer 6.

A useful formalization, borrowed from evolutionary biology and native to the direct response tradition's own testing discipline: every deployable campaign is a genotype - a tuple of (structure version, content versions, targeting definition, parameter values) - and the executed campaign, as experienced by a cohort of prospects, is its phenotype. Fitness attaches to genotypes. Each stimulus within the campaign is itself a small genotype of two to four needle-mover variables (a cold email: subject, first line, body mechanism, call to action), and the tradition's hardest-won heuristic applies at every altitude: most variables are noise; the skill is identifying the few that, if changed tenfold in either direction, would visibly move the metric. Self-improvement machinery pointed at noise variables is expensive decoration.

2.2 Substrate versus projection - and where learning must live

The second dichotomy is architectural, and it determines whether learning compounds or evaporates.

Every artifact in the system is either sovereign substrate - authored once, versioned, canonical, irreplaceable - or a derived projection, regenerable on demand from substrate and never authoritative in its own right.

The identity test is the regeneration question: can this artifact be reconstructed, in full, from other artifacts plus a definition? If yes, it is projection - a dashboard, a vendor-side campaign object, a rendered email, a vector index, a scorecard. If no, it is substrate - a decision record, an authored messaging framework, an observed outcome event, the definition of the ideal customer profile, the campaign graph the organization committed to. Substrate corruption is information loss; projection corruption is a render cost.

This distinction exists independently of marketing (databases have tables and views; event-sourced systems have logs and read models; version control has commits and log output), but it has a specific and under-appreciated consequence for self-improving campaigns: the location of learning determines its half-life. A scoring weight tuned inside a sending platform's proprietary settings, a "winning variant" flagged in an email tool's UI, an audience insight living only as a saved segment in an ad platform - all of these are learning recorded in projections. They are hostage to the vendor, invisible to any other tool, unauditable, and lost on migration. The same learning recorded as substrate - a dated amendment to the ICP definition, a versioned change to the scoring configuration, an appended entry in a decision log with the evidence attached - survives every vendor change, feeds every downstream consumer, and can be reasoned over by humans and agents alike.

The reference discipline for maintaining such a substrate is the Organizational Context Protocol (OCP) - an open, MIT-licensed specification (in the lineage of CommonMark and the Model Context Protocol) for how an organization exposes its structure, knowledge, and access rules to AI agents, with git and plain markdown as the single source of truth and everything else treated as a disposable projection of it. OCP is summarized here because it supplies, off the shelf and with zero vendor lock-in, exactly the properties a self-improving campaign system's memory requires:

• A readable, diffable source of truth. Organizational knowledge - ICP definitions, offer doctrine, messaging frameworks, decision records, campaign definitions, learnings - lives as versioned markdown, reviewable in a pull request, forkable, portable, and equally legible to a human and to an agent. Not locked inside an opaque vector store that cannot be inspected, corrected, or exited.
• Structure agents can reason over. A recursive organizational hierarchy and the closed OCP set of five artifact types - note, adr (decision record), prompt, template, report - replace the flat "bag of chunks." Closure at the type level is the point: five known types means five known schemas, so every tool works against every conformant organization; extensibility lives in an open role discriminator and metadata object that require no ecosystem coordination.
• Access and trust declared in the substrate, versioned with the content. Trust tiers and visibility scopes ride on each artifact's frontmatter, with the invariant that derivation does not launder trust: a fact inherits the trust tier of its lowest-trust source no matter how many model passes rewrite it. For a learning system this is the defense against a subtle attack and a subtle self-deception alike - an inference derived from unvetted data does not become canon merely because it landed in a nicer document.
• Versioning where the commit is the truth. Every artifact has a history; every belief has a date; "what did we know when we made that decision" is answerable by replay. Section 6 shows why this is not hygiene but a precondition for honest credit assignment.

Under this axiom, the relationship between the knowledge layer and the execution layer becomes crisp. The substrate holds the campaign's genome: doctrine, definitions, structures, content sources, decisions, and observed outcomes. Every vendor tool - the sending platform, the CRM, the ad account, the dashboard - holds projections: regenerable renderings of the genome into a runtime. Self-improvement, in its final analysis, is the disciplined return path by which market responses flow back and amend the substrate. A system that only tunes projections is not self-improving; it is redecorating rented rooms.

2.3 Determinism first: the executable spine

The third principle is an ordering: build the deterministic scaffolding before adding a single stochastic component, and keep every stochastic component isolated behind an explicit contract.

The deterministic spine of a modern campaign system consists of five commitments.

The campaign is a directed acyclic graph, not a sequence. Linear step lists collapse real prospect behavior into a false model; reality branches. A prospect replies positively, negatively, with an objection, with a referral, or not at all, and each outcome should route somewhere different. A sequence is merely the degenerate graph in which every node has one outgoing edge. Modeling the general case - nodes as actions (send, wait, enrich, route, escalate, exit), edges as outcome-labeled transitions - is what makes structure a first-class, versionable, analyzable object rather than an implicit property of tool configuration.

Outcomes live on edges, drawn from one shared taxonomy. A node's identity is what it does; the campaign's structure is where each outcome goes next. Fusing them destroys reusability and analysis. The outcome vocabulary itself must be a single, governed taxonomy shared by three consumers - the classifier that labels responses, the node's declaration of what it can produce, and the edge predicates that route on results - because a system with three implicit outcome vocabularies will drift three inconsistent truths. A practical refinement keeps the graph tractable: route on a small closed set of outcome categories (desired, retriable, redirectable, terminal-negative, terminal-neutral) while recording the specific outcome (not interested, meeting requested, out of office, referral to colleague, undeliverable) as data on the transition event. The branch count stays bounded by the category alphabet; the analytical richness stays unbounded in the data. And the taxonomy should be built with a negotiator's ear rather than an engineer's enum: "not interested" and "not interested but asked how it works" are different states deserving different edges. Calibrated outcome granularity is where listening becomes structure. Edges are also where expectations live: every transition carries its expected conversion band - green, yellow, red - and, authored at design time rather than in the panic of a red dashboard, a pre-committed intervention for the day the band is breached. The worst moment to decide what a failing metric means is while staring at it; a well-built graph already knows.

Telemetry is emitted once, at a chokepoint, and projected many times. One canonical transition event per edge traversal - carrying the genotype identifiers, the outcome, the timestamps - feeds every audience: the client's "is it working" view, the operator's "where is it leaking" view, the engineer's "what is failing" view. A system that emits audience-specific telemetry will drift audience-specific truths. Equally load-bearing: the business-outcome event stream and the operational telemetry stream are categorically different and must never be merged. Operational telemetry (latencies, errors, costs) may be sampled for affordability; business outcomes (replies, bookings, closes) must be recorded unsampled and durably, because a count sourced from a sampled stream is wrong by construction - off by the reciprocal of the sampling rate, while looking authoritative. Record outcomes, not steps; and let the field vocabulary of these events be a governed, bounded registry rather than an open bag of ad-hoc keys, because the learning system downstream is only as good as the dimensional consistency of the events it consumes. Every dimension the learner will ever want to slice by must be present on the event at emit time; a dimension cannot be added retroactively to events that already fired.

Published versions are immutable and content-addressed. When a campaign version goes live, the system freezes exactly what shipped: the graph structure, the resolved content of every template (including everything transitively referenced), pinned versions of every external reference (prompts, assets), and the values of every variable resolved at publish time - bundled, hashed, and stored, with the hash serving as the genotype's fingerprint. Two publishes with identical content produce identical hashes; content-hash equality is a machine-checkable answer to "did anything change between these versions?" Running cohorts pin their frozen bundle; mid-flight mutation is forbidden - not as an engineering convenience but as the ceteris paribus law made structural, because a test whose stimulus changed mid-run produced no information. This is the laboratory notebook property: every fitness reading in the system attaches to an exactly reconstructible stimulus.

Records are bi-temporal. Every substrate event carries two timestamps: when the event occurred in the world (occurred_at) and when the system recorded it (recorded_at). For real-time capture they coincide; for backfills, corrections, and late-arriving webhooks they diverge, and the divergence is information. Bi-temporality unlocks three orthogonal query axes - the filter axis ("what happened in this window?"), the audit axis ("what did we know as of this moment?"), and the lineage axis ("how has our understanding of this evolved?") - and Section 6 will show that the audit axis is what makes it possible to evaluate past decisions against what was actually knowable at decision time rather than against hindsight.

Only on top of this spine do stochastic components earn a place, and each one enters behind a typed contract at a named seam: a classifier that maps raw replies into the outcome taxonomy; a generator that produces content candidates for a template slot; a judge that pre-screens candidates; a router that reads subtext a regex cannot. The contract discipline is what keeps the system analyzable: when a stochastic component misbehaves, it misbehaves inside a boundary, visible in the telemetry of that boundary, replaceable without touching structure. Intelligence is added to the spine, never instead of it.

3. The anatomy of self-improvement: nine questions

Strip away the technology and every self-improving system - a thermostat, a gradient-descent trainer, a century of keyed coupons, the layered machine this paper describes - is an assignment of answers to the same small set of questions. Making the questions explicit does three jobs: it gives a complete checklist for designing a loop, a diagnostic for debugging one, and a precise way to say what distinguishes one layer of sophistication from the next (each layer is a more ambitious answer-set). Any nuance of a self-improving campaign system can be located under one of these nine.

Q1. Variation - what is permitted to change? The choice set. Scoring weights only? Selection among pre-authored content variants? The content itself? The targeting frontier? The journey graph? The doctrine that generates all of the above? The breadth of the variation surface is the single biggest determinant of both the system's ceiling and its risk. A loop's power and its blast radius scale together.

Q2. Observation - what does the market say back, and is the telemetry faithful? The sensor suite. Which events are captured, at what grain, with which dimensions, on which stream, deduplicated against which identity resolution? Every downstream inference is bounded above by observation fidelity, and the classic corruptions live here: machine-generated opens counted as human attention, duplicate contacts counted as independent samples, replies classified into the wrong taxonomy bucket. No learner outruns its sensors.

Q3. Evaluation - what is fitness? The objective function, and its distance from revenue. Every campaign metric sits in a proxy cascade - open → reply → positive reply → meeting booked → meeting held → qualified → closed → retained - and each step left of "closed" is cheaper, faster, and more corruptible. The governing rule from the tradition: designate one north-star metric per campaign, as revenue-proximal as the volume allows, and treat everything upstream as sub-metrics that are diagnostic but never optimized in isolation, because sub-metrics live inside a non-linear system and maximizing one routinely destroys the whole (the clickbait subject line that lifts opens and kills replies). Fitness is a design decision, and Goodhart's law is its standing adversary.

Q4. Attribution - which component receives the credit or blame? The assignment problem. A closed deal in month four touches a dozen stimuli across three channels; a scoring model, a subject line, a follow-up cadence, and an offer all claim the win. Attribution is the hardest question of the nine in high-value, long-cycle markets, hard enough to earn its own section (Section 6). A system that varies many things but attributes crudely will confidently learn the wrong lessons.

Q5. Update - what mechanism converts evaluation into change? The learner itself: a human editing a config; a bounded arithmetic rule nudging a weight; a bandit reallocating traffic; an embedding model redrawing a similarity frontier; a language model drafting the next variant; a structured amendment pass proposing a diff to doctrine. The mechanism must be matched to the evidence mass available - the deepest recurring design error is deploying a data-hungry learner in a sample-starved regime.

Q6. Cadence and blast radius - how often may change occur, and how much at once? The step-size discipline. Per-send, per-batch, per-cohort, per-quarter; one variable or many; clamped or unbounded. Two of the tradition's laws live here. Ceteris paribus: change one needle-mover per cycle and hold every other variable as a named constant, or the cycle produced zero information. Do nothing during the run: a test observed early, inside its latency window, or "improved" mid-flight is a discarded test. Cadence discipline is what separates a learning system from a thrashing one.

Q7. Constraint - what must never change, no matter what the optimizer wants? The inviolable region: consent and suppression lists, regulatory boundaries, brand voice floors, deliverability budgets (bounce and complaint ceilings), spend caps, and the structural invariants of published versions. Constraints are not penalty terms to be traded against fitness; they define the feasible region the optimizer operates inside. A self-improving system without hard constraints will eventually discover that the highest-scoring action is one you never intended to permit.

Q8. Memory - where does the learning live, and does it survive vendor change? The substrate question of Section 2.2, elevated to a design axis. Learning can live in a platform setting (dies with the vendor), in a model's weights (opaque, unauditable, hard to correct), in an operator's head (dies with the operator), or in versioned substrate (durable, diffable, transferable, and readable by every future human and agent). Only the last compounds. The memory answer also determines whether learning transfers - across campaigns, clients, and channels - or must be re-purchased with fresh samples each time.

Q9. Exploration - how is novelty generated, and what budget may it burn? Selection without variation converges and then stagnates precisely as the market moves on; variation without a budget burns the audience. Every loop needs a declared exploration policy: what fraction of traffic tests challengers, how challengers are generated (human creativity, systematic recombination of proven winners, model generation), and - crucially in finite outbound audiences - an accounting that recognizes the true unit of exploration cost is not dollars but prospects consumed, which leads directly into Section 4.

The nine questions are orthogonal design axes. Two systems with identical models can occupy opposite corners of the space; a "less intelligent" system with faithful observation, honest attribution, tight cadence discipline, and durable memory will outperform a frontier-model system that is sloppy on Q2, Q4, Q6, and Q8 - not occasionally, but as a rule.

4. The economics of learning

Machine learning's default mental model assumes data is abundant and the bottleneck is compute or algorithm. High-value B2B marketing inverts every term of that assumption, and the inversion is so complete that it deserves its own section before any layer is described, because it determines which layers a given operation may honestly run.

4.1 The sample-starved regime, and the audience as the budget

Consider the arithmetic of a typical high-ticket B2B motion. The total addressable market that survives a serious ideal-customer-profile filter is often measured in the low thousands of accounts - sometimes hundreds. A relationship-led, compliance-clean outbound program touches a few hundred prospects a quarter. Reply rates in single-digit percentages yield tens of responses per cohort; positive replies yield a handful; closed deals, with a three-to-twelve-month sales cycle, yield single digits per year. Meanwhile the statistical machinery of casual "growth hacking" - the A/B test with a significance calculator, the bandit, the trained response model - assumes hundreds of conversions per arm.

Three consequences follow, and they are the spine of honest system design in this regime.

First: the scarce resource is prospects, not compute, and every test spends them irreversibly. A consumer app can rerun an experiment on tomorrow's traffic; an outbound program that sends a weak variant to three hundred of its two thousand qualified accounts has consumed fifteen percent of a finite, slowly-renewing audience - and, worse than consuming attention, has conditioned it (Section 4.2). Exploration budgets must therefore be denominated in audience, and the cost side of every learning decision must include the option value of the untouched prospect. This also explains a design pattern that looks like timidity and is actually optimality: gating expensive and irreversible pipeline stages behind cheap reversible ones, running enrichment and outreach only on the scored shortlist, metering every costly operation against an explicit budget with a durable ledger. The pipeline's economics are part of the learning system.

Second: priors must do the work data cannot. When n is small, the correct response is not to abandon inference but to change its character: shrink estimates toward informed priors, pool evidence hierarchically across segments and campaigns (a signal's effect in one vertical is a prior, not a verdict, for the next), decay old evidence explicitly, and - the genuinely new instrument of the current era - use large language models as compressed priors over human response. A frontier model has absorbed a substantial fraction of recorded human persuasion; asked to rank ten subject lines, critique an offer against a sophistication stage, or predict the objection a CFO raises to a mechanism claim, it supplies judgment that would otherwise cost hundreds of samples. The discipline is to treat this as what it is - a prior, useful for pruning the candidate set and improving the starting variant, never a substitute for market contact. The market remains the only ground truth; the model narrows what you ask the market.

Third: below a volume threshold, the honest learner is a human with instrumentation. This deserves to be stated without embarrassment, because the industry's marketing pretends otherwise. At tens of outcomes per cohort, no statistical update rule extracts more signal than a competent operator reading every reply - provided the system makes that human learning cheap to perform, legible to record, and durable to store. The machine's highest-value role in the sample-starved regime is not to replace judgment but to be its exoskeleton: classify and route every response, surface the anomalies, keep the constants constant, enforce the observation gates, and write what the human concluded into substrate where it compounds. The layer model's lower rungs are designed around exactly this: graceful degradation to human judgment, with the machine as amplifier and archivist.

The initial-conditions corollary belongs here too. Because each loop cycle is expensive in audience and latency, the quality of the first variant matters enormously: a campaign whose opening genotype is grounded in real niche research, a correct sophistication diagnosis, and competitor teardown reaches its target in a fraction of the cycles of one that "just starts sending." Cheap starting variants are the most expensive item in the sample-starved regime.

4.2 Non-stationarity, reflexivity, and the Red Queen

The second economic reality: the function being learned does not hold still, and - this is the part naive optimization misses - your own success is what moves it.

Schwartz's sophistication model, restated as dynamics: a market's response to a class of claims degrades monotonically with cumulative exposure to that class, across all senders, and your winning campaign contributes to the very saturation that will kill it. The five stages are diagnosable states with distinct winning strategies and distinct observable signatures:

Two related decay phenomena compound the stage dynamics. Habituation (call it stimulus numbness): repeated exposure of a niche to structurally similar stimuli produces non-response even when each instance is competently made - the "proven template" of eighteen months ago now pattern-matches to noise. Conditioning: worse than numbness, a niche burned by a genre of outreach responds to that genre with active hostility regardless of the individual message's quality; the correct move is to exit the burned genre entirely, not to write it better. And decay is cyclical: abandoned genotypes recover effectiveness as niches forget, which creates a contrarian window for the operator who re-enters a recovered pattern before consensus does.

For the learning system, four design consequences:

Evidence must decay. Observations carry a half-life; a signal's predictive weight and a variant's fitness reading from four quarters ago are priors at best. Exponential recency decay on both evidence and creative-performance estimates is not a tuning nicety but a model of the world.

Winning variants must be treated as depreciating assets. Plan the successor before the incumbent dies; maintain a portfolio of live angles rather than converging the whole population on the current champion, because convergence maximizes exactly the exposure that accelerates decay. In evolutionary terms, this is the Red Queen's regime - continuous adaptation is the price of constant relative fitness - and diversity is the hedge.

Stage transitions are detectable, and detecting them is a first-class learning objective. The signatures in the table's right column are queryable against a faithful outcome stream: declining response to benefit-led variants alongside stable response to mechanism-led ones is the Stage-2→3 boundary announcing itself. A sophisticated system does not merely select today's winner; it monitors the relative slopes of strategy classes and flags the regime change - because the correct response to a stage boundary is not iteration within the dying class (a Layer 2 move) but a strategy-class jump (a Layer 4–6 move).

Latency windows are sacred. Every channel has a lag between deployment and interpretable outcome - days for email with a full follow-up sequence, weeks for paid at the offer layer, a quarter for organic - and distributions are clumpy: a healthy system can throw a run of losses well inside normal variance. Most working systems are killed by their operators inside the latency window, panicking at noise. The loop's cadence discipline (Q6) exists to make that impossible: minimum sample floors per channel, minimum observation windows, and a hard prohibition on mid-run changes are encoded as structure, not left to operator nerve.

One counterweight before this section closes, developed fully in Section 9: reflexivity has a constructive twin. The same causal arrow that lets your success erode the response landscape lets it enrich the input pool - every delivered result densifies the proof your next campaign carries, every advocate warms the population your next cohort draws from, and persistent delivered assets keep routing self-selected prospects back long after their cohort closed. A complete theory holds both: adversarial reflexivity on the content you show the market, beneficial reflexivity on the evidence and warmth your delivery manufactures. The first you hedge with portfolios; the second you build with flywheels.

4.3 The moving bottleneck

A chain converts at the rate of its narrowest stage - the bottleneck principle, familiar from the theory of constraints and already embedded in Layer 2's test-the-weakest-stage rule. The deeper law, visible only when the whole journey is in view, is that the binding constraint migrates as the system scales, and it migrates in a predictable pattern: it sits, at any moment, at whichever stage still runs at human bandwidth. Early, that is almost always the conversion stage - a founder's calendar caps closed deals at a hard ceiling regardless of how much qualified volume acquisition produces, and a tenfold improvement anywhere else in the chain is invisible at the north star. Relieve that constraint - a hire, an automation, a delegation - and it does not vanish; it moves: to onboarding capacity, then to delivery hours per customer, and eventually, once automation has lifted the human-bound middle, back upstream to sourcing and creative throughput.

Two disciplines follow, and both are budget rules for the scarce resources of Section 4.1. First, improvement effort targets the currently binding constraint, only. Tests, tooling, hires, and layer upgrades pointed at a non-binding stage produce moving sub-metrics and zero throughput - and pre-relieving a future constraint, however confidently forecast, is the sophistication trap wearing a capacity-planning costume. Second, capacity expansions gate on evidence, not calendar. The next hire, the next automation layer, the next tool is authorized when the constraint is measurably binding - the threshold crossed and sustained - not when the plan predicted it would be. A system that expands on schedule rather than on signal amplifies whatever bottleneck actually exists instead of relieving it.

The synthesis of this section in one sentence: in high-value markets, the learning system's binding constraints are audience and time, its ground truth is scarce and perishable, its adversary is a response landscape that its own victories erode, and the correct target of the next unit of improvement effort is always the constraint that is binding now - and every layer in the next section is an answer to the question "given all that, what may honestly vary, and what machine may honestly vary it?"

5. The layer model

Self-improvement is not one capability. It is a stack of seven, ordered by what they are permitted to vary (Q1), the evidence they require (Q2–Q4), the machinery they employ (Q5), and the governance they demand (Q6–Q7). Each layer strictly contains the ones below it: a Layer 4 system still runs Layer 0's instrumentation and Layer 1's weight discipline. The ordering is also a deployment sequence - a maturity path - and skipping rungs is the most reliable way to build a confident, expensive, wrong machine.

A one-table overview, then each layer in depth.

Layer 0 - Instrumented Execution: the measurement substrate

Layer 0 varies nothing and learns nothing, and it is the most important layer in the stack, because every layer above it is bounded by its fidelity. It is Section 2.3's deterministic spine, viewed as the learning system's sensory apparatus: the campaign as a versioned DAG; the single governed outcome taxonomy with a classifier mapping every raw response into it; one transition event per edge traversal, emitted at a chokepoint, carrying the full genotype fingerprint (structure version, content hash, targeting version, weights version) and every dimension analysis will ever slice by; the unsampled business-event stream kept categorically separate from sampled operational telemetry; identity resolution and deduplication so that one human is one sample; content-addressed immutable publish snapshots so every fitness reading attaches to an exactly reconstructible stimulus; bi-temporal timestamps on everything; suppression and consent enforced as pipeline stages, not afterthoughts.

Two Layer 0 disciplines are so frequently skipped that they warrant emphasis. The cohort is the unit of statistical validity: a bounded set of prospects run against a single frozen genotype under declared constants for the channel's minimum sample size and latency window. Learning reads cohorts, never live dribbles. Iteration happens at cohort boundaries - a new version opens a new cohort - and hot-patching a running cohort is structurally forbidden. The pre-flight declaration: before a cohort launches, the hypothesis, the varied needle-mover, the enumerated constants, the north-star threshold, the sample floor, and the latency window are written down. A test without a pre-registered hypothesis is a story you will tell yourself afterward.

The layer's own maxim: a campaign that cannot be replayed cannot be improved. Everything above is inference; Layer 0 is evidence.

Layer 1 - Parametric Adaptation: the weights move

The first layer at which the system changes itself, and deliberately the humblest. What varies: scalar multipliers on the signals of a scoring model - the model that ranks discovered prospects by fit and buying-signal strength and thereby decides who receives attention next. Nothing else moves: not the lead sources, not the ICP filters, not a word of copy, not the graph. The visible effect of learning is a re-ranked queue.

The machinery is three feedback lanes of ascending automation, and the triad matters more than any one lane:

The manual lane. An operator sets a bias directly - "weight the funding-event signal up" - as a versioned configuration change. No statistics, no model; a hand on a dial, recorded in substrate with a reason.

The judgment lane. Humans render fast qualitative verdicts on individual leads - a one-word pursue / reject / wrong-fit in a review surface - and a language model acts as interpreter (its first and most modest role in the stack), mapping free-text verdicts onto structured signal feedback. Repeated consistent rejections along one dimension raise an ICP-drift alert: a suggestion routed to a human, never an automatic filter change. In the sample-starved regime this lane does the real work.

The statistical lane. At volume thresholds, a deterministic rule computes, per signal, the outcome rate among prospects carrying that signal against the baseline across all prospects, and nudges the signal's multiplier by a small fixed step in the direction of the lift - clamped to a floor and ceiling so no signal can be zeroed out or come to dominate, with recency decay so stale evidence fades. This is a bounded cousin of the multiplicative-weights update family - one of the oldest and best-understood algorithms in learning theory - chosen precisely because its failure modes are legible: small steps mean one noisy batch cannot swing the model; clamps mean a spurious correlation cannot capture it; the update is arithmetic a human can audit in a spreadsheet.

The honest volume caveat is a feature of the design, not a flaw to hide: the statistical lane's activation threshold (on the order of a thousand sent-and-resolved outcomes) may simply never arrive inside a low-volume engagement, in which case the manual and judgment lanes are the learning system, and the statistical lane is documented, dormant capacity. Demonstrating that the loop closes - engagement events flow in, the feedback stage runs, a score visibly moves, the queue re-ranks - is a different and much cheaper claim than demonstrating statistical convergence, and in most high-value engagements it is the only claim worth making.

Graduation criterion to Layer 2: a stable Layer 0, a scoring model whose re-rankings the operators demonstrably trust, and enough volume that at least one channel can fill an experimental arm inside a tolerable window.

Layer 2 - Experimental Selection: choosing among variants

What varies: which of several pre-authored, human-written content variants gets the traffic - subject lines, opening angles, call-to-action framings, send times, channel order for a segment - always within a fixed structure, always one needle-mover per test with everything else a declared constant. Layer 2 is Hopkins's keyed coupon industrialized: the system now runs the tests, allocates the traffic, and enforces the discipline humans reliably break.

The machinery, in order of increasing sophistication: fixed-horizon A/B tests with pre-registered sample floors and sequential analysis rules (so that "peeking" at interim results is mathematically licensed rather than silently corrupting); multi-armed bandits - Thompson sampling being the workhorse: maintain a posterior over each variant's rate, sample from the posteriors, send the sampled winner, update on the outcome - which reallocate traffic toward winners during the run and thereby reduce the audience cost of exploration, the binding cost of Section 4.1; and standing holdout slices that receive the incumbent (or, periodically, nothing) as the permanent ground-truth anchor against which all modeled claims are audited.

Layer 2's laws are inherited directly from the tradition and enforced as structure: sample floors set by the channel's noise floor (representative floors: a few hundred fully-sequenced sends per email variant; a hundred connections per social-DM variant; a hundred dials per call script; four figures of impressions per paid variant; double-digit posts over multiple weeks per organic style - the exact numbers are configuration, the existence of a floor is doctrine); latency windows respected to completion; a bottleneck-first rule that directs testing at the weakest stage of the state chain (testing subject lines while replies are the bottleneck is optimizing sub-system one when sub-system two is broken - the metric moved will be the one that doesn't matter); and the north-star rule that no sub-metric is ever the objective.

The bandit caveat for non-stationary worlds: classical bandit guarantees assume fixed arm rates, and Section 4.2 says rates decay - winners habituate their audience. Practical remedies (sliding-window posteriors, discounted updates, scheduled re-exploration of retired arms to catch the contrarian recovery window) are standard, but the deeper posture is portfolio thinking: the bandit picks today's traffic split; it does not absolve the operator of maintaining tomorrow's challenger population.

Graduation criterion to Layer 3: the variant library and the segmentation both start to outgrow hand curation - you have more plausible variants than test slots and more plausible sub-audiences than named segments - which is precisely the problem representation learning exists to solve.

Layer 3 - Semantic Representation: the system learns what things mean

Layers 1 and 2 treat prospects and messages as rows with categorical features. Layer 3 gives the system a geometry: dense vector embeddings of companies, people, messages, and replies, in which semantic similarity becomes measurable distance. What varies as a result is primarily the sourcing frontier - which prospects enter the pipeline at all - plus the routing of responses and the discovery of structure the ICP's author didn't articulate.

The machinery and its uses: Lookalike sourcing - embed the closed-won customers, search the universe for nearest neighbors, and let the discovery stage expand along the similarity frontier rather than only along explicit firmographic filters; the ICP stops being a static filter and becomes a learned region that moves as wins accumulate. Semantic ICP matching - score a prospect's website, news, and hiring signals against the ICP definition by meaning rather than keyword, catching the company that never uses your industry's vocabulary but exhibits its exact shape. Emergent segmentation - cluster the embedded prospect base and let sub-audiences announce themselves, then test Layer 2 variants per cluster, which is how "this angle wins with operations-led buyers and loses with finance-led ones" becomes discoverable rather than folklore. Reply intelligence - classify inbound responses into the outcome taxonomy with calibrated granularity, detect intent and objection type, route the "not interested but how does it work" reply down its own edge; embedding-based similarity also powers semantic suppression (recognizing that two differently-spelled contacts are the same person, or that a "new" prospect is a subsidiary of a client) and drift detection - monitoring the embedded distribution of who is entering, replying, and converting, and alerting when the population shifts under the model.

Layer 3's characteristic risk is the filter bubble, and it is the exploration problem in geometric clothing: a sourcing frontier expanded only toward past winners converges on a region and then over-fishes it, systematically blind to adjacent regions that would convert - and the blindness is invisible in the metrics, because the metrics only measure the region being fished. The defenses are explicit: reserve a budgeted slice of sourcing for out-of-region sampling; audit clusters with human eyes before they become targeting policy; and treat similarity-to-winners as one signal in the score, never the gatekeeper of discovery.

Graduation criterion to Layer 4: the variant demand from per-cluster testing exceeds human authoring capacity - the system now knows who to say things to at finer grain than anyone can write for by hand.

Layer 4 - Generative Variation: the machine drafts

What varies: the content itself. A language model, conditioned on the substrate - the ICP definition, the offer doctrine, the messaging framework, the sophistication-stage diagnosis, the prospect's enrichment record, and the library of past winners and losers with their fitness readings - generates candidate copy for a template slot. This is the layer the market calls "AI marketing," and it is fourth, not first, because everything it needs to be safe and useful is supplied by the layers below: structure to fill (the graph and template slots are fixed), constraints to obey (brand, claims, compliance as hard filters), a taxonomy to be measured against, and an experimental harness (Layer 2) that remains the only arbiter of what actually ships at scale.

The machinery, as a pipeline of roles - and role-separation is the design insight: the generator produces candidates inside a constrained prompt that pins the structural grammar (a hook–story–offer skeleton; a mechanism-led opening for a Stage-3 market; a value-equation-complete offer statement), retrieves the relevant substrate, and explicitly injects the needle-mover being varied while holding the rest verbatim from the incumbent - ceteris paribus enforced at generation time. The judge - a second model pass with a different prompt and ideally a different model - pre-screens candidates against a rubric (claim accuracy against the substrate, brand-voice conformance, sophistication-stage fit, banned-pattern detection), functioning as a cheap prior filter that spends tokens to save audience. The editor is human: at low volume every shipped candidate passes an operator's taste; at higher volume the human curates the exemplar library and the rubric rather than every instance - the altitude of human involvement rises with evidence, never disappears. Then Layer 2 tests what survives, and the fitness readings flow back into the exemplar library, closing the generation loop.

Two genuinely evolutionary operators become available here, mechanizing what master operators always did by hand: splicing - extract the trait that made a winner win (a transparency frame, a pacing pattern, a proof structure) and inject it into a different stimulus's genotype - and cross-pollination across channels, where a hook proven in one channel becomes the seed variant in another. Winning traits, not winning strings, are the unit of inheritance; the substrate's job is to store them as named, cited patterns rather than as folklore.

Layer 4's characteristic risks are self-referential. Model collapse: a generator trained or few-shotted predominantly on its own surviving outputs drifts toward a homogeneous house style - which then habituates the market faster (Section 4.2) precisely because it is uniform; the defense is deliberate injection of exogenous variation (human drafts, out-of-domain exemplars, scheduled style resets). Judge sycophancy and style bias: LLM judges systematically prefer LLM-shaped text, so a judge is calibrated against market outcomes - its pre-screening verdicts are themselves scored for predictive validity against what actually converted - or it is quietly optimizing for its own aesthetic. Substrate contamination: generated inferences must never be written back into doctrine as fact (the trust-laundering rule of Section 2.2); a model's summary of why a variant won is a hypothesis until the amendment protocol of Layer 6 admits it with evidence.

Graduation criterion to Layer 5: content-level variation reaches diminishing returns within the current structure - the copy is strong and the leak is demonstrably in the journey shape itself: the cadence, the branch logic, the channel mix, the sequence length.

Layer 5 - Structural Evolution: the graph itself mutates

What varies: the campaign DAG - edge rewiring (where does this outcome route?), branch pruning and grafting, follow-up cadence and sequence length, channel ordering and mix, the insertion of new node types (a mid-journey enrichment refresh; a human-escalation gate moved earlier), and, at the frontier, per-prospect next-best-action policies that choose the following touch dynamically. This is the highest-blast-radius variation surface in the operational system: a structural change alters the denominator of every downstream metric simultaneously, which is exactly why it sits above four layers of measurement and attribution machinery rather than below them.

The machinery is the analytical heavy artillery, and Section 6 treats its foundations; here, the inventory. Structural diagnosis: per-node drop-off and per-edge routing-mix analysis over the transition-event stream identifies where journeys leak and which outcomes dead-end. Multi-touch attribution (position-based heuristics graduating to Shapley-value and Markov removal-effect models as volume permits) apportions a conversion's credit across the touches that preceded it, converting "the sequence works" into "touch three's mechanism email and touch six's channel-switch carry the sequence." Survival analysis models time-to-conversion as a hazard function, which is how a system reasons honestly about six-month sales cycles - reading censored data instead of waiting for it - and how cadence changes are evaluated by their effect on the hazard curve rather than on quarter-end tallies. Uplift modeling asks the only causally honest question about any touch: not "did converters receive it" but "did it change the probability for those who received it versus statistically identical prospects who did not" - the analysis that finds the touches you could delete, or that actively suppress (the follow-up that triggers unsubscribes among a segment that was converting anyway). Off-policy evaluation (inverse-propensity and doubly-robust estimators over logged decisions, which is why Layer 0 must record the score and propensity at decision time, not just the action) estimates how a candidate structure would have performed on historical traffic before any prospect is spent on it - with the sober caveat that at B2B volumes its confidence intervals are wide, so it serves as a pruning filter for structural candidates, not a verdict. And simulation: persona-conditioned language models role-playing the prospect against a candidate journey, which is Layer 4's "LLM as prior" promoted to the structural altitude - genuinely useful for catching gross defects (a cadence that reads as harassment; a branch that strands an objection) and strictly forbidden from being confused with ground truth; sim results are priors, the market is the posterior, and the sim-to-real gap is itself measured.

Structural evolution operates under the strictest inheritance of the tradition's laws: structures version immutably, one structural mutation per cohort comparison, sample floors an order of magnitude above content tests, and a sunset protocol as the terminal phase - a structure that fails its economic viability criterion (e.g., cannot fund its own acquisition cost) across a defined number of consecutive cohorts, or that has been iterated to the top of the sophistication ladder without recovery, is archived, not nursed, its version history preserved for the record and its audience released. Refusing to sunset - nursing a dead structure because the next variant will surely fix it - is a category error the system makes structurally impossible: the cycle already produced its information, and the information was no.

Graduation criterion to Layer 6: structural learnings start rhyming across campaigns - the same cadence discovery in two verticals, the same channel-mix result for the third client - which means the knowledge wants to live above any single campaign, in the doctrine that generates them all.

Layer 6 - Doctrinal Self-Amendment: the loop closes over the generator

Every layer below improves artifacts: a weight, a traffic split, a frontier, a draft, a graph. Layer 6 improves the generator of artifacts - the organization's codified doctrine: the ICP definition, the offer architecture, the messaging framework, the sophistication-stage diagnosis per segment, the outcome taxonomy itself, the channel playbooks, the KPI thresholds, and - reflexively - the learning policies of Layers 1 through 5 (the clamps, the floors, the exploration budgets, the graduation criteria). This is the layer at which the phrase self-improving stops being a metaphor, because the "self" that improves is the organization's externalized knowledge, and the improvement is durable in exactly the way Section 2.2 demands: it lands in versioned substrate.

The structure is a tangled hierarchy - Douglas Hofstadter's strange loop, instantiated at the organizational altitude. The doctrine (top level) generates campaigns (bottom level); campaigns contact the market; outcomes flow back and amend the doctrine that generated them; the amended doctrine generates the next campaigns. Causation runs both down and up the hierarchy, and once the corpus is versioned, structured, and consumed by agents at every decision, it acquires causal force over its own authors: the operator is held to the doctrine the operator wrote; the next campaign inherits every prior campaign's lessons by construction rather than by memory. Where the selection loop (Layers 1–5) scales the business through contact with reality, the amendment loop scales the substrate that processes reality. They compound through each other: every selection result worth keeping becomes an amendment; every amendment becomes the next testable hypothesis.

The machinery is an editorial protocol, not an algorithm, and its discipline is what separates a compounding corpus from a bloated one:

The amendment pass. When an outcome falsifies a doctrinal claim - the funnel math was five times optimistic; the mechanism claim has crossed into Stage-4 fatigue; the ICP's size floor is excluding the segment that actually converts - a structured pass (drafted by a model, gated by a human) names the triggering evidence, walks the dependency graph of every artifact whose meaning the change touches, proposes a proportional amendment to each under its mutation rule, and executes as a single coherent wave. The alternative is the knowledge system's deepest failure mode, phantom canon: multiple documents claiming truth, none consistent with the latest evidence, every downstream consumer - human or agent - retrieving a stale belief with full confidence.

Mutation rules per artifact class, heterogeneous on purpose. Decision records are append-only with amendment logs - history is walkable, and a decision is a fact about a moment. Doctrine notes rewrite freely - they are current truth, with git as the audit trail. Prompts and templates version - executable artifacts pin what ran when. Reports are immutable - evidence is never edited, only superseded. Conflating these rules destroys the corpus's integrity in different directions at once (edited evidence corrupts audits; append-only doctrine becomes archaeology).

The admission standard: empirical grounding and via negativa. A candidate principle enters canon only with the instances that produced it attached, and the standing question is subtractive - what should the corpus not contain? A "principle" whose violation produces no observable failure is decoration, and decoration dilutes the force of the principles that bind. The corpus optimizes for coherence, not exhaustiveness; exhaustiveness decays, coherence compounds.

Meta-learning, soberly scoped. The slowest loop tunes the loops: reallocating exploration budget toward the layer currently producing the most information per prospect spent; adjusting statistical-lane thresholds as volume changes; retiring a graduation criterion that proved premature. And cross-campaign transfer becomes a first-class operation: hierarchically-pooled evidence lets a new campaign in an adjacent vertical start from the corpus's priors rather than from zero - the substrate is what makes sample-starved learning cumulative across engagements instead of purchased anew each time. The lineage axis of the bi-temporal substrate (Section 2.3) is this layer's native query surface: how has our belief about this segment's sophistication stage evolved across the last four cohorts, and which amendments correlated with downstream lift is a literal, executable question.

The human role at Layer 6 is constitutional editor. Models draft amendments, walk dependency graphs, and surface the aside in a working session that deserves canonization (the insight said in passing is routinely worth more than the decision the session was about - a corpus that captures only deliberate decisions learns at a fraction of its potential rate). Humans approve what becomes doctrine, and certain invariants are placed beyond any amendment's reach. The failure modes are the ones self-reference always invites - over-canonization (every clever phrase a "principle"), bureaucracy creep (every micro-change an amendment ceremony, until the cost of deciding exceeds the value of the decision), the museum corpus (revered, unfalsifiable, drifting from lived reality), and the sophistication trap's final form: polishing the doctrine while the campaigns it exists to improve go unshipped. The governing test for the whole layer is blunt: if the corpus is growing while the loop beneath it is not firing against the market, the system is writing literature, not learning.

6. Credit assignment and time

Attribution deserves its own section because it is the hardest problem in the stack and the one most often waved away. Every layer's update mechanism (Q5) is only as honest as its answer to Q4, and in high-value B2B the structure of the domain conspires against easy answers: rewards are delayed by months, journeys are multi-touch and multi-channel, sample sizes forbid brute-force randomization of everything, and the counterfactual - what would this prospect have done absent the touch? - is unobservable by definition.

The practical doctrine is an attribution ladder, climbed only as evidence permits. At the bottom, cohort-level comparison - genotype A's cohort versus genotype B's cohort on the north-star metric, with constants held - which is Layer 2's workhorse and requires no per-touch causal claims at all; when in doubt, this is the rung to stand on. Next, heuristic multi-touch models (first-touch, last-touch, position-weighted, time-decayed), cheap and directionally useful, each embedding an assumption that should be stated aloud rather than absorbed silently. Then data-driven apportionment - Shapley values over touch coalitions, Markov-chain removal effects - which distribute credit by marginal contribution but demand volumes that most high-ticket motions only reach when evidence is pooled across campaigns (a Layer 6 service to Layer 5). Alongside, survival analysis to handle delay honestly - modeling the hazard of conversion over time so that six-month cycles inform decisions six months before they close, and so that censored prospects (still in flight) contribute information rather than being dropped or, worse, counted as failures. Above that, uplift modeling, the only rung that speaks strict causal language about individual touches, reserved for the decisions that justify its cost. And at the top, permanently, the holdout: a standing slice receiving the incumbent or nothing, the court of final appeal that every model on the ladder is periodically audited against, because models drift and holdouts do not.

Three disciplines make any rung of the ladder trustworthy.

Log the decision context, not just the action. Off-policy evaluation, uplift, and even honest post-mortems require knowing why the system did what it did: the score, the propensity, the model version, the variant-selection probability at the moment of choice. An action log without decision context is a diary; with it, it is a laboratory record from which counterfactuals can be estimated.

Judge decisions against the information set at decision time. This is the audit axis of the bi-temporal substrate doing its real job. A scoring decision made in March is evaluated against what was recorded as of March - not against the enrichment data that arrived in May, not against the correction backfilled in June. Systems that evaluate past policies against present knowledge systematically flatter hindsight, learn spurious lessons, and punish decisions that were correct under uncertainty. occurred_at versus recorded_at is the two-field schema that makes honesty mechanical.

Attach fitness to genotype fingerprints, never to names. "The Q3 sequence" is not an experimental unit; content-hash 9f31… at structure version 4 under targeting version 2 with weights version 7 is. The content-addressed snapshot discipline of Layer 0 exists precisely so that every outcome event carries the exact, immutable identity of the stimulus that produced it - which is what makes results reproducible, comparisons valid, and the corpus's claims about "what worked" auditable years later.

Two further principles complete the attribution picture, and both come from taking the whole journey seriously rather than the acquisition segment alone. First, signal value rises with funnel depth. The scarcest events - closes, churns, renewals, expansions, referrals - carry the most information per event about every upstream choice, because they are the outcomes the shallow metrics merely proxy. The practical discipline is depth-weighted re-scoring: on a slower cadence than stage-local optimization, the system re-evaluates its upstream policies - sourcing weights, enrichment variables, reigning copy champions, even the ICP's boundaries - against the deepest outcome the accumulated volume can support. A copy variant that wins on replies and loses on closes is a loser; a sourcing signal that predicts opt-ins but not revenue is noise wearing a correlation; and the deep events, precisely because each one is precious, are propagated upstream as re-weighting signals to every stage that contributed to producing them. Second, the fitness window is a declaration, not a default. Any bounded evaluation window structurally underprices genotypes whose value accrues in tails - a delivered asset that keeps routing inbound interest for years, a relationship deposit, a brand impression - and a challenger that edges the incumbent on a thirty-day metric can be the wrong strategic choice against an incumbent with a compounding tail. The defense is pre-registration: declare the window in the pre-flight, name any known tail the window cannot see, and pre-commit the modifier ("the challenger must win by more than X to displace an incumbent with a measured asset tail") before the results arrive, so the evaluation cannot be re-argued after the fact by whoever prefers its outcome.

One closing calibration for the sample-starved regime: the correct response to attribution difficulty is not attribution nihilism but humility with structure. Prefer the lowest rung that answers the live question; state the assumption each model embeds; pool across campaigns before reaching for data-hungry rungs; and let the holdout, not the model, have the last word.

7. Governance and guardrails

A system licensed to change itself needs a constitution - not as compliance theater, but because every capability in Section 5 has a failure mode in Section 8, and the difference between the two is almost always a guardrail. The governance model has five pillars.

Hard constraints define the feasible region; the optimizer lives inside it. Consent and suppression (an unsubscribe is instant, global, and permanent across every channel and tool); regulatory boundaries per jurisdiction; deliverability budgets (bounce and complaint-rate ceilings treated as circuit breakers that halt sending, because sender reputation is a slow-to-earn, fast-to-burn asset the optimizer must never be allowed to spend); financial budgets with metered gates and durable cost ledgers at every expensive stage; and claim-truthfulness floors (no generated variant may assert what the substrate cannot support). These are not penalty terms traded against fitness. They are the boundaries of the space in which fitness is even computed, enforced at chokepoints the optimizer cannot route around.

Blast radius determines approval altitude. The human role is not a binary of "in the loop / out of the loop" but a gradient matched to consequence: weight nudges within clamps proceed automatically and are reviewable after the fact; content variants ship on judge-plus-editor approval at low volume, on rubric-plus-sampling at high volume; structural rewires require explicit sign-off with the diagnosis attached; doctrinal amendments require the evidence, the dependency wave, and a named approver. As evidence accumulates that a layer's automation is trustworthy - measured, not asserted - the human moves up an altitude: from approving instances, to approving policies, to approving the constitution. The direction of travel is earned autonomy, never assumed autonomy.

Change is legible or it didn't happen. Every self-modification the system performs - a weight step, a traffic reallocation, a frontier expansion, a shipped variant, a graph mutation, a doctrine amendment - is recorded in substrate with its trigger, its evidence, and its version fingerprints. The mutation rules of Layer 6 (append-only decisions, versioned executables, immutable reports, freely-rewriting doctrine with git history) are the corpus-level expression of this; the immutable cohort and snapshot rules of Layer 0 are the runtime-level expression. Silent drift - the config edited in place, the prompt tweaked without a version bump, the "small fix" mid-cohort - is the governance failure that makes every later question unanswerable.

Telemetry integrity is governed like an API. The two-stream separation (unsampled business events; sampled operational telemetry) is inviolable, because the learning system reads the business stream and a sampled count is a wrong count. The dimensional vocabulary of events is a closed, typed registry - a bounded budget of field names, each earning its slot as a correlation key, a slicing dimension, a measure, or a searched identifier - rather than an open bag, because dimensional drift (synonyms, case-variants, ad-hoc keys) silently fragments the evidence base the learner stands on. Personal data on the telemetry surface carries a declared class and is redacted or hashed at the single emission chokepoint under an explicit toggle - hashing rather than dropping, so per-subject correlation survives redaction.

The market's counterparty is protected by design. A self-improving persuasion system aimed at finite audiences of real people carries an obligation the optimization literature doesn't price: frequency caps and cadence floors that no fitness reading can override; a permanent human-escalation path so that automation exhausting its inventory routes to a person rather than silently dropping or endlessly re-touching a prospect; honesty constraints on generated claims; and the recognition, encoded in the sunset protocol, that a market segment signaling stop - through conditioning-grade hostility or complaint rates - is an answer to be respected, not an obstacle to be optimized around. This is not only ethics; it is Section 4.2's economics. The audience is the budget, and burning it is the one loss no layer can learn back.

8. A bestiary of failure modes

Every failure below has destroyed real campaigns and real learning systems. Each entry names the mechanism and the structural defense - because the lesson of the whole tradition is that these are defeated by architecture, not by vigilance.

Goodhart's law / proxy corruption. Any measure made a target ceases to be a good measure. Open rates polluted by machine prefetching; reply rates "won" by hostile replies; click rates won by curiosity-gap bait that poisons the meeting-held rate downstream. Defense: north-star discipline (optimize the revenue-proximal metric, treat sub-metrics as diagnostics), an outcome taxonomy that distinguishes reply valence, and periodic audits of every proxy against the metric it proxies.

Ceteris paribus violations. Two variables changed in one cycle; the result is attributable to neither; the cycle produced zero information while consuming audience and time - the single most common industry-wide waste. Defense: pre-flight declarations with enumerated constants, one needle-mover per cohort, publish-time freezing of everything else.

Premature iteration inside the latency window. A working system killed at day four of a fourteen-day window because early numbers looked bad - or "improved" because they looked good. Defense: minimum observation windows enforced by the workflow layer; iteration commands structurally unavailable until the gate passes.

The regression-to-the-mean trap. Outcome distributions are clumpy; a healthy 20%-conversion process can throw a dozen straight losses. Operators panic at the clump, change the script, and destroy the constants - and the process never recovers because the constants are gone. Defense: sample floors set by the channel's noise floor, and dashboards that display confidence bands rather than point estimates.

Sub-metric optimization with non-linear damage. Lifting a stage's conversion in a way that degrades the composition of who passes through it (clickbait improving opens while filling replies with the wrong people). Defense: evaluate every change at the north star, and instrument stage-to-stage composition, not just rates.

Vanity-metric optimization (the organic variant). Likes, impressions, follower counts trained into an algorithm that then serves the wrong audience - actively damaging acquisition while the dashboard celebrates. Defense: define the organic north star as qualified inbound; treat platform-native engagement as noise unless proven correlated.

Deliverability confounds. The "copy learning" that is actually measuring spam-filter behavior: variant A "lost" because it shipped from a cooler domain, at a worse hour, into a stricter provider mix. Defense: hold sending infrastructure in the constants stack; monitor placement independently of response; treat deliverability metrics as circuit breakers upstream of any learning read.

Habituation and conditioning misread as copy failure. Structural decay (the niche has gone numb to the genre, or hostile to it) diagnosed as a writing problem, answered with more polish inside the dying pattern. Defense: monitor strategy-class slopes (Section 4.2's stage-boundary signatures); respond to decay with class jumps and portfolio diversity, not intra-class iteration; respect the sunset protocol.

Exploration starvation and the sourcing filter bubble. The selection machinery converges on yesterday's winners - arms, regions, angles - and the metrics cannot see what is never tried. Defense: budgeted standing exploration at every layer (challenger traffic, out-of-region sourcing, retired-arm re-tests for the contrarian window).

Feedback collapse / model collapse. The generator learns from its own survivors until the population homogenizes - which accelerates habituation precisely because the market now sees one style everywhere. Defense: exogenous variation injected on schedule; exemplar libraries curated for diversity; human drafts kept in the gene pool.

Judge sycophancy and simulation overfitting. LLM judges preferring LLM-shaped text; persona simulations rewarding what simulates well rather than what converts. Defense: score the judge and the simulator against market outcomes; treat both as priors with a measured sim-to-real gap; never let a synthetic verdict ship a variant without a market-facing test path.

Attribution hindsight. Past decisions evaluated against present knowledge; policies that were correct under uncertainty punished; spurious lessons learned with confidence. Defense: bi-temporal records and audit-axis evaluation (Section 6); decision-context logging.

Fitness-window myopia (tail blindness). A bounded evaluation window crowns the variant that wins inside it and silently discards the one whose value accrues in a multi-year tail - the persistent installed asset, the relationship deposit, the compounding referral surface. The window's verdict looks rigorous and is structurally incomplete; read naively, it retires the strategic winner. Defense: declare the window and its known blind spots in the pre-flight; pre-register tail modifiers before results arrive (Section 6); instrument persistent assets with durable provenance so tails graduate from argument to measurement.

Phantom canon and silent drift. The knowledge-layer failures: multiple documents claiming inconsistent truth after an unpropagated decision; configs and prompts edited in place until behavior bears no relation to documentation. Trust in the corpus decays non-linearly - one caught lie and every future retrieval gets manually re-verified, which is the end of the substrate's value. Defense: the amendment-wave protocol, mutation rules per artifact class, and the rule that a decision's coherence pass closes the day the decision is made.

Over-canonization and bureaucracy creep. Every clever phrase a "principle" until none binds; every micro-change an amendment ceremony until decisions stop being made. Defense: empirical-grounding admission standards, via negativa passes, and a changeability model with a free tier (user-domain content needs no ceremony).

The sophistication trap. The meta-failure that eats engineering organizations: building elaborate loop infrastructure - the registry, the simulator, the attribution engine - before any campaign is producing revenue, mistaking machinery for progress. The loop runs on a spreadsheet and a calendar reminder if it has to. Defense: the layer model itself as a deployment discipline - each layer built only when the layer below is running against the live market and its graduation criterion has actually fired. Ship the loop; then compress its cycle time.

9. Past the close: one graph, one loop, the whole journey

Everything to this point has honored the tradition's oldest habit: treating the close as the terminal node - the place where marketing's mandate ends and someone else's ledger begins. The habit is a historical accident of agency org charts, not a property of the system, and correcting it completes the theory rather than appending to it. Two claims organize this section. First, the post-sale journey - onboarding, delivery, support, retention, expansion, advocacy - is structurally the same object as the acquisition campaign: a state-transition graph over humans, driven by designed touches, measured on outcome-labeled edges, and improvable by exactly the nine questions of Section 3. Second, the two segments are economically one system: the last phases manufacture the inputs the first phases consume, and at high market sophistication that coupling is not a nicety but the remaining source of differentiation.

9.1 The journey graph: the campaign, continued

Extend the campaign DAG through the close and the full object appears: one continuous graph per customer, spanning five phases - acquisition (first touch through booked-and-shown), conversion (shown through decision), closing (agreement through activation), value delivery (activation through first result and compounding results), and retention & advocacy (renewal, testimonial, referral, expansion). The acquisition graph and the post-close journey compose at the close edge - and nothing about the grammar changes on the far side of it. Nodes are still actions; edges still carry outcomes from the one shared taxonomy; versions still freeze immutably; telemetry still emits once per transition; and the value-equation review that gates an acquisition touch ("does this move dream-outcome up, or time-and-effort down, for the person receiving it?") gates a day-21 delivery milestone identically. A campaign graph was never anything but the first two phases of a journey graph, drawn by someone whose responsibility ended early.

Three properties of the extended graph deserve canon status, because each changes how the loop runs.

The graph is fractal. Any node decomposes into a child graph with its own edges, thresholds, and roles, recursively, down to atomic transitions. The conversion phase is the canonical demonstration: what a pipeline diagram renders as one "sales call" node is, opened up, a sub-graph - agenda-setting → qualification → discovery → pitch → price → objection handling → decision - in which each internal edge has its own conversion rate, its own needle-mover variables (the order of discovery questions moves qualification the way a subject line moves opens), and its own failure modes (pitching before the diagnosis is complete is this sub-graph's version of asking for the meeting in a cold email's first line). Sales, in other words, is not downstream of the theory; it is the theory at one level of nesting. The recursion continues past the close, where "onboarding" opens into contract-sent → contract-signed → assets-received → configuration-complete → activated, each transition instrumented and each one a place a customer can silently stall. And the nesting mirrors the knowledge substrate's own recursion (initiative → project → atomic work unit), so the execution layer and the memory layer share a shape - a symmetry that costs nothing and pays compound interest in tooling and comprehension.

A graph this honest also contains touches whose purpose is negative selection. The tradition's attract-and-repel doctrine becomes a designed node: the deliberate disqualification touch - "this may not be for you if…" - placed between booking and showing, whose green metric is a healthy rate of self-removal. It looks like sabotage to a stage-local optimizer, because it lowers show volume; it is correct at the north star, because it raises close rate and protects the scarcest resource in Section 4.3, the closer's calendar. Its presence in a graph is a standing test of whether the system has genuinely internalized north-star discipline or merely recited it.

Nodes attach to roles, never to people. Every node names a responsible role - a durable business function with its own scorecard - and the role resolves to its current occupant at runtime: a human, an agent, or a paired human-plus-agent, recorded on an occupant timeline. The consequence is that one graph serves every scale of the operation. At founder-solo scale, one person occupies most seats; at maturity, each seat holds a specialist or an agent; and the path between those states is a series of occupant reassignments with zero graph rewrites. The org chart stops being a wall poster and becomes a versioned property of the graph - which quietly converts staffing itself into something the loop can improve (Section 9.4). It also makes the automation frontier legible: which nodes are agent-occupied, which are paired, and which remain human is a queryable fact with a history, so the earned-autonomy gradient of Section 7 becomes an assignment policy rather than a sentiment.

Edges are KPI gates with a pre-committed diagnostic order. Section 2.3 established that edges carry expected conversion bands and pre-authored interventions; the journey graph is where that discipline pays in full, because the chain is now long enough that a red north star underdetermines its own cause. The complete discipline adds a declared diagnostic sequence: when the north star goes red, walk the chain upstream from the first red edge, testing hypotheses in a pre-committed order - transient anomaly first (Section 8's clumpy-distribution trap says most reds are noise), input quality second (did the population entering this stage change?), process drift third (did a constant silently move?), environmental shift last (has the market's sophistication stage advanced?). The order matters because cheap explanations are checked before expensive ones, panicked rewrites are structurally delayed, and a pre-registered ladder - like a pre-registered hypothesis - cannot be re-argued after the fact by whoever has a favorite culprit.

One node-class in the delivery phase deserves individual notice, because it is empirically the strongest single predictor in the post-sale graph: the adoption threshold - the tracked, binary moment the customer crosses from "I bought a service" to "I am running this method." It is a state transition, not a sentiment: define its observable criteria, timestamp its crossing, and treat its absence past a declared deadline - in practice, the first two to three weeks after activation - as the highest-severity churn signal the system emits. It is the post-sale twin of the pre-sale truth that a prospect who books but never confirms has already given you the answer.

9.2 The flywheel: the last phase feeds the first

The five phases are drawn left to right and operate as a circle. Retention & advocacy produces four output classes, and every one is an input to the acquisition phase of the next cohort: case studies, which become copy substrate and mechanism proof; testimonials and third-party reviews, which densify the proof wall cold traffic is judged against; referrals, which enter the top of the funnel warm and convert at multiples of the cold baseline; and reactivation and expansion demand, the cheapest revenue in the system. This is the flywheel, and it resolves a tension Section 4.2 left standing. At sophistication stages four and five - claims exhausted, mechanisms tired - the remaining moves are identity, story, and proof density. Proof is not written; it is manufactured, and the manufacturing line is the delivery phase. Which yields this section's sharpest reframe: in a sophisticated market, the delivery operation is not adjacent to marketing - it is marketing's upstream supplier, and an hour spent making a customer measurably successful is the highest-leverage copywriting available, because it produces the one asset a saturated market still believes.

The flywheel is also Section 4.2's beneficial reflexivity made mechanical. Adversarial reflexivity says your winning content erodes its own landscape; the flywheel says your delivered outcomes enrich your own input pool - each cohort's proof is denser than the last, each advocate warms the population the next cohort draws from, and the intent distribution of inbound prospects shifts favorably as a function of the system's own outputs. Two measurement consequences follow. Traffic temperature becomes a first-class fitness covariate: self-selected, proof-warmed inbound converts at multiples of cold outbound, so every fitness reading must be conditioned on the temperature of the cohort that produced it, or the system will attribute to a copy change what a warmth change caused. And persistent assets acquire measurable tails: where a delivered artifact remains installed in the customer's world - an embedded tool, a durable resource carrying attribution back to its source - it keeps routing self-selected demand for years after its originating cohort closed. That is precisely the tail Section 6's window-declaration discipline and Section 8's tail-blindness entry exist to keep visible: a genotype whose asset compounds off-window can lose every thirty-day comparison and still be the correct strategic champion, and the pre-registered tail modifier is what lets the evaluation say so.

9.3 The isomorphism: support and operations run the same loop

Here is the claim this section exists to make, at full strength: the iterative improvement of client acquisition before the sale and the iterative improvement of operational delivery after it are the same mechanism, applied to two segments of one graph. Not analogous - isomorphic: every load-bearing concept in the acquisition loop has a structure-preserving image in the operations loop.

The nine questions apply without translation. What varies: SOP steps, touch cadence, escalation logic, health-score weights. What is observed: the same wide-event discipline over operational events, unsampled wherever the outcome is business truth. What is fitness: retention-proximal, with time-to-first-value leading the way booked-and-shown leads revenue. Attribution: which onboarding touch drove adoption is the same multi-touch problem - and churn analysis is not like survival analysis, it is survival analysis, the identical hazard mathematics with renewal as the event. Memory: the runbook is the post-sale campaign template, living in the same substrate under the same mutation laws. And the layer ladder transfers rung for rung: Layer 0 is operational instrumentation - milestone and ticket events carrying genotype fingerprints, where which playbook version onboarded this customer is the content-hash question again; Layer 1 is health-score weight tuning with the same three lanes, a delivery lead's one-word verdict on an at-risk flag being the judgment lane verbatim; Layer 2 is controlled comparison of onboarding variants across customer cohorts; Layer 3 is embeddings for ticket routing, customer segmentation, and churn-risk similarity; Layer 4 is generated support responses and playbook drafts under accuracy and tone guardrails; Layer 5 is structural evolution of the delivery pipeline itself; Layer 6 is the operations doctrine amending from outcomes, in the same corpus, through the same amendment waves.

The isomorphism bends - without breaking - at four asymmetries, and naming them is what makes the mapping honest rather than hopeful. Samples are scarcer still: customers number a fraction of prospects, so the loop leans even harder on priors, hierarchical pooling, and judgment lanes; the statistical lane may never fire, and that is the design, not the failure. Blast radius per variation is higher: a weak email variant costs a reply; a weak onboarding variant risks a paying account and the referral tree behind it - so governance altitude per unit of variation rises, true holdouts are often contractually or ethically unavailable, and exploration shifts toward staged rollouts, natural experiments, and reversible-first sequencing. Latency is longer: renewal cycles dwarf reply windows, forcing the system onto leading indicators - the adoption threshold, time-to-first-value - whose predictive validity against the lagging truth is itself a monitored, periodically recalibrated model. But the fourth asymmetry runs the other way and pays for the first three: the counterparty is cooperative. A prospect owes you nothing; a customer is invested in your improvement and will tell you directly - in tickets, in reviews, in the kickoff call's offhand remark - what a thousand cold sends cannot reveal. Voice-of-customer is a high-bandwidth judgment lane that pre-sale marketing never gets at scale, and the deepest-funnel signals (renewal, churn, expansion, referral) are, per Section 6, the richest per event in the entire journey. Post-sale learning trades statistical volume for signal quality and cooperation. The loop does not change; its lane weighting does.

9.4 The seed: toward the self-improving organization

Once acquisition and operations are recognized as two segments of one improvable graph, the pattern refuses to stop at the customer. Every business function, examined at the same altitude, has the same anatomy: a population moving through a state-transition graph, driven by designed touches, measured on outcome-labeled edges, gated by thresholds with pre-committed interventions, staffed by roles with occupant timelines, and improvable by the nine questions over a shared substrate. Marketing and sales run the graph over prospects. Support and operations run it over customers. Hiring runs it over candidates and seats - and here the roles-not-people design pays its deepest dividend: because nodes bind to roles and roles carry occupant histories, staffing becomes an experiment the loop can run. The seat trial - multiple candidates, the founder included, working identical inputs against a single pre-declared metric for a fixed sample of attempts, the metric deciding the occupant - is Layer 2's experimental selection applied to people, with the same pre-registration, the same sample floor, and the same protection against narrative re-argument after the results land. Product development runs the graph over features and hypotheses; finance runs it over cash and commitments. One substrate, many journey graphs, one loop discipline, one governance model.

That composition - every function's graph reading from and writing to one sovereign, versioned substrate; every function's outcomes feeding every other function's priors; a single doctrine layer amending across all of them under one amendment protocol - is the self-improving AI-native organization, and it is the subject of a successor to this paper. This document is, deliberately, that theory's first worked instance: the marketing-and-sales subset, developed in full because client acquisition is where feedback is fastest, the empirical tradition is deepest, and the measurement discipline was codified first - a century ago, with a coupon and a key. The organizational theory will require no new machinery: the same nine questions, the same layer ladder, the same substrate and mutation laws, instantiated per function and composed at the doctrine layer. The claim to carry forward is the one this section has demonstrated from both directions: the close is an edge, not an endpoint - and the loop that improves how you win a customer is, mechanism for mechanism, the loop that improves how you keep one.

10. The asymptote

What does fully self-improving mean, at the limit - and what does it not mean?

It does not mean autonomy from human judgment. At every layer the human role transformed rather than vanished: sensor designer, judge of last resort, author of variants, curator of exemplars, editor-in-chief, architect, constitutional editor. The direction of travel is that humans stop performing the loop and start governing it - spending their attention where it is genuinely irreplaceable: taste, strategy, relationships, and the constitutional questions no fitness function can settle. A system that promises to remove the human has misunderstood which resource was scarce; it was never keystrokes, it was judgment, and the machine's job is to spend judgment more efficiently - to make one hour of human discernment propagate through a thousand decisions instead of one.

It does not mean convergence. Section 4.2's Red Queen forecloses that: the landscape erodes under the winners' feet, by the winners' own success. The asymptotic system is not one that has found the answer but one whose improvement rate durably exceeds its environment's decay rate - whose portfolio always contains the successor before the incumbent dies, whose stage-boundary detectors fire before the response curves flatten, and whose corpus makes every mechanism transition a coherence pass instead of a crisis.

What it does mean is the two loops running continuously, at their two speeds, over one substrate. The fast loop - variation, selection, inheritance against the market - improving the artifacts. The slow loop - reflection, amendment, canonization against the evidence - improving the generator of artifacts, including the fast loop's own policies. The fast loop without the slow one yields a tactically excellent operation that is fragile to every mechanism shift and leaks its lessons through personnel and vendor churn: the fate of most agencies. The slow loop without the fast one yields a beautifully documented ghost - doctrine untested by contact, a museum: the fate of most "thought leadership." Together, compounding through each other, they produce the thing this paper has been circling: an organization whose accumulated, versioned, structured, agent-readable knowledge is its competitive asset - more than any individual campaign, tool, or model, all of which are regenerable projections of it.

The grand unification, then, in one paragraph. A marketing campaign is a hypothesis about human response, generated from a knowledge substrate, expressed through a deterministic structure, filled with perishable content, and projected into a market whose response function its own success erodes. A self-improving campaign system is the disciplined return path: market responses, captured faithfully and attributed honestly, flow back to amend - in strict order of increasing evidence, consequence, and rarity - the weights, the traffic, the frontier, the content, the structure, and finally the doctrine that generated all of them. The deterministic laws of direct response govern every step because they are the conditions under which feedback yields information rather than confident error. The AI at every layer is an instrument - interpreter, prior, generator, judge, simulator, editor - that compresses the loop's cycle time and extends its reach, and earns each promotion in altitude with evidence. And the substrate is the point: learning recorded in projections evaporates with the vendor; learning recorded in sovereign, versioned, human-and-agent-readable substrate compounds without bound. And the graph does not terminate at the close: the journey's final phases manufacture the proof, the warmth, and the referred demand its first phases consume, so a complete system improves the winning of customers and the keeping of them as one object, and each loop funds the other. The models will keep improving; that was never the bottleneck. The organizations that pull away will be the ones whose loops close - every cycle, at every layer, onto ground they own.

Appendix A: schemas and pseudo-code

Illustrative sketches, not implementations. Field names are suggestive; the shapes are the point.

A.1 - The outcome event (Layer 0). One event per edge traversal, unsampled, dimension-rich at emit time, genotype-fingerprinted, bi-temporal.

event: prospect.replied.v1
occurred_at: 2026-07-08T14:22:05Z      # when it happened in the world
recorded_at: 2026-07-08T14:22:31Z      # when the system learned of it
subject:
  prospect_id: p_84c2                   # post identity-resolution
  cohort_id: cohort_2026w27_a
genotype:                               # the exact stimulus, reconstructible
  structure_version: 4
  content_hash: "9f31be07…"             # content-addressed publish bundle
  targeting_version: 2
  weights_version: 7
decision_context:                       # for honest off-policy evaluation
  score_at_selection: 81
  variant_propensity: 0.42
classification:
  category: DESIRED                     # closed routing alphabet
  outcome: meeting_requested            # open taxonomy, recorded as data
  confidence: 0.93
dimensions: { channel: email, segment: fintech_ops, touch_index: 3 }

A.2 - Clamped multiplicative weight update (Layer 1, statistical lane).

for signal in scoring_signals:
    n_s   = count(resolved_outcomes where prospect_has(signal))
    if n_s < MIN_EVIDENCE: continue                    # priors hold below floor
    lift  = decayed_rate(desired | signal) / decayed_rate(desired | all)
    step  = LEARN_RATE * sign(lift - 1.0)              # e.g., 0.1, direction only
    w[signal] = clamp(w[signal] + step, W_MIN, W_MAX)  # e.g., [0.3, 3.0]
log_amendment(signal_weights, evidence=cohort_ids, trigger="batch_1000")

A.3 - Thompson sampling over content variants (Layer 2).

for arm in variants:  θ[arm] ~ Beta(α[arm], β[arm])    # sample belief
send(argmax_arm θ)                                      # exploit the sample
on outcome: (α,β)[arm] += (success, failure)            # update belief
# non-stationarity: decay (α,β) toward priors on a sliding window

A.4 - The campaign genotype (Layer 5's unit of evolution).

Genotype = (G, C, T, W)
  G: DAG(nodes: actions, edges: outcome-category-labeled transitions)   # structure
  C: map(node → content_hash)          # immutable content refs
  T: targeting predicate (typed AST)   # who enters
  W: scoring weight vector             # who ranks where
Phenotype  = execution of Genotype over a cohort under declared constants
Fitness(g) = north_star(outcomes | cohort(g)), read only after
             sample_floor(channel) and latency_window(channel) are met

A.5 - The amendment record (Layer 6).

amendment:
  date: 2026-07-10
  approver: <named human>
  trigger:
    claim_falsified: "1 close per 1,000 sends (canonical planning figure)"
    evidence: [cohort_2026w14..w18, event_ids…]
    corrected_claim: "1 close per ~5,000 sends at current stage/segment"
  coherence_wave:                      # every artifact whose meaning changed
    - artifact: revenue_model.md        (rewrite §2 projections)
    - artifact: adr/tier2_unlock.md     (append amendment log; supersede metric)
    - artifact: playbook.md §3.1        (update threshold; cite this record)
  captured_principles:                  # the aside that outlived the session
    - "Volume thresholds set against optimistic funnel math are unreachable
       by construction; thresholds cite their funnel-math version."
  out_of_scope: [ …deferred, with rationale ]

Appendix B: a diagnostic

Ten questions to locate any system - yours, a vendor's, a case study's - on the layer ladder, honestly.

1. Can you reconstruct, byte-for-byte, the exact stimulus behind any outcome from last quarter? (No → you are below Layer 0, whatever the dashboard says.)
2. Is there one outcome taxonomy shared by the classifier, the nodes, and the routing - or three implicit ones drifting apart?
3. Are business counts sourced from an unsampled stream? Would anyone notice if they weren't?
4. When a score changes, can you name the lane - hand, judgment, or statistics - and the evidence that moved it?
5. What was the last test discarded for a broken gate (sample floor, latency, constants)? (If the answer is "never," the gates aren't real.)
6. What fraction of sourcing and traffic is budgeted exploration - and who defends that budget when the quarter tightens?
7. Has a generated variant ever been blocked by the judge, and has the judge ever been scored against the market?
8. When did a structure last get sunset on criteria rather than nursed on hope?
9. Does the loop survive the close - do onboarding, delivery, and support run on outcome-labeled edges, versioned playbooks, and health-score lanes with the same discipline as acquisition, or does measurement end at the signature?
10. Where would a lesson learned this week live, such that a new campaign - or a new hire, or an agent - inherits it next quarter without anyone remembering to tell them?

The last question is the whole paper in miniature. If the answer is a tool's settings page or a veteran's memory, the system may be optimizing, but it is not self-improving. If the answer is a versioned, readable, owned substrate - it compounds.

This framework synthesizes a century of direct response doctrine (Hopkins, Caples, Ogilvy, Halbert, Schwartz, Kennedy, Brunson, Hormozi, Voss, and the systems formulations of client acquisition), the engineering canon of feedback and architecture (ports and adapters, domain-driven boundaries, event sourcing, antifragility, the theory of constraints, lean iteration), the statistics of sequential experimentation and causal inference, and Hofstadter's account of tangled hierarchies - applied to the question of how a marketing organization's contact with its market can be made to improve the organization itself. It is free to adopt, adapt, and extend. The substrate discipline it assumes is specified independently as the Organizational Context Protocol (OCP), an open standard; everything else here composes with any conformant substrate. Structure endures; content decays; doctrine amends; the loop is the point.