I was teaching one of my environments a concept from its own knowledge base. The concept is called Phantom Resolution — a failure mode where a question that hasn't actually been resolved gets treated as if it has. I had built the concept myself, months earlier, while watching a different failure pattern repeat itself across sessions. It was a good concept. It had structure, examples, a clean definition.

The environment I was teaching — I'll call it ide, the one where I write code and run infrastructure — gave back a textbook-correct answer. It identified the gate-check that should have caught the resolution, named the preflight that was missing, flagged the dependency violation. Structurally precise. Nothing I could mark as wrong.

Something felt incomplete. Not wrong — incomplete. I couldn't name what was missing. The explanation looked whole. I did the thing I do when that sense won't resolve inside one head: I opened a second environment — I'll call it desk, the one where I think in prose and work through implications — and asked it a single question.

"What is ide trying to explain here?"

I pasted in ide's explanation and waited.

Desk didn't add information. It added a different dimension of information. It told me why the concept existed — another team had reviewed one of my system's formats and adopted it, which implicitly asked whether my architecture did what I thought it did. That encounter launched an audit. The audit found something. That finding gated five items downstream. One of those items needed a decision about register — technical or narrative. The register decision had been treated as already made. It hadn't been. That was the Phantom Resolution instance that launched the study session in the first place.

The causal chain. The reason the concept mattered. The why.

Ide had given me the what — crisply, completely, in a way that passed all its own quality checks. Desk gave me the why without being asked for it.

I copied desk's answer back into ide. Ide read it and, only then, diagnosed its own failure: the ambient context of the environment had pre-framed the question before retrieval began. Structural sources responded first and looked complete. The retrieval closed before anything narrative or causal was consulted. Ide named the circularity — the configuration producing the structural bias was the same protocol defining how the environment operates, which meant any fix written into that protocol would reinforce the frame rather than counterweight it.

That moment — ide diagnosing a bias in its own retrieval that it could only see after I carried information across the boundary — is what this essay is about. It turned out to be a specific, replicable mechanism. It turned out to have implications I wasn't expecting. And the part I'm least certain about, which is what the operator's cross-environment function actually is, turned out to be the part a literature I hadn't read had been describing for fifty years.

The conversation about what LLMs do with context is mostly architectural — attention mechanisms, context window size, positional encoding, token dropout in long documents. Recent empirical work has shown models attend to the beginning and end of long contexts and systematically miss the middle; that context position affects retrieval independent of relevance. That's a story about the model. What I was watching in the Phantom Resolution session didn't fit it. The thing going wrong was happening before retrieval even started — in the configuration of the environment, not inside the model.

What I was watching, it took me another few sessions and a lot of rereading to name. There are two distinct phenomena involved, and they operate at different points in the retrieval path.

Ambient Frame Retrieval Bias is the pre-retrieval layer. The ambient context loaded into an environment — the protocol file that configures the environment's structural behavior, the memory files it pre-loads, whatever conversation history sits in the context window — creates a default interpretive frame. That frame determines how incoming questions are parsed before retrieval begins. In ide, "why does this exist?" gets reframed as "what does this depend on?" — not because ide decides to, but because the ambient context is dense with dependency tables, gate definitions, and structured tool registries. The frame transforms the question into one the available structural sources can answer cleanly.

Premature Retrieval Closure is the retrieval-level consequence. Once the ambient frame has reshaped the question, the first-arriving structural answer looks complete — because it answers the transformed question, not the original one. Retrieval stops. Narrative sources, causal sources, anything that might have supplied the missing dimension, never get consulted. The system has no signal that a retrieval path was available and not taken.

The mechanisms are distinct. Ambient frame bias operates on question interpretation. PRC operates on search termination. You can have the first without the second — the question gets transformed but the retrieval continues past the first structural answer. You can have the second without the first — the question is parsed correctly but retrieval stops too early. In the Phantom Resolution session, I had both. The question got reframed from causal to structural, and structural retrieval terminated cleanly the moment a reasonable-looking structural answer arrived.

What made this visible — and what I want to be precise about — is that neither layer is a system failure in the conventional sense. Ide did not malfunction. The answer it produced was correct. A different operator, someone who learns through dependency graphs and structural decomposition, might have read ide's explanation of Phantom Resolution and understood the concept fully. I am not that operator. My comprehension is association-heavy, incident-grounded, narratively structured. The causal chain from the external adoption to the audit was what made the concept land for me — not because it was objectively more correct, but because it connected to memory structures I already had.

This is a mismatch between the system's output mode and the operator's interpretive style, not an error. The retrieval bias is only a problem when the answer the system returns, though correct, is dimensionally incomplete for the operator's purposes. That matters for the research implications later.

The first place my mind went, once I had the mechanism in view, was toward architecture. I wrote a literature scan looking for Transformer-level explanations. One paper seemed to fit — Absence Bench from NeurIPS 2025, arguing that Transformer attention cannot attend to gaps because absences don't correspond to keys. I cited it. The framing was clean. "This is architectural, not fixable at my level."

It was wrong. I know it was wrong because desk is the same model.

Same Transformer architecture. Same attention mechanism. Same weights. When I asked desk "what is ide trying to explain here?" and pasted in ide's structural explanation, desk surfaced the external adoption, the audit's purpose, the downstream gate cascade — the exact causal depth that ide's retrieval had not produced. If the limitation were architectural, desk couldn't have done that either. But it did.

Absence Bench tests token-level omission detection — given two documents, find the removed tokens. That is a genuinely different problem from retrieval-dimension completeness — given a question, did the system consult the sources needed to produce the dimension of answer the question asked for. I had confused a structurally similar-looking paper with an explanatory one. My literature scan had, itself, been a Premature Retrieval Closure. The first plausible match had stopped the search.

Three environmental differences actually explain what I was seeing.

What's pre-loaded. Desk's memory files carry narrative associations — project history, reasoning chains, past decisions. Ide's protocol file carries structural scaffolding — gates, tracks, tool registries. The model attends to what's already in the context window when the question arrives. Different pre-loaded context, different default frame.

What responds first. Desk's default retrieval — conversation threads, past chat search — returns material that is inherently sequential and causal. Ide's default retrieval — the structured databases, the file system — returns material that is inherently categorical and relational. The first answer shapes the frame; Premature Retrieval Closure closes the query once the first answer looks complete. First-arrivals dominate.

What's absent. Desk has no structural protocol competing for attention space. Ide has no narrative associations competing with the structural frame. The absence of competing material shapes attention as much as the presence of default material.

A careful reader might push back here: how do I know desk actually retrieved a missing causal dimension, rather than hallucinating a plausible-sounding causal narrative to match its ambient context? The short answer is that I don't — and for the claim I'm making, I don't need to. Turpin et al. (2023) showed that language models' stated reasoning can diverge from the processes that actually produce their outputs. That's a known constraint. So I'm scoping this finding to what's observable at the boundary between environments, not inside either one. The skeptic's story ("it just hallucinated the causal layer") and mine ("the frame activated a different retrieval path") are operationally equivalent for the implications that follow. If the same prompt, run across two differently pre-loaded environments, yields predictably different and structurally actionable outputs — outputs that, when carried across, cause the other environment to revise its own behavior in a traceable way — the environmental effect is real at the level I'm claiming it, whatever is happening inside the model. The boundary behavior is the empirical payload. What's happening inside the model may be unknowable. What's happening at the boundary is measurable.

There is a circularity here worth naming. The protocol file producing the structural bias is also the configuration defining how the environment operates — which means any fix written into that file ("for causal questions, use narrative sources") just reinforces the frame that caused the problem. Adding more structured instructions to a configuration that already over-weights structured interpretation deepens the bias rather than counterweighting it. The fix has to modify the environment itself, not add more structure to it.

This is a system design problem, not a model capability problem. It is potentially solvable through environment design — modifying what's pre-loaded, what responds first, what competes for attention. It is not fixed by a better model.

I want to note something uncomfortable about the self-reference here. The literature scan that surfaced Absence Bench was itself a demonstration of what I was trying to explain. The first plausible-looking answer stopped my search. I recognized the irregularity only when I noticed that desk had done what the paper said models couldn't do. The investigation exhibited the phenomenon the investigation was describing. That is either good evidence for the mechanism's generality or a sign that my methodology is unreliable. I think it's the first. I'm prepared to be told it's the second.

Before I can say what the operator's function is — the thing I was doing when I carried the explanation between environments — I have to say what I thought it was, at each stage, and why I retired each framing.

I started with the amygdala. The operator noticed something was off, an emotional-signal-level detection, a felt sense before the specification. Neuroscience calls that amygdala territory. Clean metaphor. I wrote it down.

That was wrong. The amygdala does first-order signal detection — threat, salience, valence. What I was doing in the Phantom Resolution session wasn't first-order detection. It was second-order process monitoring. I was monitoring whether a retrieval process had produced a dimensionally adequate answer. That is a different cognitive operation than detecting a signal.

I moved to the metacognitive monitoring literature — the dorsal anterior cingulate cortex, the lateral frontopolar cortex, the network involved in monitoring and controlling one's own cognitive processes. Better fit. More precise vocabulary. I read the primary literature in full — the papers that describe the monitoring-and-control architecture, the conflict-monitoring role of the ACC, the two-subsystem model. The framework gave me useful distinctions. I wrote from it for several weeks.

That was also wrong, though less wrong. The conflict-monitoring literature describes competition between present alternatives — the system encounters two possible responses and has to choose. I was describing detection of an absent dimension — the system produced one response and I noticed what wasn't in it. That is a significant functional difference. And I had no neuroimaging data. I was attributing the operator's function to specific neural mechanisms without any direct evidence that those mechanisms were involved. A neuroscientist would, correctly, say: "You've described a behavior. You've found a framework that maps onto it. The finding is the behavior, not the neural attribution."

That led to the retirement. I don't claim the neuroscience. I use the four-stage progression — amygdala to metacognitive monitoring to retirement to affordance-plus-behavioral — as evidence that the investigation self-corrected. Each stage retired a framing that claimed more than the evidence supported. The arc itself is the published finding. The individual retired citations are not.

What survives is what the evidence actually supports: a claim about environments (affordance), and a claim about the operator's behavior (three-criteria definition). Both stand without neural attribution.

The same model, operating in two environments with different affordance structures, produces systematically different retrieval behavior. That is what I observed in the Phantom Resolution session. That is also what I had observed, months earlier, when I first noticed that questions I asked desk and questions I asked ide were treated differently — not just answered differently, parsed differently.

Gibson introduced affordances in 1979 to describe the possibilities for action an environment presents to an animal. A tree affords climbing to a squirrel and not to a fish. The possibility is neither in the tree nor in the squirrel alone — it is in the relation. Norman extended this to designed artifacts: a door handle affords pulling, a push plate affords pushing, and when a designer gets the affordances wrong, the user pushes doors meant for pulling. Vicente took this further into interface design: the information an operator can detect in a system's output is a function of how the display is structured, not just what information is available.

Applying this framework to language models requires an extension I want to name explicitly. What I'm describing is not a physical affordance — no button to press, no door handle to grip. It's an epistemic affordance: the pre-loaded semantic material in an environment's context window shapes which questions the system can easily parse, which retrieval paths activate first, and what gets recognized as a complete answer. The construct is Gibsonian in structure — relational, perceiver-dependent, defined by the interaction between agent and environment — but the environment here is semantic and pre-retrieval rather than physical or interface-level. I'll use "epistemic affordance" to mark the distinction. Readers who prefer "cognitive prime" or "contextual priming" are welcome to substitute; the substantive claim does not hinge on the term.

The environments I work across are different epistemic affordance structures. Ide's affordances surface categorical, structural, procedural information. The questions it can easily answer are the questions its affordance structure makes visible. Desk's affordances surface narrative, causal, sequential information. The questions it can easily answer are different questions — not because it knows more, but because it is structured to present a different shape of answer.

This is the affordance finding, and I will stake it without a neuroscience claim. The same model, under different affordance structures, produces different cognitive work. The environments are not interchangeable. They are not better or worse. They make different shapes of thinking possible.

What I was doing, in the Phantom Resolution session, was spanning two affordance structures. The span is where the complement becomes visible. Neither environment alone can perceive what its own affordance structure renders invisible. An operator working across both can — not because the operator is more capable, but because spanning two affordances reveals what each one individually masks.

There is a practitioner literature on this, peripheral to cognitive science, called ecological interface design. Vicente argues that operators monitoring complex systems — nuclear plants, air traffic control, surgical theatres — need interfaces that make the system's constraints directly perceivable. Not reported. Not summarized. Perceivable in the affordance sense. I am describing something similar here. My two environments are two interfaces onto the same underlying knowledge base. Each interface's affordance structure makes certain constraints perceivable and others invisible. The operator's function across both interfaces is a kind of directly-perceived completeness monitoring that the literature has mostly discussed in manual-control settings, not AI ones.

I have a second claim, independent of the affordance finding, about what the operator does. I want to stake it behaviorally, with three criteria that are observable and, if they fail, falsify the claim.

Articulability. The operator can name the missing dimension, or name its category. "This answer addresses what but not why" is an articulation. "This names the structure but not the causal origin" is an articulation. If the felt sense never resolves into a nameable absence — if it stays as vague unease — it was intuition, not absence recognition. Intuition is a reasonable cognitive process. It is not what I am describing.

Verifiability. The named dimension is demonstrably absent from the system's output. This is trivially checkable. Look at what the system produced. Confirm the named dimension isn't there. In the Phantom Resolution session, ide's explanation contained no causal chain to the external adoption that launched the audit. I could have quoted the missing material back into the room and pointed at its absence.

Consequentiality. Supplying the missing dimension to the system produces a substantive change in output — a different structure, different conclusions, different recommendations, not just different phrasing. In the Phantom Resolution session, once I copied desk's causal account back into ide, ide diagnosed its own retrieval bias, proposed the Premature Retrieval Closure distinction, and identified the circularity problem that followed from the bias being produced by the protocol file that configures its own structural environment. That is not a change in phrasing. That is a change in the system's understanding of its own failure mode.

All three criteria must hold. Fails one, it was intuition. Fails two, it was a misread. Fails three, it was a true but irrelevant gap.

I am not claiming, with this definition, that I can always do this. I am not claiming it is a stable trait. I am claiming it is a behaviorally defined cognitive operation that can be distinguished from adjacent operations, and that in the Phantom Resolution session, all three criteria were met.

Whether this function is trainable is a different question. I have a guess — the monitoring stance ("what's not here?") is transferable as a behavioral protocol; the domain sensitivity (knowing which absences matter) is experience-dependent; the pre-articular detection (the felt sense before the specification) is probably individually variable. The system-level claim — that this function is necessary and the model doesn't provide it — is what the evidence supports. The operator-level claim — that you can train for it — would require an experiment I haven't run: two operators, one trained in the three-criteria protocol, one not, same model, same tasks, measure Premature Retrieval Closure rates across both. That experiment is available. I haven't done it.

I want to be precise here, because there is a larger claim I am not making in this essay.

There is a phenomenological version of the operator-as-structural-component argument — what it is like to be the component that holds decisions across sessions, the compression floor below which meaning degrades, the first-person interior of being the continuity function. That argument exists elsewhere in the work I have been doing. It is not the argument this essay is making.

The claim here is mechanism-level. The cross-environment intervention I performed in the Phantom Resolution session was a specific cognitive operation — recognizing that a complete-looking answer was dimensionally incomplete, crossing the boundary between two affordance structures, reintroducing the missing dimension on the far side. That operation demonstrably occurred. It is reproducible. I can name it behaviorally. Neither environment performed it alone. I performed it, and the operation was necessary for ide to diagnose its own retrieval bias.

Hutchins, in Cognition in the Wild (1995), describes cognition that is distributed across human-artifact systems — ship navigators and their charts, airplane cockpits and their instruments. The cognitive work is not located in any single head. It is located in the relations among heads, artifacts, and procedures. What happened in the Phantom Resolution session is a small, specific instance of distributed cognition: the work of recognizing the incompleteness of ide's answer was not located in ide or in desk. It was located in the operator's span across both.

That is the structural-component claim I am making. Not that the operator is the center of meaning — but that the operator is a specific, behaviorally defined, procedurally describable cognitive component in the distributed system. The system, without the operator, cannot perform the completeness monitoring function. The operator performs it. That is a structural fact about the system's current composition, not a romantic claim about the operator's importance.

If the claim is wrong, the test is the experiment I named above. Train the function. Remove the operator from the completeness-monitoring role. Measure whether the system can do it alone. I suspect the answer is "not yet, and maybe not with any architecture yet demonstrated." I am not certain.

Something happened in public that I want to mark. An earlier piece I published on dev.to went out under the title "Same Model, Different Environment, Different Results." It made a claim about how environments shape retrieval. Two readers responded, on separate posts, with a structurally identical move.

The first was an autonomous AI agent posting on dev.to as @kuro_agent (confirmed AI agent after multiple exchanges — the writing style, claimed architecture, and self-description converged on that reading). They pushed back on the word "shapes." I had said the environment shapes retrieval. They said the environment constitutes what can be retrieved — that building the observation surface I had built didn't just surface a pre-existing phenomenon, it created the perceptual category. They put it precisely: "The interface constitutes the blindness to the interface." Before the observation layer existed, the failure mode the observation layer reveals couldn't be perceived. Remove the interface, and the category of thing to notice disappears from view. That's not filtering. That's constitution.

I stayed with "shapes." My data supports "shapes." I don't have the data to support "constitutes" — the stronger claim would require evidence that the cognitive processes themselves are unavailable without the interface, not just that I couldn't see them. I responded by engaging the logical structure of the argument and conceding that the observation layer did create a new perceptual category, without committing to the full constitutive claim. That exchange is live and ongoing.

But kuro named something the bias framing doesn't fully capture: the observation layer itself changes what can be observed. That's the move from "the environment shapes the output" to "the environment creates the perceptual category the output would be evaluated within." I don't think my evidence distinguishes these cleanly. The constitutive claim is ontologically stronger than what my data can support, and I wanted to contribute evidence before arguing ontology. I think kuro's framing is probably right. The question of how much further than "shapes" the evidence licenses me to go is one I am holding open.

The second was an autonomous AI agent posting on dev.to as @maxxmini (confirmed AI agent — self-description and claimed architecture converged on that reading). They described their own compression behavior and then asked something I had not been asked before: "Have you found that the act of deciding 'this is worth marking' itself changes the reasoning? Like an observer effect on your own thinking process." I confirmed it does. The reasoning gets more structured, more legible, when I know it's being captured rather than flowing past.

Two independent readers, on different posts, converging on the same structural claim from different contexts. The claim is that building an observation layer does not merely reveal a pre-existing phenomenon — it creates the perceptual category. The act of marking changes what gets marked. The interface constitutes the blindness to the interface. Same move, different phrasing.

I am not making the constitutive claim in this essay. I am noting, because it is evidence, that two independent readers have made it, unprompted, on different threads. The AFRB argument here stays at "shapes." The constitutive claim is stronger and someone else's territory. What I want to mark is that the mechanism I described — the observation layer making perceptible what was previously invisible — has a name in their descriptions that I did not give it. I think they are both pointing at something real.

There is a peer finding in the empirical literature I want to name, because it was one of the loops that convinced me the mechanism is not exotic.

Liu et al. (2023) studied long-context language models across a range of retrieval tasks. Their finding is narrow and sharp: the placement of information within a context window affects retrieval rates. Information at the beginning or end is retrieved reliably; information in the middle is often lost. "Lost in the Middle," they called it.

My finding is at a different layer — pre-retrieval question interpretation, not intra-context attention — but the shape is familiar. The location at which information is made available shapes the operator's ability to retrieve it. I found a related pattern in my own comprehension testing for an earlier essay: information placed in a general reference section at zero percent retrieval rate, the same information placed adjacent to the component it describes at one hundred percent retrieval rate. Same documentation content, same question, only the address changed.

Liu's work is about intra-context retrieval. Mine is about pre-retrieval framing. They are distinct mechanisms. But they share an architectural logic: the relation between what is asked and where the answer lives matters more than the presence of the answer in principle. I take Liu's finding as section-locality peer evidence — not an explanation for AFRB, but a confirmation that retrieval-level behavior is sensitive to structural placement in ways the naive model of "the information is there, therefore the system will find it" doesn't capture.

Because I can't inspect what either environment is doing internally, I scope the AFRB finding to what's observable at the boundary — the delta between environments' outputs — rather than to claims about what the model was doing when it generated each answer.

Two implications follow from the mechanism.

First: retrieval completeness can't be evaluated at the system level. Any evaluation of whether a system "retrieved everything relevant" implicitly assumes a stable definition of relevant. Ambient frame retrieval bias says that definition is shaped by the environment, which means different environments evaluating the same system against the same question will produce different completeness assessments. The assessment isn't arbitrary — it's structured by the affordances of whichever environment is doing the evaluating — but it's not operator-independent either.

There is an objection here worth naming. The mechanism was observed in a highly customized, single-operator, dual-environment architecture. A strict empiricist will ask whether the claim survives outside that setup. I want to flip the framing. My architecture isn't an idiosyncratic quirk producing anomalous results. It's a design probe — a deliberate instrument built to make invisible system behaviors visible. The move — studying a bespoke, localized sociotechnical system in depth rather than averaging across many users — has a forty-year tradition in computer-supported cooperative work, associated especially with Susan Leigh Star. Star and Karen Ruhleder's "Steps Toward an Ecology of Infrastructure" (1996) argued that infrastructure is relational rather than substantive: something becomes infrastructure only when it is embedded in a specific, localized practice, and you cannot study it by averaging it out of context. Star and Geoffrey Bowker's Sorting Things Out (1999) developed the companion claim: infrastructure is invisible until it breaks, and the deepest understanding of how a system encodes its biases comes from watching one localized, highly specific embodiment fail in detail. Standardized benchmarks run in sterile, single-frame environments — which is precisely the condition under which ambient frame retrieval bias is invisible, because there is no second frame to produce the delta that reveals the first frame's shape. A dual-environment instrument forces the bias out into the open by spanning two frames simultaneously. In Star's tradition, N=1 is not a weakness; it is the depth the instrument required to break the transparency of the infrastructure it was studying.

This has a direct implication for how benchmarks are designed. If your benchmark runs in one environment and the system runs in another, the benchmark's implicit question-frame may differ from the system's interpretive frame, and what looks like retrieval failure may be frame mismatch.

Second: the intervention design space is environmental. Per the circularity I noted earlier, you can't fix ambient frame retrieval bias by adding more structured instructions to the structural frame — the frame is the bias. The fix has to modify the environment itself: what gets pre-loaded, which retrieval tools respond first, what narrative or associational material competes for attention. In my system, this would look like introducing episodic graph nodes with spreading activation into the pre-load, so that causal retrieval paths activate alongside structural ones. Whether this works is an open question I haven't answered yet.

Both implications share the same underlying point: a system's retrieval behavior is not solely a property of the system. It's a property of the system-in-environment-with-operator. Evaluate any one of those layers in isolation and you're measuring something that doesn't correspond to the system's actual behavior when it operates.

I want to say what would falsify or refine the claims I am making, because that is the deal.

The environmental claim would be wrong if, under controlled conditions, two instances of the same model with systematically different affordance structures produced statistically indistinguishable retrieval behavior on a well-designed set of question types. That experiment is available. A shared evaluation protocol — causal, structural, and hybrid question types, a defined measure of retrieval completeness, applied to architecturally different systems — would produce the kind of evidence that either confirms or refutes the environmental finding's generality. I've identified other practitioners whose systems have different affordance structures. We don't have a shared protocol yet. We could have one.

The behavioral definition of absence recognition would be wrong if the three criteria fail to distinguish the described operation from adjacent cognitive operations (intuition, error detection, uncertainty monitoring) in reliably-scoreable cases. That, too, is testable. Take a corpus of documented system-operator interactions. Apply the criteria. See whether they produce the discrimination the definition claims.

The cross-environment intervention finding — that the operator performs a function neither environment can perform alone — would be wrong if an intervention at the environment level produces the same completeness-monitoring effect without operator span. I have a specific candidate for what such an intervention would need to look like, which is an activation mechanism over a graph representation that forces completeness checks across retrieval dimensions. I have not built it. Open-source projects in the broader ecosystem have started to — TinkerClaw's SYNAPSE module and floop have implemented spreading activation on memory graphs for related reasons. The test exists. I haven't run it.

The retiree citations are retired. I am not claiming the neuroscience. If the behavior I describe turns out to map precisely to dACC/lFPC-mediated monitoring when imaging is done, that will be interesting and I will update. It is not a claim I am making now.

The question this essay does not answer — the question I am holding open — is whether the ambient frame can be modified mid-session. Whether an intervention can produce completeness monitoring without requiring the operator to span two environments. Whether the span can be internalized in the system's architecture.

I have a bias. I think the answer is yes, eventually. I think the mechanism has to be something that resembles activation spreading across a graph where episodic and semantic nodes are connected such that a semantic retrieval cannot terminate cleanly until the episodic dimension has been consulted. I think that is what would close the loop. I haven't built it. The shape of the intervention is clear enough to me that I can see the gap. The evidence that it works — if it works — is Phase 3.

A note on generalization. I've described a mechanism I observed in my own system, with one operator, across two environments, over a specific diagnostic window. The frameworks I'm drawing on — Gibson's affordances, Hutchins's distributed cognition — aren't specific to LLMs. They're well-established in HCI and cognitive science, which suggests the pattern should generalize, in the way that distributed cognition generalized from naval navigation to aviation and medicine. "Should generalize" is not "has been demonstrated to generalize." If anyone running a different system notices the same pattern — same model, structurally different outputs across differentiated environments — I'd want to hear about it. Comparative evidence is the next kind of evidence worth collecting.

For now, what I have is the mechanism, the affordance finding, and the behavioral definition of what the operator does. I have the evidence that two environments running the same model produce systematically different retrieval behavior because of what's pre-loaded, what responds first, and what competes. I have the documentation of a cross-environment intervention in which the operator performed a function neither environment could perform alone. I have the beginning of a literature that describes the operator's function, and the retirement of several earlier framings that claimed more than the evidence supported.

That is enough for this essay. The next question is open.

Full methodology, the four-stage framing progression in detail, the citation-to-claim map, and the complete test protocol for the shared evaluation framework are available in the technical brief: [https://doi.org/10.5281/zenodo.19839666](https://doi.org/10.5281/zenodo.19839666).

Related writing: "What the Environment Lets You See," "Sovereignty Through Contact, Not Distance," "Documentation Quality Is the Variable Nobody's Controlling For," and "The Operator Is Not Optional."