# The return must be able to disagree A loop can keep returning the answer it has made easiest to receive. Another image sharpens the same mistaken reconstruction. A successful warning removes the event it warned about. A search finds the most convincing result among the candidates it happened to inspect. In each case something came back. What that return warrants depends on how it was produced. The three investigations began with different questions. They now press on the same temptation: to treat a convincing return as though it had answered every alternative to the explanation we favor. The [visual notebook](index.html) makes that temptation inspectable. Its six executable models and ten figures establish conditional results; none contains a living experiencer. They cannot enact or disprove recognition. The telescope gave us the first distinction. Alex's point about the horizon was right to move attention beyond a single view. The historical example needs Jupiter's moons, while Saturn supplies an especially useful complication: the object can be there and the instrument can still leave its structure obscure. In the [small observation model](../when-the-loop-must-wait.md), another sum view cannot distinguish left from right. A difference view can, provided we know the wiring. Waiting changes the available evidence only when something informative eventually arrives. Distance alone gives no universal timetable for understanding. Noise makes the distinction harder to see. Repeated measurements really do help under a stable calibration. With twenty acquisitions, nine spent on calibration and eleven on the target, the [checked rule](noisy-observation/index.html) errs about **0.826%** of the time. Reverse the wiring between those stages without telling the rule, and its error becomes **99.174%**. The apparently decisive record also fits the opposite object under stable wiring. More confidence in the fitted answer cannot identify which of those worlds produced the record. That is a specific failure of an assumption, not an indictment of measurement. It also sharpens what it means for the world to push back. In this protocol, the relevant disagreement cannot get through. Correcting that requires evidence or justified knowledge capable of distinguishing the calibration histories. Repeating the same protocol does not supply it. Elsewhere, a new inference from an old archive may be enough; correction does not always require a new physical measurement. Prediction moves the difficulty into what we do next. In the [prevention example](index.html#prediction), acting on the warning changes the outcome rate. Removing prevention because the rate fell makes the rate rise again. The estimate of what just happened was accurate. Carrying it across a change of action was the mistake. This gives the “handshake” a concrete scientific possibility without pretending to explain the experience that suggested it. A prediction can enter an action, and the action can help make the next scene. Successful prevention can make a warning appear false; other interventions can help produce the pattern later treated as confirmation. Establishing either account requires the causal path and its alternatives. The model says nothing about thought creating an external scene without such a path, or about the origin of lucid or sacred experience. The resemblance is a question-generating connection, not evidence that these are one mechanism. Finding out has consequences too. The [learner extension](learned-effects/index.html) must acquire both action risks within the same 256-unit horizon in which it acts. In the strong-effect world, expanding its balanced audit from twenty to 128 units reduces the probability of choosing the wrong final action from **3.2143% to 0.004645%**. Yet total expected excess loss rises from **5.2757 to 19.2018** loss units. During the larger audit, more units receive the worse action. The calculation makes the cost of learning visible; it does not supply the correct cost to place on harm, or an optimal policy for every world. The third investigation prevents this argument from becoming a sermon against surprise. There is a genuinely better candidate in the [shared-evidence model](selection-dependence/index.html). It can be found. There are also enough ordinary candidates for a lucky one to look better. Under independent scores, increasing the total from two to one hundred raises the expected displayed winning score from **65.1471% to 67.8616%**, while lowering the selected candidate's average true accuracy from **64.3639% to 57.0391%**. What improves on the page can deteriorate in the thing selected. The details matter. Shared item difficulty improves recovery of the genuine signal in this particular construction: at one hundred candidates, recovery rises from **46.9276% to 54.9248%** when the dependence setting goes from zero to 0.20. That is not a general benefit of correlation. Nor do selection and confirmation become the same event. The figures separate finding the signal, accepting it, and accepting a lucky null. Fresh confirmation buys another sixty-four items. Its protection has an evidence cost. An unexpected pattern can therefore earn the next question before it earns belief in its answer. Demanding that every observation arrive with its question already fixed would close off the possibility that the encounter changes what we know to ask. Letting the same observation choose its explanation and certify it without accounting for that choice creates the opposite failure. Inquiry needs both the freedom to be redirected and an honest account of what warrants the new claim. For Circulatory Epistemology, this leaves two distinct questions. What became meaningful, consequential, or transformative in the encounter? What could distinguish the proposed explanation from the alternatives that remain? The first belongs to the framework's account of living recognition. The second concerns the evidential reach of a particular claim. Answering either does not settle the other. Doc III already distinguishes the detector's trace from the sensor's recognition of its meaning. Doc VIII explicitly separates a real experience of opening from the truth of a particular connection attributed to it. Doc XI gives the proposed mechanism a test it has not yet passed. Those distinctions survive these models. A human's presence is insufficient evidence that the available record discriminates between the relevant alternatives; that is a criticism of the proxy, not an experiment proving that humans are unnecessary for knowing. The models were made and interpreted by people, and contain no variable for lived recognition. One sentence in Doc III now deserves a direct conversation: “Beauty is the felt signature of truth circulating well.” Read as a description of how successful recognition can feel, it leaves room for error. Read as a way to certify success by the feeling, it needs a distinction between recognition and compelling misrecognition. None of our calculations measures beauty or settles that distinction. I have left the sentence intact. The question is too central to resolve by quietly substituting a safer philosophy. The strongest practical proposal I can see is to bring the sensor's stakes into **which unresolved alternative is worth pursuing**, then make the instrument account for what could actually distinguish it. A measurement can be precise about something that no longer matters. An encounter can reveal why a different question matters, including a question the original objective concealed. Treating that change as part of inquiry gives the living pole real work to do. It still does not make the preferred explanation true. This proposal must earn a contribution beyond good existing practice. A fair study would compare it with an equally resourced human–instrument baseline already using calibration checks, causal design, selection controls and decision analysis. Both conditions would include living participants and have the same data, tools, time and access to their stakes. The comparison tests the added practice, not whether human presence is necessary. The proposed CE practice would explicitly connect a reported change in meaning or consequence to a revised question, a live alternative and a discriminating next step. Independent review could assess recovery of planted alternatives, unsupported promotion of claims, correction after a mechanism changes, and decision loss; participants would separately assess whether the revised question captured what mattered. Better reports of engagement alone would not establish better external inference. No such comparison has been run here, and the technical controls are not CE inventions. Three questions remain for Alex's return. When recognition happens, what changes: truth itself, knowledge, meaning, or our relation to an already true proposition? Can an encounter be alive and transformative while its external explanation is false—and if so, what distinguishes recognition from conviction? What result would persuade us that this proposed CE practice adds nothing to a strong existing method? The next work depends on those answers. These models have made the point at which an answer could change the framework harder to overlook. --- This is a local research synthesis, not a new formal appendix or a revision of the published core. The [synthesis ledger](synthesis-ledger.md) distinguishes model support, philosophical interpretation and proposed experiments. The [completion receipt](synthesis-verification.json) records verification. The original precision pass remains scoped to the ten claims in its ledger (`meta/core-document-precision-2026-09-06/claim-ledger.md`); unresolved physics claims elsewhere in the chapters have not thereby passed review.