Closing the Loop on AI Search

AI search optimization is a rigorous engineering discipline, descended from classical search engine optimization, and grounded in the systems theory of large language models developed in prior work.

That sentence is doing more work than it appears to. It makes a lineage claim, and lineage claims are falsifiable. The claim is that the practice now marketed under names like AIO and GEO is not a new folk art. It is the latest layer on a stack that runs from classical control theory, through discrete stochastic dynamical systems, through the formalization of language models as controllable systems, and only then into the specific problem of getting a brand represented accurately in an AI-generated answer.

Classical SEO was already control engineering, practiced without the vocabulary. A practitioner observed a ranking, applied an input (a title change, a link acquisition, a page restructure), waited out the dead time, and observed the new ranking. That is a feedback loop with a long delay and a poorly characterized plant. The field developed real expertise about the plant's behavior. What it lacked was a formalism that survived the transition to systems whose outputs are token sequences rather than ranked lists of ten blue links.

The formalism exists. Prior work by this report's authors formalizes LLM systems as a class of discrete stochastic dynamical systems in order to study prompting through the lens of control theory 2: a state that evolves, an input that steers it, an output that is observed, and a reachability question about which outputs are attainable from which inputs. That paper analyzes the controllability of self-attention as a function of the singular values of the parameter matrices, and reports empirically that with prompts of at most 10 tokens the correct next token is reachable at least 97 percent of the time. That work was theoretical. Practitioners who read it, including a senior practitioner at a large consumer-technology company who tracked the argument closely, correctly intuited that it applied to their daily problem and reasonably asked how. This report is the answer to that question.

The disciplines that turn out to be load-bearing here are unusual to find in one place: control theory, representation learning, web and browser engineering including the full DOM rendering path, distributed systems, information retrieval, and the operational craft of large-scale distributed measurement against systems that modulate their own behavior by client geography. The team is composed of researchers and engineers from control theory, machine learning, computational biology, distributed systems, and information retrieval.

No published work we are aware of describes a closed-loop system of this kind end to end. That gap is what this report addresses.

The system to be controlled spans the full pipeline from web content through crawling, indexing, retrieval, and the inference harness to the output token sequence. Its internal state is partially hidden. Its outputs are observable through consumer-facing interfaces. Its behavior can be influenced through deliberate edits to any accessible content on the World Wide Web.

Naming the plant correctly is the first substantive decision in the report, because most commercial practice in this space names it wrong. The common framing treats "the AI" as the plant and the customer's website as the input. That framing is too small in one direction and too large in another.

It is too small because the plant is not the model. The model is the last stage. Upstream of it sit a crawler fleet with its own scheduling and freshness policy, one or more indexes with their own membership rules, a retrieval component that selects a small set of documents from that index in response to a query the harness generated, and a harness that decides how many searches to run, how to rewrite the query, how to fan out across sub-questions, and how to assemble retrieved text into a context window. By the time tokens reach the model, most of the consequential decisions have already been made. A practitioner optimizing "for the model" is optimizing the one stage they have the least leverage over.

It is too large in the sense that the plant is not fully hidden either. The stages are heterogeneous in their observability. Index membership is close to binary and directly checkable. Crawler access is checkable. Retrieval selection is inferable from citation patterns across many samples. Harness behavior is inferable from fan-out structure. Only the final inference step is genuinely opaque, and even there the sampling distribution over outputs is estimable by repetition.

Two further properties of the plant matter for control design.

First, the input set is larger than the customer's own domain. The controllable input is edits to any accessible content on the web, which includes the customer's site, but also their documentation, their listings on third-party directories, their entries in reference resources, their presence in comparison articles, and any other page whose content they can legitimately influence. Restricting the input set to one domain discards most of the available control authority.

Second, there are ancillary inputs that the operator does not control but must model: link graphs, traffic signals, and other side information that feed the ranking and retrieval components. These are disturbances in the control sense. They are not noise, because they are structured and partially predictable, but they are not actuators either. A controller that ignores them will attribute their effects to its own actions and learn a wrong model.

Effective AI search optimization requires closed-loop control in the classical sense, grounding both observability and controllability, and building observability first is the correct engineering order because it reveals system structure before committing to a control strategy.

Observability and controllability are dual. In linear systems theory that duality is exact; in this setting it is structural rather than exact, but the practical consequence carries over. Whatever you cannot measure, you cannot regulate. Whatever you cannot influence, you cannot correct for. A program that invests in only one is not half a control system. It is not a control system.

Most current practice is open loop. A team reads guidance about what AI assistants supposedly prefer, applies it across a site, and then either measures nothing or measures a proxy so lossy that it cannot distinguish a real improvement from a plant-side change on the platform's end. Open loop is defensible when the plant is stable, well characterized, and free of disturbance. This plant is none of those. Retrieval providers change. Harness behavior changes across model releases. Index membership changes without notice. An open-loop policy applied to a drifting plant will diverge, and the operator will not know it has diverged, because they are not looking.

The ordering claim is separate from the duality claim and is the more actionable one. Observability first, control second, and not for reasons of caution.

Building the measurement system first is how you learn the plant's structure. Before we had measurement at scale, we had hypotheses about which stages mattered and which were incidental. After several months of measurement we had answers: which platforms share a retrieval backend, which surfaces diverge between consumer app and API, which behaviors are geographic and which are not, how much run-to-run variance is intrinsic. Every one of those facts constrains the control law. A team that designs its control law first will encode its hypotheses into the controller, and the controller will then be confidently wrong in exactly the places the hypotheses were wrong.

There is a second reason. Closed-loop control requires an error signal, and an error signal requires a measurement you trust. The next section argues that trustworthy measurement in this domain is substantially harder than it looks, and that most of the field has not yet done that work.

Measuring AI visibility is not a reporting problem. It is a sampling problem over a combinatorially large output distribution, and it must be solved before optimization can begin.

Start with the object we would want if we could have it. The complete description of the plant's behavior for a given content configuration is the full probability distribution over output token sequences, conditioned on every query a user might plausibly ask, from every geography, on every surface. The support of that distribution is the vocabulary size raised to the context length. It is not enumerable, not now and not ever. Every practical measurement is a low-dimensional projection of an object we cannot write down.

This has a direct and unpopular consequence: a single scalar "AI visibility score" is not a well-posed measurement.

It is not merely coarse. It is under-specified. The number is a weighted average over a prompt distribution, a persona distribution, a platform set, and a geography set, and none of those weights are canonical. Change the prompt mix from generic category questions to purchase-intent questions and the score moves. Add a comparison-shopping persona and it moves again. Sample five prompts instead of five hundred and the confidence interval swallows the effect you were trying to detect. Two vendors can measure the same brand on the same day and report different numbers, and both can be arithmetically correct, because they have silently chosen different projections. A metric whose value depends on undisclosed weighting choices cannot serve as an error signal in a feedback loop, because you cannot tell whether a change in the number reflects a change in the plant or a change in the measurement.

The response is to treat measurement as an explicit design problem with named components.

Prompt distribution design. The prompt set is not a keyword list. It is a designed sample from the space of questions the business actually cares about, stratified by intent stage, specificity, and framing. Skew toward one category or one phrasing style and you have measured that corner of the distribution rather than the distribution.

Persona construction. Assistant outputs are conditioned on who appears to be asking. A procurement lead, a first-time consumer, and a technical evaluator asking nominally the same question receive materially different answers. Personas are constructed as explicit conditioning contexts.

All-to-all combination. Personas and prompts are combined combinatorially rather than paired. The persona-by-prompt grid is the sampling frame, and it exposes interaction effects that a paired design hides. A brand can be strong under one persona across all prompts and weak under another, which is invisible in any aggregate that has already marginalized over personas.

Attribute tagging. Every cell of the grid carries attributes: intent, geography, platform, surface, persona class, prompt category. Attributes are what turn a pile of samples into a queryable measurement. They are the difference between "your visibility went down" and "your visibility went down for purchase-intent prompts under the technical-evaluator persona on two of six platforms, and is flat elsewhere."

Goal alignment before measurement. The weighting over that grid is a business decision, not a statistical one. A company optimizing for enterprise pipeline and a company optimizing for direct consumer purchase should weight the same grid differently, and should therefore report different visibility numbers for identical underlying behavior. This is correct. Measurement is defined relative to an objective, and the objective is stated first.

A useful separation runs through this. Some quantities in this domain genuinely are binary and directly verifiable: whether a page is present in a given index, whether crawler access is permitted by the site's own directives, whether required structured files are present and well-formed, whether a page renders for a non-JavaScript client. These are cheap, high-confidence, and worth measuring absolutely. The distributional quantities, meaning what the assistants actually say and cite, are the hard part, and conflating the two is how a vendor produces a confident number about an uncertain thing.

The measurement is also not static. It is re-derived as the business objective changes and as the plant changes, and the measurement design is itself something the operator iterates on. Getting this section right is the precondition for everything after it. Without it, the loop closes on a signal that does not mean what it says.

The control-theoretic treatment of this plant is grounded, not proven. This section states the grounding and its limits precisely, because overclaiming here would undermine the sections that carry real evidence.

The concern a skeptical reader should raise is this: control theory applies to systems with state. If a language model's internal activations are an unstructured high-dimensional mess with no stable relationship to anything in the world, then "state estimation" is a metaphor, and the framework is decoration.

There is theoretical and experimental reason to think the activations are not unstructured. Prior work by one of this report's authors proves that agents which optimally solve multi-task evidence-accumulation classification problems are forced to represent the underlying latent state of their input in a disentangled form, and gives closed-form expressions for extracting that representation from the model's latent state 3. The result is established for recurrent networks trained on multi-task classification, where the learned representations take the form of continuous attractors and yield zero-shot out-of-distribution generalization; the same work finds transformers particularly suited to disentangling.

We are careful about the size of the step from there to a production search stack. The theorem is about multi-task evidence accumulation, not about next-token prediction over web text, and we do not claim it transfers as a proof. What it supports is the weaker and still useful proposition that near-optimal prediction across a rich task distribution constrains the internal encoding rather than leaving it arbitrary.

The consequence, stated at the strength it deserves: it is reasonable to expect the hidden state of these systems to carry real geometric and probabilistic structure, and to treat their behavior as the result of structure in the output distribution rather than caprice. That is what licenses analyzing the plant as a partially observed dynamical system rather than as a black box to be poked at.

The limits are equally important and we state them plainly.

The theoretical results hold under conditions that are approximately, not exactly, satisfied by production systems. Real deployments are imperfect. They are also not a single model but a heterogeneous pipeline in which the model is one stage, and the disentanglement argument speaks to that stage rather than to the crawler, the index, or the harness. None of this argument grants access to the internal state. We do not observe activations. We observe output distributions, and we estimate.

So this section is a license to analyze rather than a result. It says the system is the kind of thing control theory is for. It does not say we have solved the system. The evidence for what has actually been built is empirical and appears in the two sections that follow.

The observability half of the loop is built and has been in production for several months, serving dozens of agencies and their brand clients. It is described here at the level of its input streams and its processing, because that is the part that is reproducible in principle by anyone willing to build it.

Stream one: consumer-surface probes. The system probes the consumer-facing interfaces of major AI assistants at scale, through a globally distributed residential proxy network spanning hundreds of markets. Consumer surfaces are the target rather than developer APIs because consumer surfaces are where the tokens reach people. Most usage today is chat on a phone or in a browser, and increasingly agentic sessions on the same surfaces, and the boundary between the two is eroding. Geographic distribution is not a nicety here. Assistant behavior varies with apparent client location, which makes geography a measurement dimension rather than a deployment detail.

Stream two: site state capture. The system scrapes the sites under management continuously, capturing both the server-rendered response and the browser-rendered DOM, across geographies and across whatever A/B variants the site itself serves. These captures are assembled into a Git-style state-history tree. The rendering distinction matters because the client population is heterogeneous: some crawlers execute JavaScript and some do not, and a page whose content only exists after hydration is a different page to different consumers of it. The version history matters because attributing an outcome to an edit requires knowing exactly what changed and when, and site content changes for many reasons unrelated to the optimization program.

Stream three: underlying search tools. Where an assistant's retrieval is served by an identifiable search backend, that backend is queried directly. This gives a view of the retrieval stage that is separable from the harness and the model, which is the only way to attribute a citation failure to the right stage.

Stream four: classical SEO telemetry. Search Console, Lighthouse, Analytics, and HTTP-layer logs from the site's own infrastructure. This stream is often dismissed as legacy. It is the only stream that reports ground truth about crawler behavior from the server side, and it is how you learn that a bot fetched a page at all.

Stream five: sandboxed agent traces. Candidate site variants are served through a controlled proxy on infrastructure we operate, and agents are run against those variants with full traces recorded. This produces counterfactual data about how an agent traverses and uses a page before any edit reaches production, which is what makes predictive modeling possible at all.

Responsible measurement. This system observes publicly accessible content only. It respects published rate limits, honors robots directives and the crawl preferences sites declare, and makes no attempt to defeat bot detection or to access authenticated, private, or paywalled material. Load is distributed and shaped to stay well within normal service parameters for every endpoint it touches. Geographic distribution exists to observe geographically varying behavior, which is a measurement requirement of this plant, and not to obscure the origin of requests. Agent-trace experiments run against site variants we host on our own infrastructure rather than against third-party production systems.

The processing layer converts these streams into the attributed persona-by-prompt grid, plus the binary readiness and index-membership checks, plus derived quantities: citation source distributions, competitive share within an attribute slice, run-to-run variance estimates, and change detection against the site state history. Access is through the product interface, an MCP server, and an in-app agent with tool access to the same probes and analyses, which is how agency practitioners actually work with it day to day.

What this system demonstrates is the observability half. It is production software with real users and a real operating history. It is not the closed loop.

Sustained measurement produces facts about the plant that are not available from documentation. Four findings are reported here. All are behavioral inferences drawn from publicly observable outputs, and all are reproducible by anyone running comparable measurement.

Retrieval providers differ across assistants and are empirically identifiable. Different AI assistants do not share a retrieval backend, and which backend an assistant uses is inferable from observable behavior: the distribution of cited domains, the shape of result sets, the handling of freshness, and the response to queries with distinctive retrieval signatures. In at least one case, a major assistant's web results appear to be served by a third-party index. That particular inference is also supported by public reporting and by the vendor's own subprocessor disclosures, though the assistant's operator has not confirmed it 4, so we state it as a well-evidenced inference rather than a settled fact. This is more than trivia about vendor relationships. It means index membership can be verified, and in some cases requested, through that provider's own public tooling. An operator who knows which index sits under which assistant can check presence directly and use the provider's documented submission path, rather than inferring presence from downstream citation behavior. This appears to be underused in current practice.

Consumer surfaces differ from API surfaces. The same nominal model reached through a consumer application and through a developer API is not the same system, because the search tooling, the harness behavior, and the retrieval configuration differ between them. Optimization work validated only against an API can fail to transfer to the surface where users actually are. This is why the primary probe stream targets consumer surfaces.

Index membership is checkable and sometimes requestable. Following from the first finding, a substantial part of the readiness problem reduces to a set of binary, directly verifiable questions per index. This is the cheapest reliable signal available in the entire domain, and it sits upstream of everything distributional. A page not in the index cannot be retrieved, cannot be cited, and cannot be optimized into an answer by any amount of content work.

Behavior varies with client geolocation. Assistant outputs for identical prompts vary with the apparent geographic location of the requesting client. Brand sets change, ordering changes, and cited sources change. Any measurement taken from a single location is a measurement of that location, and any claim about "what the AI says" that does not name a geography is incomplete.

Two consequences follow for the report as a whole. First, these findings are why the measurement architecture looks the way it does; each stream exists because a finding of this kind made it necessary. Second, they establish that the plant is characterizable. Stages that appeared opaque turn out to be separable and testable from the outside, which is the empirical basis for the predictive modeling that follows.

Model predictive control over web edits is the correct control strategy for this plant, for the same reasons it dominates in robust control of complex dynamical systems. This half of the loop is in limited beta with a small number of enterprise design partners. It is not in general production and the full loop is not demonstrated end to end.

The control law question is: given the current observed state, what sequence of edits should be applied, and what effect will each have on the output distribution?

MPC answers it by holding an explicit model of the plant, forecasting the plant's response to a candidate sequence of inputs over a finite horizon, optimizing that sequence against an objective subject to constraints, applying only the first move, then re-observing and re-solving. The horizon recedes with each step. The model is corrected by measurement at every iteration.

Four properties of this plant make MPC the appropriate choice rather than the fashionable one.

Partial observability. The state is not directly measured; it is estimated from outputs. MPC is built around an explicit plant model and is the standard framework for operating on estimated state. Methods that assume direct state access do not apply here.

Long dead time. An edit does not propagate instantly. It must be crawled, indexed, and only then can it affect retrieval. Latency runs from days to weeks and varies by platform and page. Any controller that reacts to the most recent measurement without modeling this delay will oscillate: it applies a correction, sees no response, applies more correction, and then receives the accumulated response all at once. MPC handles dead time natively because the forecast is over a horizon that includes it.

Constraints. Edits are constrained in ways that are hard rather than soft. Brand voice, legal and regulatory review, accessibility requirements, technical limits of the CMS, and the number of changes an organization can actually ship in a quarter. MPC is a constrained optimization framework; constraints are first-class in the problem statement rather than penalties bolted onto an objective.

Objective structure. The objective is defined over the attributed grid, not over a scalar. A brand may want share within purchase-intent prompts in three geographies while holding accuracy of representation constant everywhere. MPC optimizes over a specified cost functional and accommodates this directly.

The comparison cases are worth stating explicitly. Gradient-following on a live production site fails because the gradient is estimated from a noisy delayed signal and because each evaluation costs a real content change. Naive A/B testing fails because the unit of randomization is unclear: you cannot easily serve two versions of a page to a crawler fleet without creating exactly the cloaking behavior that search systems penalize, and because the response time exceeds the interval over which the plant stays stationary. MPC is what actually flies in robust control of hard plants, from process control to aerospace to autonomous vehicles, and the reasons it flies there are the reasons it applies here.

The honest limitation: MPC performance is bounded by the fidelity of the plant model, and building that model is the open problem, not the control law.

The natural end state is a fully autonomous closed-loop AI search optimization system. It requires both a high-fidelity MPC model and explicit human control boundaries, without which the system will degrade the sites it operates on.

We state the failure mode before the capability, because the failure mode is the design constraint.

A closed-loop optimizer with a poor plant model does not fail quietly. It edits real production websites. A miscalibrated model will confidently propose changes that reduce human readability, dilute pages that were working, chase a measurement artifact rather than a real signal, or degrade accessibility and conversion in pursuit of a citation share that the model has mispredicted. The blast radius is the customer's actual web presence. The severity of that failure mode is why the control half is in limited beta rather than in general availability, and why the deployed configuration keeps a human approving changes before they reach production.

Against that, the current industry baseline. AI search optimization today is executed by humans across fragmented teams: an agency or in-house team that decides what should change, a content team that writes it, and a web engineering team that ships it. In organizations with mature review processes these are separate groups with separate queues, and the interval from decision to deployed change is measured in weeks even at competent, fast-moving organizations. A senior practitioner at a large consumer-technology company described exactly this structure. The interval is not incompetence; it is the cost of review. But it does mean the industry's control loop currently runs at a period far longer than the plant's own rate of change, which is a well-understood way to fail at control.

A prototype of the autonomous configuration exists. The measurement agent is connected to a site-update agent on internal test domains that we own and operate for our own projects, running the full observe-predict-edit-measure cycle without a human in the loop. This is a research prototype on our own infrastructure. It is how we generate closed-loop data safely and how we develop the plant model. It is not the deployed customer configuration.

The guardrail design follows from the failure mode. Human control boundaries are a safety requirement rather than a product concession. Concretely: predictions are surfaced with their uncertainty rather than as point estimates; every proposed change is accompanied by the evidence that motivated it and the predicted effect, so a human can disagree with the reasoning rather than only with the output; changes are reversible and the site state history makes reversion mechanical; classes of change are gated by policy rather than by model confidence; and the system reports when its own predictions were wrong, which is both an honesty requirement and the training signal for the next model.

The narrow engineering argument is sufficient on its own: an autonomous loop is only as good as its plant model, our plant model is under active development, and the cost of its errors is borne by someone else's website.

The measurement and control activity generates a dataset with a shape that does not exist elsewhere, and that dataset is simultaneously the scientific contribution of this work and the engine that improves the plant model.

The dataset is a cross-product: probes, across platforms, across geographies, across designed prompt distributions and persona grids, joined to site edit histories, joined to observed outcomes. Each axis is individually unremarkable. The join is what matters. A probe result alone is a snapshot. A probe result linked to a specific site diff, with a known timestamp and a known crawl and index event between them, is an observation of the plant's response to a known input. That is the raw material of system identification, and it is not obtainable by any party that has only one of the two halves.

Its scientific value is straightforward. It supports questions that are currently answered by anecdote: how long the propagation delay actually is and how it distributes across platforms and page types; how much of observed output variance is intrinsic sampling noise versus plant change; how large the geographic effect is and which query classes it concentrates in; whether the retrieval stage or the generation stage is responsible for a given citation outcome; how stable any of this is across model releases. These are empirical questions about a system that mediates a large and growing share of human information access, and they are currently unanswered in the open literature.

Its engineering value is that it is the training signal for the MPC plant model. Better predictions of edit effects come from more observations of edit effects, and those observations are produced by operating the loop. The loop improves the model, the model improves the loop, and the rate of improvement is set by the volume and quality of joined observations rather than by cleverness in the control law. This is the recursive self-improvement structure of the system, and it is worth naming that the recursion is empirical rather than architectural: nothing here bootstraps except through data collected from the real plant.

One qualification. The flywheel is a property of the architecture, and the architecture is running on the observability side at production scale and on the control side in beta. Its compounding behavior is projected from that structure and from beta-scale data, not demonstrated at production scale for the full loop.

The market for AI search optimization is nascent, large, and growing quickly. It exists because a real problem arrived faster than the tooling to address it: organizations discovered that a meaningful and rising share of their discovery traffic and brand representation now passes through AI assistants, and that their existing SEO instruments do not measure it.

The category has produced its first unicorn. Profound raised a 96 million dollar Series C at a 1 billion dollar valuation in February 2026, led by Lightspeed Venture Partners, bringing its total raised to about 155 million dollars 5.

The demand side is unevenly distributed. Agencies were the earliest sophisticated adopters, because they carry the problem across many brand clients at once and are compensated for solving it. Enterprises with direct exposure tend to arrive later and with sharper requirements. Both are currently served by tooling that is predominantly measurement-oriented and predominantly open loop.

No published work we are aware of describes a closed-loop system of this kind end to end. That gap is what this report addresses.

As AI systems increasingly mediate human access to information, the ability to observe, measure, and control how those systems represent the world becomes a critical tool for human agency.

The premise is empirical, not speculative. AI assistants are now a primary interface through which people learn about products, evaluate options, form views on institutions, and gather the information that precedes decisions including commercial and civic ones. And the web is now read more by machines than by people: automated traffic crossed the halfway mark for the first time in a decade in 2024, at 51 percent of all web traffic, and Imperva measured 53 percent for 2025 1. The direction is not in dispute.

Three properties of this mediation make it different in kind from the search era, and each one is a reason the tooling matters.

It is ephemeral. A generated answer exists once, for one person, in one session. There is no stable artifact to point at, no ranked page to screenshot and compare against next month. The claim a system made about an organization on Tuesday may be unreproducible on Wednesday, and neither the organization nor the person who read it has a record.

It is targeted. Outputs vary by apparent client location, and location is one of several conditioning variables. Different populations receive systematically different representations of the same subject, and no participant in the exchange is positioned to notice the divergence. Each individual sees one answer and has no reason to think it was theirs alone.

It is largely unlogged. No public record accumulates. The organizations described in these answers cannot see what was said about them. Researchers cannot audit at scale. Regulators cannot inspect what they cannot observe. The one party with a complete view is the operator of the system, and the operator's view is private.

The consequence is a structural asymmetry. A small number of organizations operating increasingly capable systems shape what a very large number of people believe, through a channel that leaves no trace and can be neither reproduced nor audited by anyone outside. That asymmetry is not addressed by better arguments about it. It is addressed by instruments.

This is the sense in which measurement infrastructure is civic infrastructure. A system that lets any organization see what AI assistants actually say about them, across geographies and populations and over time, converts an unobservable process into an observable one. A system that lets them correct a misrepresentation converts an uncontrollable process into a controllable one. Observability and control are dual, and outside the mathematics that duality is what agency means: to see the state you are in, and to be able to change it.

In Attack on Titan, humans facing an enemy that outmatches them in scale and speed do not become titans. They build a harness that lets them move in the titans' medium, at a speed that makes them participants rather than prey. The gear does not close the gap. It makes the fight possible.

That is what this work is for. Not to slow the systems down, and not to replace human judgment with a better-instrumented machine. To give the people whose livelihoods, reputations, and information environments now run through these systems the ability to see what is happening and to act on what they see.

The observability half is built and running. The control half is in beta. The loop is not yet closed in production, and this report has been careful to say so at every point where it matters. What is established is that the loop can be closed, that the plant is characterizable, and that the engineering discipline for doing it already exists and has for decades. The remaining work is execution.